Processing math: 100%
Research article Special Issues

Differential diagnosis of breast cancer assisted by S-Detect artificial intelligence system


  • Received: 09 February 2021 Accepted: 13 April 2021 Published: 27 April 2021
  • Objective 

    Traditional breast ultrasound relies too much on the operation skills of diagnostic doctors, and the repeatability in different doctors was low. This study aimed to evaluate the assistant diagnostic value of S-Detect artificial intelligence (AI) system in differentiating benign from malignant breast masses.

    Methods 

    The ultrasound images of 40 patients who underwent ultrasound examination in our hospital were collected. The conventional ultrasound images, elastic images, and S-Detect mode of breast lesions were analyzed. The breast imaging reporting and data system recommended by the American Society of Radiology (BI-RADS) classification for each breast mass was evaluated both by the doctor and AI. The receiver operator characteristics (ROC) curves were drawn to compare the diagnostic efficiency.

    Result 

    Among the 40 lesions, 16 were benign, and 24 were malignant. The S-Detect AI system had a high diagnostic efficiency for malignant mass, with sensitivity, specificity, and accuracy of 95.8%, 93.8%, and 89.6%. The accuracy of AI was higher than the elastic image and then than the conventional gray-scale image. With the assistance of the S-Detect AI system, the accuracy of BI-RADS classification was improved significantly.

    Conclusion 

    The S-Detect AI system will enhance breast cancer diagnostic accuracy and improve ultrasound examination quality.

    Citation: Qun Xia, Yangmei Cheng, Jinhua Hu, Juxia Huang, Yi Yu, Hongjuan Xie, Jun Wang. Differential diagnosis of breast cancer assisted by S-Detect artificial intelligence system[J]. Mathematical Biosciences and Engineering, 2021, 18(4): 3680-3689. doi: 10.3934/mbe.2021184

    Related Papers:

    [1] Bo An . Construction and application of Chinese breast cancer knowledge graph based on multi-source heterogeneous data. Mathematical Biosciences and Engineering, 2023, 20(4): 6776-6799. doi: 10.3934/mbe.2023292
    [2] Jiajia Jiao, Xiao Xiao, Zhiyu Li . dm-GAN: Distributed multi-latent code inversion enhanced GAN for fast and accurate breast X-ray image automatic generation. Mathematical Biosciences and Engineering, 2023, 20(11): 19485-19503. doi: 10.3934/mbe.2023863
    [3] Lingli Gan, Xiaoling Yin, Jiating Huang, Bin Jia . Transcranial Doppler analysis based on computer and artificial intelligence for acute cerebrovascular disease. Mathematical Biosciences and Engineering, 2023, 20(2): 1695-1715. doi: 10.3934/mbe.2023077
    [4] Yali Ouyang, Zhuhuang Zhou, Weiwei Wu, Jin Tian, Feng Xu, Shuicai Wu, Po-Hsiang Tsui . A review of ultrasound detection methods for breast microcalcification. Mathematical Biosciences and Engineering, 2019, 16(4): 1761-1785. doi: 10.3934/mbe.2019085
    [5] Jun Gao, Qian Jiang, Bo Zhou, Daozheng Chen . Convolutional neural networks for computer-aided detection or diagnosis in medical image analysis: An overview. Mathematical Biosciences and Engineering, 2019, 16(6): 6536-6561. doi: 10.3934/mbe.2019326
    [6] Muhammad Bilal Shoaib Khan, Atta-ur-Rahman, Muhammad Saqib Nawaz, Rashad Ahmed, Muhammad Adnan Khan, Amir Mosavi . Intelligent breast cancer diagnostic system empowered by deep extreme gradient descent optimization. Mathematical Biosciences and Engineering, 2022, 19(8): 7978-8002. doi: 10.3934/mbe.2022373
    [7] Jian-xue Tian, Jue Zhang . Breast cancer diagnosis using feature extraction and boosted C5.0 decision tree algorithm with penalty factor. Mathematical Biosciences and Engineering, 2022, 19(3): 2193-2205. doi: 10.3934/mbe.2022102
    [8] Xiao Zou, Jintao Zhai, Shengyou Qian, Ang Li, Feng Tian, Xiaofei Cao, Runmin Wang . Improved breast ultrasound tumor classification using dual-input CNN with GAP-guided attention loss. Mathematical Biosciences and Engineering, 2023, 20(8): 15244-15264. doi: 10.3934/mbe.2023682
    [9] Xiaoli Zhang, Ziying Yu . Pathological analysis of hesperetin-derived small cell lung cancer by artificial intelligence technology under fiberoptic bronchoscopy. Mathematical Biosciences and Engineering, 2021, 18(6): 8538-8558. doi: 10.3934/mbe.2021423
    [10] Xi Lu, Xuedong Zhu . Automatic segmentation of breast cancer histological images based on dual-path feature extraction network. Mathematical Biosciences and Engineering, 2022, 19(11): 11137-11153. doi: 10.3934/mbe.2022519
  • Objective 

    Traditional breast ultrasound relies too much on the operation skills of diagnostic doctors, and the repeatability in different doctors was low. This study aimed to evaluate the assistant diagnostic value of S-Detect artificial intelligence (AI) system in differentiating benign from malignant breast masses.

    Methods 

    The ultrasound images of 40 patients who underwent ultrasound examination in our hospital were collected. The conventional ultrasound images, elastic images, and S-Detect mode of breast lesions were analyzed. The breast imaging reporting and data system recommended by the American Society of Radiology (BI-RADS) classification for each breast mass was evaluated both by the doctor and AI. The receiver operator characteristics (ROC) curves were drawn to compare the diagnostic efficiency.

    Result 

    Among the 40 lesions, 16 were benign, and 24 were malignant. The S-Detect AI system had a high diagnostic efficiency for malignant mass, with sensitivity, specificity, and accuracy of 95.8%, 93.8%, and 89.6%. The accuracy of AI was higher than the elastic image and then than the conventional gray-scale image. With the assistance of the S-Detect AI system, the accuracy of BI-RADS classification was improved significantly.

    Conclusion 

    The S-Detect AI system will enhance breast cancer diagnostic accuracy and improve ultrasound examination quality.



    Multisite phosphorylation cycles are common in cell regulation systems [1]. The important distributive sequential n-site phosphorylation network is given by the following reaction scheme:

    S0+Ekon0koff0Y0kcat0S1+E  Sn1+Ekonn-1koffn-1Yn1kcatn1Sn+ESn+Fonn-1offn-1Un1catn1Sn1+F  S1+Fon0off0U0cat0S0+F. (1)

    Arrows correspond to chemical reactions. The labels in the arrows are positive numbers, known as reaction rate constants. This reaction scheme represents one substrate S0 that can sequentially acquire up to n phosphate groups (giving rise to the phosphoforms S1,,Sn) via the action of the kinase E, and which can be sequentially released via the action of the phosphatase F, in both cases via an intermediate species (denoted by Y0, Y1, , Yn1, U0, U1,,Un1) formed by the interaction of the substrate and the enzyme.

    The kinetics of this network is deduced by applying the law of mass action to the labeled digraph of reactions (1), which yields an autonomous polynomial system of ordinary differential equations depending on the positive reaction rate constants. This system describes the evolution in time of the concentrations of the different chemical species. We denote with lower case letters the concentrations of the 3n+3 species. The derivatives of the concentrations of the substrates satisfy (see e.g. [2] for the general definition):

    ds0dt=kon0s0e+koff0y0+cat0u0,dsidt=kcati1yi1konisie+koffiyi+catiuioni1sif+offi1ui1,i=1,,n1,dsndt=kcatn1yn1onn1snf+offn1un1,

    the derivatives of the concentrations of the intermediate species satisfy:

    dyidt=konisie(koffi+kcati)yi,i=0,,n1,duidt=onisi+1f(offi+cati)ui,i=0,,n1,

    and dedt=dy0dtdyn1dt, dfdt=du0dtdundt, which give two linear conservation relations. Indeed, there are three linearly independent conservation laws:

    ni=0si+n1i=0yi+n1i=0ui=Stot,e+n1i=0yi=Etot,f+n1i=0ui=Ftot, (2)

    where the total conservation constants Stot,Etot,Ftot are positive for any trajectory of the system intersecting the positive orthant.

    A steady state of the system is a common zero of the polynomials in the right-hand side of the differential equations. There are 6n+3 parameters: the reaction rate constants and the total conservation constants (which correspond to points in the positive orthant R6n+3>0). It is known that for any n2, the sequential n-site system shows multistationarity. This means that there exists a choice of parameters for which there are at least two positive steady states satisfying moreover the linear conservation relations (2) for the same Stot,Etot and Ftot. In this case, it is said that the steady states are stoichiometrically compatible, or that they lie in the same stoichiometric compatibility class. This is an important notion, because the occurrence of multistationarity allows for different cell responses with the same total conservation constants (that is, the same total amounts of substrate and enzymes).

    The possible number of positive steady states of the n-site phosphorylation system (for fixed total conservation constants) has been studied in several articles. In the case n=2, it is a well known fact that the number of nondegenerate positive steady states is one or three [3,4]. The existence of bistability is proved in [5]. In [6] and [7], the authors give conditions on the reaction rate constants to guarantee the existence of three positive steady states based on tools from degree theory, but this approach does not describe in general the total conservation constants for which there is multistationarity. In [7] Conradi and Mincheva give a sufficient condition on the reaction rate constants to guarantee multistationarity for suitable choices of total conservation constants. These total amounts are precisely given in implicit form. Specifically, after writing Stot,Etot and Ftot in terms of the concentrations at steady state of E,F and S0, as in the present paper, the values of total amounts that give rise to multistationarity are derived from those for which an explicit polynomial becomes negative, which allows to construct as many witnesses as wished. When n3, there could be more than three positive steady states, and our methods allows us to give lower bounds bigger than three for n>3.

    For an arbitrary number n of phosphorylation sites, it was shown in [4] that the system has at most 2n1 positive steady states. In the same article, the authors showed that there exist reaction rate constants and total conservation constants such that the network has n (resp. n+1) positive steady states for n odd (resp. even); that is, there are 2n2+1 positive steady states for any value of n, where . denotes integer part.

    In [8] (see also [9]) the authors showed parameter values such that for n=3 the system has five positive steady states, and for n=4 the system has seven steady states, obtaining the upper bound given in [4]. In the recent article [10] the authors show that if the n-site phosphorylation system is multistationary for a choice of rate constants and total conservation constants (Stot,Etot,Ftot) then for any positive real number c there are rate constants for which the system is multistationary when the total conservation constants are scaled by c. Concerning the stability, in [11] it is shown evidence that the system can have 2n2+1 positive steady states with n2+1 of them being stable. Recently, a proof of this unlimited multistability was presented in [12], where the authors find a choice of parameters that gives the result for a smaller system, and then extend this result using techniques from singular perturbation theory.

    In the previous paper [13], open parameter regions on all the parameters are given for the occurrence of multistationarity for the n-site sequential phosphorylation system, but no more than three positive steady states are ensured. These conditions are based on a general framework to obtain multistationary regions jointly in the reaction rate constants and the total conservation constants. The description of these open sets is explicit on a subset of the reaction rate constants and the total conservation constants, while some other reaction rate constants can vary in an open set whose form is described but which is not completely quantified. Our approach in this article uses the systematic technique in [13], which we briefly summarize in Section 3.

    The removal of intermediates was studied in [14]. More specifically, the emergence of multistationarity of the n-site phosphorylation system with less intermediates was studied in [15]. It is known that the n-site phosphorylation network without any intermediates complexes has only one steady state for any choice of parameters. In [15], the authors showed which are the minimal sets of intermediates that give rise to a multistationarity system, but they do not give information about how many positive steady states can occur, and also, they do not describe the parameter regions for which these subnetworks are multistationary.

    In this paper, we work with subnetworks of the sequential n-site phosphorylation system that only have intermediates in the E component (that is, in the connected component of the network where the kinase E reacts), see Definition 1. In case of the full mechanism on the E component or if we only assume that there are intermediate species that are formed between the phosphorylated substrates with parity equal to n (that is, roughly only 14 of the intermediates of the n sequential phosphorylation cycle), we obtain precise conditions on all the parameters to ensure as many positive steady states as possible. Indeed, we show in Proposition 7 that the maximum number of complex solutions to the steady state equations intersected with the linear conservation relations is always n+1, the maximum number of real roots is also n+1, that could be all positive when n is even, while only n of them can be positive when n is odd, so the maximum number of positive steady states equals 2n2+1 for any n. In Theorem 2 and Corollary 5 we describe open regions in all the parameters so that the associated phosphorylation/dephosphorylation system admits these number of positive steady states. This follows from Theorem 4. As we pointed out before, we give precise new conditions on some of the parameters, involving reaction rate constants and total conservation constants, while some other rate constants vary on open sets defined by inequalities that are not exactly quantified. In order to state these results, we need to introduce some notations.

    Definition 1. For any natural number n, we write In={0,,n1}. Given n1, and a subset JIn, we denote by GJ the network whose only intermediate complexes are Yj for jJ, and none of the Ui. It is given by the following reactions:

    Sj+EkonjkoffjYjkcatjSj+1+E,if jJ,Sj+EτjSj+1+E,if jJ,Sj+1+FνjSj+F,0jn1. (3)

    where the labels of the arrows are positive numbers. We will also denote by GJ the associated system of differential equations with mass-action kinetics.

    For all these systems GJ, there are always three linearly independent conservation laws for any value of n:

    ni=0si+jJyj=Stot,e+jJyj=Etot,f=Ftot, (4)

    where the total conservation constants Stot,Etot,Ftot are positive for any trajectory of the system of differential equations which intersects the positive orthant. Note that the concentration of the phosphatase F is constant, equal to Ftot.

    To get lower bounds on the number of positive steady states with fixed positive total conservation constants, we first consider the network GIn, that is, when all the intermediates in the E component appear. It has the following associated digraph:

    S0+Ekon0koff0Y0kcat0S1+E  Sn1+Ekonn-1koffn-1Yn1kcatn1Sn+ESn+Fνn1Sn1+F  S1+Fν0S0+F. (5)

    Theorem 2. Let n1. With the previous notation, consider the network GIn in (5), and suppose that the reaction rate constants kcati and νi, i=0,,n1, satisfy the inequality

    maxieven{kcatiνi}<miniodd{kcatiνi}.

    For any positive values Stot, Etot and Ftot of the total conservation constants with

    Stot>Etot,

    verifying the inequalities:

    maxieven{kcatiνi}<(StotEtot1)Ftot<miniodd{kcatiνi}, (6)

    there exist positive constants B1,,Bn such that for any choice of positive constants λ0,,λn1 satisfying

    λjλj1<Bj  for j=1,,n1,1λn1<Bn, (7)

    rescaling of the given parameters konj by λjkonj, for each j=0,,n1, gives rise to a system with exactly 2n2+1 nondegenerate positive steady states.

    Remark 3. We will also show in the proof of Theorem 2, that for any reaction rate constants and total conservation constants satisfying (6), there exist t0>0 such that for any value of t(0,t0), the system GIn has exactly 2n2+1 nondegenerate positive steady states after modifying the constants konj by tjnkonj for each j=0,,n1.

    We now consider subnetworks GJ, with JJn where

    Jn:={iIn:(1)i+n=1}, for n1, (8)

    that is, subsets J with indexes that have the same parity as n.

    Theorem 4. Let n1, and consider a subset JJn. Let GJ be its associated system as in (3). Assume moreover that

    Stot>Etot.

    A multistationarity region in the space of all parameters for which the system GJ admits at least 1+2|J| positive steady states can be described as follows. Given any positive value of Ftot, choose any positive real numbers kcatj,νj, with jJ satisfying

    maxjJ{kcatjνj}<(StotEtot1)Ftot. (9)

    Then, there exist positive constants B1,,Bn such that for any choice of positive constants λ0,,λn1 satisfying

    λjλj1<Bj  for j=1,,n1,1λn1<Bn, (10)

    rescaling of the given parameters konj by λjkonj, for jJ and τj by λjτj if jJ gives rise to a system with at least 1+2|J| positive steady states.

    The following immediate Corollary of Theorem 4 implies that we can obtain a region in parameters space with n2 intermediates, where the associated system has 2n2+1 positive steady states.

    Corollary 5. Let n1, and consider the network GJn as in (3), with Jn as in (8). Assume moreover that

    Stot>Etot.

    Then, there is a multistationarity region in the space of all parameters for which the network GJn admits 2n2+1 steady states (with fixed total conservation constants corresponding to the coordinates of a vector of parameters in this region), described in the statement of Theorem 4.

    Similar versions of all our results work if one assumes that some or all of the reverse unbinding reactions do not occur, so that the rate constants koffi are set to zero (see Remarks 9 and 15). We thank one of the referees for pointing this out.

    We explain in Section 5 the computational approach to find new regions of multistationarity. This is derived from the framework in [13] that we used to get the previous results. We find new regions of multistationarity for the cases n=2,3,4 and 5, after some manual organization of the automatic computations. Our computer aided results are summarized in Propositions 17, 18, 19, 20, 21 and 22.

    The paper is organized as follows. In Section 2 we prove Proposition 7 and we show how to lift regions of multistationarity from the reduced system GJ to the complete sequential n-site phosphorylation system. We also show in Lemma 8 that even with a single intermediate Y0 it is possible to make a choice of all parameters such that the system has 2n2+1 positive steady states. This result has been independently found by Elisenda Feliu (personal communication). This says that while Corollary 5 is optimal, the regions obtained for any subset J with indexes of the same parity of n in Theorem 4 properly contained in Jn, only ensure 2|J|+1 positive steady states. However, note that we are able to describe open regions in parameter space and Lemma 8 only allows us to get choices of parameters.

    In Section 3 we briefly recall the framework in [13], which is the basis of our approach. In Section 4 we give the proofs of Theorems 2 and 4. Finally, as we mentioned above, we explain in Section 5 how to implement the computational approach to find new regions of multistationarity. We end with a discussion where we further compare our detailed results with previous results in the literature.

    In this section we collect three results that complete our approach to describe multistationarity regions giving lower bounds for the number of positive steady states with fixed linear conservation relations for the systems GJ in Definition 1 for any JIn. We first show a positive parametrization of the concentrations at steady state of the species of the systems GJ, which allows us to translate the equations defining the steady states in all the concentrations plus the linear conservation relations into a system of two equations in two variables. In Proposition 7, we prove that any mass-action system GJ, has at most 2n2+1 positive steady states. Lemma 8 shows that having a single intermediate is enough to get that number of positive steady states, for particular choices of the reaction rate constants. However, this computation by reduction to the univariate case is not systematic as the general approach from [13] that we use to describe multistationarity regions in Theorems 2 and 4, which can be applied to study other quite different mechanisms. It is known that if a system GJ has m nondegenerate positive steady states for a subset JIn, then it is possible to find parameters for the whole n-site phosphorylation system that also give at least m positive steady states (see [14]). In Theorem 10, we give precise conditions on the rate constants to lift the regions of multistationarity for the reduced networks to regions of multitationarity with 2n2+1 positive steady states (with fixed total conservation constants) of the complete n-sequential phosphorylation cycle.

    The following lemma gives a positive parametrization of the concentration of the species at steady state for the systems GJ, in terms of the concentrations of the unphosphorylated substrate S0 and the kinase E. It is a direct application of the general procedure presented in Theorem 3.5 in [16], and generalizes the parametrization given in Section 4 of [13].

    Lemma 6. Given n1 and a subset JIn, consider the system GJ as in Definition 1. Denote for each jJ

    Kj=konjkoffj+kcatj,τj=kcatjKj, (11)

    set T1=1, and for any i=0,,n1:

    Ti=ij=0τjνj. (12)

    Then, the parametrization of the concentrations of the species at a steady state in terms of s0 and e is equal to:

    si=Ti1(Ftot)is0ei, i=1,,n,yj=KjTj1(Ftot)js0ej+1, jJ, (13)

    The inverses K1j of the constants Kj in (11) in the statement of Lemma 6 are usually called Michaelis-Menten constants.

    Let n1 and a subset JIn. If we substitute the monomial parametrization of the concentration of the species at steady state (13) into the conservation laws, we obtain a system of two equations in two variables s0 and e. We have:

    s0+jJ(Tj(Ftot)j+1+KjTj1(Ftot)j)s0ej+1+jJTj(Ftot)j+1s0ej+1Stot=0,e+jJKjTj1(Ftot)js0ej+1Etot=0. (14)

    We can write system (14) in matricial form:

    C(s0es0es0e2s0en1)t=0, (15)

    where CR2×(n+3) is the matrix of coefficients:

    C=(10T0Ftot+β0T1(Ftot)2+β1Tn1(Ftot)n+βn1Stot01β0β1βn1Etot), (16)

    with

    βj=KjTj1(Ftot)j for jJ, and βj=0 if jJ. (17)

    Note that if we order the variables s0, e, the support of the system (that is, the exponents of the monomials that occur) is the following set A:

    A={(1,0),(0,1),(1,1),(1,2),,(1,n),(0,0)}Z2, (18)

    independently of the choice of JIn.

    We first recall Kushnirenko Theorem, a fundamental result about sparse systems of polynomial equations, which gives a bound on the number of complex solutions with nonzero coordinates. Given a finite point configuration AZd, denote by ch(A) the convex hull of A. We write vol to denote Euclidean volume, and set C=C{0}.

    Kushnirenko Theorem [17]: Given a finite point configuration AZd, a sparse system of d Laurent polynomials in d variables with support A has at most d!vol(ch(A)) isolated solutions in (C)d (and exactly this number if the polynomials have generic coefficients.)

    We also recall the classical Descartes rule of signs.

    Descartes rule of signs: Let p(x)=c0+c1x++cmxm be a nonzero univariate real polynomial with r positive real roots counted with multiplicity. Denote by s the number of sign variations in the ordered sequence of the signs sign(c0),,sign(cm) of the coefficients, i.e., discard the 0's in this sequence and then count the number of times two consecutive signs differ. Then rs and r and s have the same parity, which is even if c0cm>0 and odd if c0cm<0.

    We then have that 2n2+1 is an upper bound for the number of positive real solutions of the system of equations defining the steady states of any system GJ in Definition 1:

    Proposition 7. For any choice of rate constants and total conservation constants, the dynamical system GJ associated with any subset JIn has at most 2n2+1 isolated positive steady states. In fact, the polynomial system of equations defining the steady states of GJ can have at most n+1 isolated solutions in (C)d.

    Proof. The number of positive steady states of the subnetwork GJ is the number of positive solutions of the sparse system (14) of two equations and two variables. The support of the system is (18) whose convex hull has Euclidean volume equal to n+12. By Kushnirenko Theorem, the number of isolated solutions in (C)2 is at most 2!(n+1)2=n+1. In particular, the number of isolated positive solutions is at most n+1.

    Moreover, when all positive solutions are nondegenerate, their number is necessarily odd by Corollary 2 in [6], which is based on the notion of Brouwer's degree. Indeed, in our case, we can bypass the condition of nondegeneracy because we can use Descartes rule of signs in one variable. In fact, from the first equation of system (14), we can write:

    s0=Stotp(e), (19)

    where p(e) is the following polynomial of degree n on the variable e:

    p(e):=1+n1i=0(αi+βi)ei+1, (20)

    with

    αi=Ti(Ftot)i+1,i=0,,n1, (21)

    and βi=KjTj1(Ftot)j if jJ or βj=0 if jJ were defined in (17). Note that for any e>0 it holds that p(e)>0, and so s0>0. If we replace (19) in the second equation of (14), we have:

    e+n1i=0βiStotp(e)ei+1Etot=0. (22)

    The number of positive solutions of Eq (22) is the same if we multiply by p(e):

    q(e):=ep(e)+n1i=0βiStotei+1Etotp(e)=0. (23)

    This last polynomial q has degree n+1, with leading coefficient equal to αn1+βn1>0 and constant coefficient equal to Etot<0. The sign variation of the coefficients of q has the same parity as the sign variation of the leading coefficient and the constant coefficient, which is one. So, by Descartes rule of signs, as the sign variation is odd, the number of positive solutions is also odd.

    As we mentioned in the introduction, the following result has been independently found by Elisenda Feliu (personal communication).

    Lemma 8. If J={0}, then there exists parameter values such that the system GJ admits 2n2+1 positive steady states.

    Proof. Assume n is even, then n+1 is odd. As we did in the proof of Proposition 7, the positive solutions of the system (14) are in bijection with the positive solutions of the polynomial q(e) in (23). Here β0=K0 and βi=0 for i0. We will consider the polynomial ˜q(e):=q(e)Etot, with constant coefficient equal to 1.

    Consider any polynomial of degree n+1

    cn+1en+1+cnen++c1e1, (24)

    with n+1 distinct positive roots, and with constant term equal to 1. Then, we have that ci(1)i+1>0, and in particular, cn+1>0.

    Our goal is to find reaction rate constants and total conservation constants such that the polynomial (24) coincides with the polynomial ˜q(e). Comparing the coefficient of ei, for i=1,,n+1 in both polynomials, we need to have:

    αn1Etot=cn+1,αi2Etotαi1=ci, for i=3,,n,α0+K0Etotα1=c2,1+StotK0Etot(α0+K0)=c1. (25)

    Keep in mind that the values of ci are given. We may solve (25) in terms of Etot and of the ci,i=1,,n+1:

    αn1k=Etotki=0cn+1i(Etot)ki, for each k=0,1,,n2,α0+K0=Etotn1i=0cn+1i(Etot)n1i,1+StotK0=Etotni=0cn+1i(Etot)ni. (26)

    Note that if we take any value for Etot>0, then the values of αi for i=1,,n1, α0+K0 and StotK0 are completely determined. So, we find an appropriate value of Etot such that the previous values αi, K0 and Stot are all positive. For that, we choose K0=1, and we take Etot large enough such that

    ki=0cn+1i(Etot)ki>0, for each k{0,1,,n2} with k odd,Etotn1i=0cn+1i(Etot)n1i>1,Etotn1i=0cn+1i(Etot)ni>1.

    This is possible since cn+1>0 and that imposing the first condition just on k odd is enough to ensure that it holds for all kIn1 as well because of the signs of the ci. With these conditions, and using Eqs (26), the values of αi for each i=0,,n1 and Stot are determined and are all positive.

    Now, it remains to show that we can choose reaction rate constants such that the values of αi,i=0,,n1 are the given ones. Recall that αi=Ti(Ftot)i+1, where Ti=ij=0τjνj for i=0,,n1 and T1=1, where τ0=kcat0K0=kcat0 (we have chosen K0=1). Take for example Ftot=1, kon0=2, koff0=1 and kcat0=1 (to obtain K0=1). Then, τ0=1, so we take ν0=1α0. As αi+1=αiτi+1νi+1, for i=0,,n2, we can choose any positive values of τi+1,νi+1 such that τi+1νi+1=αi+1αi, and we are done.

    When n is odd, with a similar argument, we can find reaction rate constants and total conservation constants such that the polynomial ˜q(e) gives a polynomial like (24) (but with n distinct positive roots and one negative root).

    Remark 9. If we assume that koff0=0 (so, if there is no reverse unbinding reaction), Lemma 8 is still true. In this case, it is enough to set K0=kon0kcat0 and then choose kon0=1 instead of kon0=2 at the end of the proof.

    Multistationarity of any of the subsystems GJ can be extended to the full n-site phosphorylation system (for instance, by Theorem 4 in [14]). We give a precise statement of this result in Theorem 10.

    Consider the full n-site phosphorylation network (1), with a given vector of reaction rate constants κR6n:

    κ=(kon0,koff0,kcat0,,konn1,koffn1,kcatn1,on0,off0,cat0,,onn1,offn1,catn1).

    We define the following rational functions of κ:

    τj(κ)=kcatjμj(κ) if jJ and νj(κ)=catjηj(κ) for j=0,,n1, (27)

    where μj(κ) and ηj(κ) are in turn the following rational functions:

    μj(κ)=konjkoffj+kcatj if jJ and ηj(κ)=onjoffj+catj for j=0,,n1. (28)

    We denote by φ:R6n>0R2n+2|J|>0 the function that takes κ and gives a vector of (positive) reaction rate constants with the following order: first, the constants konj,koffj,kcatj,jJ, then τ(κ),jJ, and then νj(κ),j=0,,n1.

    Given a subset JIn and a vector of reaction rate constants κR6n>0, we consider the subnetwork Gφ(κ)J as in Definition 1, with rate constants φ(κ):

    Sj+EkonjkoffjYjkcatjSj+1+E,if jJSj+Eτj(κ)Sj+1+E,if jJSj+1+Fνj(κ)Sj+F,0jn1. (29)

    Applying a version of Theorem 4 in [14] that will appear in our forthcoming paper [18], we get the following lifting result.

    Theorem 10. Consider the full n-site phosphorylation network (1) with fixed reaction rate constants κ0 and the network Gφ(κ0)J, both with total conservation amounts Stot, Etot, Ftot>0. Suppose that system Gφ(κ0)J admits m nondegenerate positive steady states.

    Then, there exists ε0>0 such that for any choice of rate constants κ such that φ(κ)=φ(κ0) and

    maxjJμj(κ),maxjInηj(κ)<ε0, (30)

    the n-site sequential phosphorylation system admits m positive nondegenerate steady states in the stoichiometric compatibility class defined by Stot, Etot and Ftot>0. Moreover, the set of rate constants κ verifying φ(κ)=φ(κ0) and (30) is nonempty.

    Note that the function φ above is surjective, that is, any vector of reaction rate constants for the reduced network GJ can be obtained as φ(κ), for some vector κ of reaction rate constants of the full n-site phosphorylation network. For instance, given τjR>0, we can take konj=2τj,koffj=kcatj=1, and then τj(κ)=τj. Similarly, we can do this for the other reaction rate constants of GJ. Then, Corollary 10 allows us to obtain multistationary regions for the complete n-site phosphorylation system, combining the conditions on the parameters given in Theorem 2 and Theorem 4, with conditions (30) of Corollary 10. In particular, let JnIn as in (8). By lifting a multistationarity region for the system GJn in Corollary 5, we get a multistationarity region of parameters of the n-site phosphorylation cycle with 2n2+1 positive steady states in the same stoichiometric compatibility class.

    As we saw in Subsection 2.1, the steady states of the systems GJ correspond to the positive solutions of the sparse polynomial system (14) in two variables. In this section, we briefly recall the general setting from [13,19] to find lower bounds for sparse polynomial systems, that we will use in the proof of Theorems 2 and 4 in Section 4 and also in Section 5. For detailed examples of this approach, we refer the reader to Section 2 in [20].

    Consider a finite point configuration A={a1,,an}Zd, with nd+2. A sparse polynomial system of d Laurent polynomial equations with support A is a system f1(x)==fd(x)=0 in d variables x=(x1,,xd), with

    fi(x)=nj=1cijxajR[x1,,xd], i=1,,d, (31)

    where the exponents belong to A. We call C=(cij)Rd×n the coefficient matrix of the system and we assume that no column of C is identically zero. Recall that a zero of (31) is nondegenerate when it is not a zero of the Jacobian determinant of f=(f1,,fd).

    Our method to obtain a lower bound on the number of positive steady states, based on [13,19], is to restrict our polynomial system (31) to subsystems which have a positive solution and then extend these solutions to the total system, via a deformation of the coefficients. The first step is thus to find conditions in the coefficient matrix C that guarantee a positive solution to each of the subsystems. The hypothesis of having subsystems supported on simplices of a regular subdivision of a finite point configuration A is the key to the existence of an open set in the space of coefficients where all these solutions can be extended. We recall below only the concepts that we need for our work. A comprehensive treatment of this subject can be found in [21]. Following Section 3 in [19], we define:

    Definition 11. Given a matrix MRd×(d+1), we denote by minor(M,i) the determinant of the square matrix obtained by removing the i-th column of M. The matrix M is called positively spanning if all the values (1)iminor(M,i), for i=1,,d+1, are nonzero and have the same sign.

    Equivalently, a matrix M is positively spanning if all the coordinates of any nonzero vector in ker(M) are non-zero and have the same sign.

    A d-simplex with vertices in a finite point configuration A is a subset of d+1 points of A which is affinely independent. By Proposition 3.3 in [19], a system of d polynomial equations in d variables with a d-simplex as support, has one non-degenerate positive solution if and only if its d×(d+1) matrix of coefficients is positively spanning. We further define, following [19]:

    Definition 12. Let CRd×n and a finite point configuration A={a1,,an}Zd, with nd+2.. We say that a d-simplex Δ={ai1,,aid+1} is positively decorated by C if the d×(d+1) submatrix of C with columns {i1,,id+1} is positively spanning.

    Given a finite point configuration A, take a height function

    h:AR,h=(h(a1),,h(an)).

    Consider the lower convex hull L of the n lifted points (aj,h(aj))Rd+1,j=1,,n (see Figure 1). Project to Rd the subsets of points in each of the faces of L. These subsets define a regular subdivision of A induced by h. When the height vector h is generic, the regular subdivision is a regular triangulation, in which all the subsets are simplices. For the specifics on this notions we refer the reader to Section 2.2 of [21].

    Figure 1.  Regular triangulation.

    It can be proved that the set of all height vectors inducing a regular subdivision of A that contains certain d-simplices Δ1,,Δp is defined by a finite number of linear inequalities, see Chapter 2 of [21]. Thus, this set is a finitely generated convex cone CΔ1,,Δp in Rn with apex at the origin, which can be obtained as follows. Given a d-simplex Δ with vertices in A, we consider height vectors hRn, where each coordinate hj of h gives the value of the corresponding lifting function at the point aj of A. Denote by φΔ,h the unique affine function that agrees with h on the points of Δ, that is, φΔ,h(aj)=hj for all ajΔ. We associate with Δ the following cone:

    CΔ={h=(h1,,hn)Rn:φΔ,h(aj)<hj for all ajΔ}.

    Note that each inequality φΔ,h(aj)<hj is linear in h and it can thus be written as m,h>0, for a certain vector m that depends on the points in Δ (, denotes the standard inner product in Rn). As CΔ1,,Δp=pi=1CΔi, this gives a set of linear inequalities that define the cone CΔ1,,Δp. We refer the reader to Example 2.5 in [20] for a complete computation. The set of all height vectors inducing a regular subdivision Γ of A is a finitely generated convex cone in Rn, which we call CΓ. In particular, if the simplices Δ1,,Δp completely determine the regular subdivision Γ, the cone CΔ1,,Δp is equal to the cone CΓ

    Given a finite point configuration AZd, consider a system of sparse real polynomials f1=f2==fd=0 as in (31) with support A. Let Γ be a regular subdivision of A and h any height function that induces Γ. We define the following family of real polynomial systems parametrized by a positive real number t>0:

    nj=1cijth(aj)xaj=0, i=1,,d. (32)

    We also consider the following family of polynomial systems parametrized by γRn>0:

    nj=1cijγjxaj=0, i=1,,d. (33)

    Note that each polynomial system in the families (32) and (33) has again support A.

    Theorem 13 (Theorem 3.4 of [19] and Theorem 2.11 of [13]). Consider Δ1,,Δp d-simplices which occur in a regular subdivision Γ of a finite point configuration AZd, and which are positively decorated by a matrix CRd×n.

    1. Let h be any height function that defines Γ. Then, there exists a positive real number t0 such that for all 0<t<t0, the number of nondegenerate positive solutions of (32) is at least p.

    2. Let m1,mRn be vectors that define a presentation of the cone CΔ1,,Δp :

    CΔ1,,Δp={hRn:mj,h>0,j=1,,}.

    Then, for any ε(0,1) there exists t0(ε)>0 such that for any γ in the open set

    U=ε(0,1){γRn>0;γmj<t0(ε)εj,j=1,},

    system (33) has at least p nondegenerate positive solutions.

    We remark that in the first item in Theorem 13 (Theorem 3.4 in [19]), we describe a piece of curve in the space of coefficients as we vary t>0 where the associated system has at least p positive solutions, while in the second item in Theorem 13 (Theorem 2.11 in [13]), we describe a subset with nonempty interior in the space of coefficients where we can bound from below by p the number of positive solutions of the associated system.

    We start this section with a lemma.

    Lemma 14. Consider A={(1,0),(0,1),(1,1),(1,2),,(1,n),(0,0)}Z2. The triangulation Γ of A with the following 2-simplices:

    {(1,0),(1,1),(0,0)},{(1,1),(1,2),(0,0)},,{(1,n1),(1,n),(0,0)},{(1,n),(0,1),(0,0)}

    is regular (see Figure 2).

    Figure 2.  Triangulation Γ of A.

    Proof. We can take h:AR, with h(0,0)=0, h(0,1)=n and h(1,j)=j(j1)2, for j=0,,n1. It is easy to check h defines a regular triangulation that is equal to Γ.

    The idea in the proofs of Theorem 2 and Theorem 4 is to detect positively decorated simplices in the regular triangulation Γ.

    Proof of Theorem 2. By Proposition 7, the number of positive solutions of the system GIn is at most 2n2+1. So, it is enough to prove that this number is also a lower bound.

    The number of positive steady states of the system GIn is the number of positive solutions of the system (14). As we saw before, the support of this last system is

    A={(1,0),(0,1),(1,1),(1,2),,(1,n),(0,0)}Z2,

    with coefficient matrix C (16). Note that if one multiplies a column of C by a positive number, then a simplex is positively decorated by C if and only if it is positively decorated by the new matrix. After multiplying the columns by convenient positive numbers, we obtain the following matrix from C:

    Csimple=(10M0Mn1Stot0111Etot),

    where Mi=kcatiνiFtot+1, for each i=0,,n1. We will work with this new matrix Csimple.

    We consider the regular triangulation Γ in Lemma 14. The simplex {(1,0),(1,1),(0,0)} of Γ is positively decorated by Csimple if and only if EtotM0Stot<0. The simplex {(1,j),(1,j+1),(0,0)}, for j=1,,n1, corresponds to the submatrix

    Csimplej=(Mj1MjStot11Etot),

    and it is positively decorated by Csimple if and only if EtotMj1Stot and EtotMjStot have opposite signs. The last simplex {(0,1),(1,n),(0,0)} is positively decorated by Csimple if and only if EtotMn1Stot>0.

    Therefore we always have at least n positively decorated simplices using all simplices of Γ but the last one, just by imposing

    (EtotMiStot)(1)i<0,  for i=0,,n1. (34)

    We can include the last simplex if and only if n is even (because otherwise the inequalities are not compatible), and in this case we have at least n+1 positively decorated simplices. We can obtain 2n2+1 positively decorated simplices if the inequalities (34) are satisfied. These inequalities are equivalent to the inequalities (6) in the statement.

    Assume (6) holds. Given any height function h inducing the triangulation Γ, by item (1) in Theorem 13 there exists t0 in R>0 such that for all 0<t<t0, the number of positive nondegenerate solutions of the deformed system as in (32) with support A and coefficient matrix Ct, with (Ct)ij=th(αj)cij (with αjA, C=(cij)), is at least 2n2+1. In particular if we choose h as in the proof of Lemma 14, there exists t0 in R>0, such that for all 0<t<t0, the system

    s0+n1j=0(Tj(Ftot)j+1+KjTj1(Ftot)j)tj(j+1)2s0ej+1Stot=0,tne+n1j=0KjTj1(Ftot)jtj(j+1)2s0ej+1Etot=0, (35)

    has at least 2n2+1 positive solutions. If we change the variable ˉe=tne, we get the following system:

    s0+n1j=0(Tj(Ftot)j+1+KjTj1(Ftot)j)t(j+1)(j2n)s0ˉej+1Stot=0,ˉe+n1j=0KjTj1(Ftot)jt(j+1)(j2n)s0ˉej+1Etot=0. (36)

    It is straightforward to check that if we scale the constants Kj by

    tjnKj,j=0,,n1, (37)

    while keeping fixed the values of the constants kcatj, νj for j=0,,n1 and the values of the total conservation constants Etot, Ftot and Stot (assuming that condition (6) holds), the intersection of the steady state variety and the linear conservation equations of the corresponding network is described by system (36). It is easy to check that in order to get the scaling in (37), it is sufficient to rescale only the original constants konj as follows: tjnkonj, for j=0,,n1. Then, for these choices of constants, the system has at least 2n2+1 positive steady states.

    Now, we will show how to obtain the more general rescaling in the statement. The existence of the positive constants B1,,Bn follows from the inequalities that define the cone CΓ of heights inducing the regular triangulation Γ and item (2) in Theorem 13. For instance, we can check that CΓ is defined by n inequalities:

    CΓ={h=(h1,,hn+3)Rn+3:mj,h>0,j=1,,n},

    where m1=e12e3+e4, mj=ej+12ej+2+ej+3, for j=2,,n1 and mn=e2+en+1en+2en+3, where ei denotes the i-th canonical vector of Rn+3. Fix ε(0,1)n+3. As (6) holds, item (2) in Theorem 13 says that there exist positive numbers Bj for j=1,,n (depending on ε), such that the system

    γ1s0+n1j=0(Tj(Ftot)j+1+KjTj1(Ftot)j)γj+3s0ej+1γn+3Stot=0,γ2e+n1j=0KjTj1(Ftot)jγj+3s0ej+1γn+3Etot=0, (38)

    has at least 2n2+1 nondegenerate positive solutions, for any vector γRn+3 satisfying γmj<Bj, for all j=1,,n. In particular, this holds if we take γ1=γ2=γn+3=1 and

    γ23γ4<B1,γj+1γ2j+2γj+3<Bj, for j=2,,n1,γn+1γ1n+2<Bn. (39)

    If we call λ0=γ3 and λj=γj+3γj+2 for j=1,,n1, the inequalities in (39) are equivalent to the conditions (7). Then, if λj, j=0,,n1, satisfy these inequalities, the rescaling of the given parameters konj by λjkonj for j=0,,n1, gives rise to a system with exactly 2n2+1 positive steady states.

    The proof of Theorem 4 is similar to the previous one.

    Proof of Theorem 4. Again, the number of positive steady states of our system is equal to the number of positive solutions of the system (14). Recall that the support of the system is

    A={(1,0),(0,1),(1,1),(1,2),,(1,n),(0,0)}Z2.

    In this case, the coefficient matrix C (16) is equal, after multiplying the columns by convenient positive numbers, to the matrix

    Csimple=(10M0Mn1Stot01D0Dn1Etot),

    where Mi=kcatiνiFtot+1 and Di=1, for each iJ, and Mi=1 and Di=0, for each iJ.

    We consider again the regular triangulation Γ in Lemma 14. Recall that JJn, see (8), and therefore each jJ has the same parity as n, in particular 0jn2. For each jJ, consider the simplices Δj={(1,j),(1,j+1),(0,0)} and Δj+1={(1,j+1),(1,j+2),(0,0)}. Note that if jj then {Δj,Δj+1} and {Δj,Δj+1} are disjoint since j,j and n have the same parity.

    The simplices are positively decorated by Csimple (and then by C) if and only if the submatrices

    Csimplej=(1MjStot01Etot),Csimplej+1=(Mj1Stot10Etot),

    are positively spanning, and this happens if and only if EtotMjStot<0, where Mj=kcatjνjFtot+1, since jJ. The simplex Δn={(0,1),(1,n),(0,0)} is trivially positively decorated by Csimple. Then, by imposing the inequalities EtotMjStot<0 for jJ, which are equivalent to the ones in the statement (9), we can obtain 2|J|+1 positively decorated simplices.

    Assume (9) holds. Given any height function h inducing the triangulation Γ, by item (1) in Theorem 13 there exists t0 in R>0, such that for all 0<t<t0, the number of positive nondegenerate solutions of the deformed system with support A and coefficient matrix Ct, with (Ct)ij=th(αj)cij (with αjA, C=(cij)) is at least 2|J|+1. In particular if we choose h as in the proof of Lemma 14, there exists t0 in R>0, such that for all 0<t<t0, the system

    s0+jJ(Tj(Ftot)j+1+KjTj1(Ftot)j)tj(j+1)2s0ej+1+jJTj(Ftot)j+1tj(j+1)2s0ej+1Stot=0,tne+jJKjTj1(Ftot)jtj(j+1)2s0ej+1Etot=0, (40)

    has at least 2|J|+1 positive solutions. If we change the variable ˉe=tne, we get the following system:

    s0+jJ(Tj(Ftot)j+1+KjTj1(Ftot)j)t(j+1)(j2n)s0ˉej+1+jJTj(Ftot)j+1t(j+1)(j2n)s0ˉej+1Stot=0,ˉe+jJKjTj1(Ftot)jt(j+1)(j2n)s0ˉej+1Etot=0. (41)

    Similarly as we did in the previous proof, if we scale the original parameters konj, for jJ, and τj if jJ by

    tjnkonj if jJ,tjnτj if jJ, (42)

    respectively, and if we keep fixed the values of the remaining rate constants and the values of the total conservation constants Etot, Ftot and Stot, the intersection of the steady state variety and the linear conservation relations is described by system (41). Then, for these choices of constants the system GJ has at least 2|J|+1 positive steady states. The general rescaling that appears in the statement can be obtained in a similar way as we did in the proof of Theorem 2.

    Remark 15. If some or all rate constants koffi are zero (so, if the corresponding reactions are not assumed to happen), Theorems 2 and 4 hold. The proofs are very similar, except that if koffi=0, the constant Ki is equal to Ki=konikcati.

    In this section we explore a computational approach to the multistationarity problem, more precisely we find new regions of mulstistationarity. The idea is to find good regular triangulations of the point configuration corresponding to the exponents of the polynomial system given by the steady state equations and conservation laws, and deduce from them some regions of multistationarity. We first give the idea and then apply it for the n-site phosphorylation system for n=2,3,4, and 5, where we have successfully found several regions of multistationarity. This approach can be, in principle, applied to other systems if they satisfy certain hypotheses, see [13,Theorem 5.4], and are sufficiently small in order for the computations to be done in a reasonable amount of time.

    The strategy is the following. Given a polynomial system with support A and matrix of coefficients C, one first computes all possible regular triangulations of A with the aid of a computer. The number of such triangulations can be very large depending on the size of A, thus the next step is to discard in each such triangulation the simplices that obviously will not be positively decorated by C. With the reduced number of triangulations one can now search through all of them for the ones giving the biggest number of simultaneously positively decorated simplices. Each set of k simultaneously positively decorated simplices gives a candidate for a region of multistationarity with k positive steady states. If one finds m of such sets, then it is possible to have up to m such regions. Have in mind, however, that among these regions can be repetitions.

    Next we apply this to the n-site phosphorylation system with all intermediates and explain more concretely this procedure in this case.

    In Corollary 5 we obtained regions of multistationarity with 2n2+1 positive steady states each using only 1/4 of the intermediates, our objective now is to understand if it is possible to find more such regions with more intermediates. Consider the network G of the n-site phosphorylation system with all possible intermediates:

    S0+Ekon0koff0Y0kcat0S1+E  Sn1+Ekonn-1koffn-1Yn1kcatn1Sn+ESn+Fonn-1offn-1Un1catn1Sn1+F  S1+Fon0off0U0cat0S0+F

    In Section 4 of [13], the concentration at steady state of all species are given in terms of the species s0,e,f:

    si=Ti1s0eifi,i=1,,n,yi=KiTi1s0ei+1fi,i=0,,n1,ui=LiTis0ei+1fi,i=0,,n1,

    where Ki=konikoffi+kcati,Li=onioffi+cati,Ti=ij=0τjνj for each i=0,,n1 (recall that K1i and L1i are usually called Michaelis-Menten constants) and T1=1, where τi=kcatiKi and νi=catiLi, for each i=0,,n1.

    This looks very similar to (13), where Ftot is replaced by f, which is now a variable, and so we need to work in dimension 3. The main difference is that the networks GJ considered in the previous sections have intermediates only in the E component and the network G we consider in this section has all intermediates.

    Note from (2) that the support A of this system, which has 2n+4 elements, ordering the variables as s0,e,f, is given by the columns of the following matrix

    A=(1001111100101n12n00011n011n0),

    and the corresponding matrix of coefficients for the system is

    C=(100T0Tn1K0+L0T0K1T0+L1T1Kn1Tn2+Ln1Tn1Stot01000K0K1T0Kn1Tn2Etot00100L0T0L1T1Ln1Tn1Ftot).

    Recall that if one multiplies a column of a matrix C by a positive number, then a simplex is positively decorated by C if and only if it is positively decorated by the modified matrix. So, in order to test whether a simplex with vertices in A is positively decorated by C is enough to test if it is positively decorated by the following matrix

    Csimple=(10011111Stot01000N0N1Nn1Etot001001N01N11Nn1Ftot),

    where 0<Ni=KiTi1KiTi1+LiTi=(1+kcatilcati)1<1 for i=0,1,,n1. Here the matrix Csimple is obtained by dividing the fourth until the last column by its first entry.

    Now we compute all possible regular triangulations of A and search through them looking for the ones with the maximal possible number of simplices simultaneously positively decorated by Csimple. Since the number of such triangulations grows very fast with n we approach it with the following strategy:

    Algorithm 16. 1. Compute L1:={all possible triangulations of A}.*

    *We are calling L1,,L7 the sets defined in Algorithm 16. They are completely unrelated to the rational functions of the rate constants denoted with the same letters.

    2. With L1 compute L2 by discarding all simplices which do not have the last vertex (0,0,0). In fact we only need these simplices since a simplex not containing the last vertex cannot be positively decorated, because the corresponding coefficients of Csimple will be all positive.

    3. Compute L3 from L2 by removing all simplices with the corresponding 3×4 submatrix of Csimple having a zero 3×3 minor. The reason for this is clear, such simplices will never be positively decorated by Csimple.

    4. Compute L4 from L3 using the symmetry of Csimple. More precisely, change any index 4,5,,n+3 on the simplex to 1 because on Csimple they yield the same column. Here we are using the easy-to-check fact that changing the order of indexes does not change the conditions for a simplex to be positively decorated.

    5. Compute L5 from L4 removing all TL4 such that there is another TL4 with TT.

    6. For each TL5 and each set ST of simplices check if there is viable N0,,Nn1 such that all ΔS are positively decorated by Csimple at the same time. Call L6 the list of such Ss.

    7. If the maximum size of an element in L6 is k, set L7:={TL6:#T=k}. This k is the number of positive steady states and m:=#L7 is the number of candidates for regions of multistationarity.

    Step (1) can be done with the package TOPCOM inside SAGE [22], the other steps are quite simple to implement, for instance in MAPLE [23]. We show in Table 1 the number of elements in some of the lists and an approximation of the computation time for small values of n.

    Table 1.  Computation time and size of some lists while running Algorithm 16.
    n #L1 #L2 #L3 #L4 #L5 #L7 k regions of multistationarity computation time
    2 44 25 15 7 6 1 3 1 negligible
    3 649 260 100 21 18 6 3 6 about 1 sec
    4 9094 2728 682 62 53 5 5 4 about 2 min
    5 122835 28044 4560 177 149 23 5 15 about 3 hours

     | Show Table
    DownLoad: CSV

    The most computationally expensive part is to compute all regular triangulations, taking at least 90% of the time. These computations were done in a Linux virtual machine with 4 MB of RAM and with 4 cores of 3.2 GHz of processing. With a faster computer or more time one probably can do n=6 or even n=7 but probably not much more than this. For n=5 just the file for the raw list L1 of regular triangulations already has 10Mb.

    An alternative path to Steps (6) and (7) is to set a number k and look for sets TL5 and ST with #Sk such that there is viable N0,,Nn1 such that all ΔS are positively decorated by Csimple at the same time. We actually used this with k=2n2+1. This other route depends upon a good guess one may previously have at how many positive steady states to expect.

    After Step (7) one has to determine if there are any repetitions among the candidates for regions of multistationarity in L7 and also if there are any superfluous candidates of regions, that is conditions C1 and C2 such that C1 implies C2. In our case we did it by hand since the #L7 was quite small.

    Once Step (7) is done, one has a list of inequalities for each element S of L7. These come from the conditions imposing that the simplices in S are positively decorated by Csimple. We are going to use these conditions to describe the regions of mulstistationarity. Because of the uniformity of Csimple the only kind of conditions that appear are

    (I)i,j NiNj>0

    (II)i StotNiEtot>0

    (III)i EtotNi+FtotNiEtot>0

    (IV)i StotNiFtot+Stot>0

    (V) Stot>Etot+Ftot,

    or the opposite inequalities, and these translate from the Ni to the kcati,cati as follows

    (I)i,j kcatjcatj>kcaticati

    (II)i StotEtotEtot>kcaticati

    (III)i FtotEtot>kcaticati

    (IV)i FtotStotFtot<kcaticati.

    Note that

    ● conditions (III)i and (V) together imply (II)i;

    ● the opposite of condition (II)i together with condition (V) imply the opposite of (III)i;

    ● the opposite of condition (III)i together with condition (V) imply (IV)i;

    ● the opposite of condition (IV)i together with condition (V) imply (III)i;

    ● condition (III)i and the opposite of (III)j together imply (I)i,j.

    Using these properties it is easy to describe in a nice manner the regions of multistationarity and discard the repeated and superfluous ones. We sum up our findings on the following results which are proved in the same fashion as Theorems 2 and 4, once you have the regular triangulation obtained with the computer script. In the following propositions we describe the regions of multistationarity for n=2,3,4 and 5. Note that we are not describing the precise rescalings; they are potentially different for different rate constants and similar to those in (7) and (10).

    Proposition 17. Let n=2. Assume that Stot>Etot+Ftot. Then there is a choice of reaction rate constants for which the distributive sequential 2-site phosphorylation system admits 3 positive steady states. More explicitly, given rate constants and total concentrations such that

    kcat0cat0<FtotEtot<kcat1cat1, (43)

    after rescaling of the kon's and on's the distributive sequential 2-site phosphorylation system has 3 positive steady states.

    Proposition 18. Let n=3. Assume that Stot>Etot+Ftot. Then, there is a choice of rate constants for which the distributive sequential 3-site phosphorylation system admits at least 3 positive steady states. More explicitly, if the rate constants and total concentrations are in one of the regions described below

    (R3.1) kcat0cat0<FtotEtot<kcat1cat1,

    (R3.2) kcat0cat0<FtotEtot<kcat2cat2,

    (R3.3) kcat1cat1<FtotEtot<kcat2cat2,

    then after rescaling of the kon's and on's the distributive sequential 3-site phosphorylation system has at least 3 positive steady states.

    Proposition 19. Let n=3. If the rate constants and total concentrations are in one of the regions described below

    (R3.4) max{FtotEtot,FtotStotFtot}<min{kcat0cat0,kcat2cat2}, Stot>Ftot,

    (R3.5) max{FtotEtot,FtotStotFtot}<min{kcat1cat1,kcat2cat2}, Stot>Ftot,

    (R3.6) min{FtotEtot,StotEtotEtot}>max{kcat1cat1,kcat2cat2}, Stot>Etot,

    then after rescaling of the kon's and on's the distributive sequential 3-site phosphorylation system has at least 3 positive steady states.

    Proposition 20. Let n=4. Assume that Stot>Etot+Ftot. Then, there is a choice of rate constants for which the distributive sequential 4-site phosphorylation system has at least 5 steady states. More explicitly, if the rate constants and total concentrations are in one of the regions described below

    (R4.1) kcat2cat2<FtotEtot<min{kcat1cat1,kcat3cat3},

    (R4.2) kcat0cat0<FtotEtot<min{kcat1cat1,kcat3cat3},

    (R4.3) max{kcat0cat0,kcat2cat2}<FtotEtot<kcat3cat3,

    (R4.4) max{kcat0cat0,kcat2cat2}<FtotEtot<kcat1cat1,

    then after rescaling of the kon's and on's the distributive sequential 4-site phosphorylation system has at least 5 steady states.

    Proposition 21. Let n=5. Assume that Stot>Etot+Ftot. Then, there is a choice of rate constants for which the distributive sequential 5-site phosphorylation system has at least 5 steady states. More explicitly, if the rate constants and total concentrations are in one of the 13 regions described below

    (R5.(I,J))maxiI{kcaticati}<FtotEtot<minjJ{kcatjcatj},

    with (I,J) in the following list (where we write e.g. 14 instead of {1,4}):

    (0,14),(0,24),(1,24),(2,13),(2,14),(3,14),(3,024),(02,3),(02,4),(03,1),(03,2),(13,2),(13,4),

    then after rescaling of the kon's and on's the distributive sequential 5-site phosphorylation system has at least 5 steady states.

    Proposition 22. Let n=5. If the rate constants and total concentrations are in one of the regions described below

    (R5.1) max{FtotEtot,FtotStotFtot}<min{kcat1cat1,kcat2cat2,kcat4cat4}, Stot>Ftot,

    (R5.2) min{FtotEtot,StotEtotEtot}>max{kcat0cat0,kcat2cat2,kcat3cat3}, Stot>Etot,

    then after rescaling of the kon's and on's the distributive sequential 5-site phosphorylation system has at least 5 positive steady states.

    Note that the conditions in this section describe different regions from the ones described by the inequalities in Theorem 2 and Theorem 4. For instance, in Propositions 17, 18, 20, 21 the inequalities between the reaction rate constants and total conservations constants do not involve the value of Stot. In Propositions 19 and 22, the conditions onse the total conservation constants are also different (e.g. on FtotEtot and StotEtot1 instead of the product Ftot(StotEtot1)). The inequalities in Theorem 2 and Theorem 4 hold for reactions rate constants of a reduced system GJ, but if we use Theorem 10 to extrapolate these conditions to the full n-site phosphorylation network, the regions are different as well.

    The problem of finding multistationarity regions for chemical reaction networks is in theory effectively computable but it is both hard to compute and to describe such regions in a useful fashion. It is well known than even for small systems, where multistationarity regions can be explicitly found, the expressions are very complex. So, partial approaches that give sufficient conditions on a subset of parameters are useful and relevant.

    We developed in this paper both the theoretical setting based on [13,19] and the algorithmic approach that follows from it, to describe multistationarity regions in the space of all parameters for subnetworks of the n-site sequential phosphorylation cycle, where there are up to 2n2+1 positive steady states with fixed total conservation constants. We chose to assume that the subnetworks we consider have intermediate species only in the E component, but of course there is a symmetry in the network interchanging E with F, each Si with Sni, the corresponding intermediates and rate constants, and completely similar results hold if we assume that there are only intermediates in the F component. These regions can be lifted to regions of multistationarity with at least these number of stoichiometrically compatible positive steady states of the full n-site sequential phosphorylation cycle. The main feature of our polyhedral approach is that it gives a systematic method applicable to other networks (for instance, to enzymatic cascades [20]), and it provides open conditions and not only choices of parameter values where high multistationarity occurs. Moreover, we show that it can be algorithmically implemented. However, as we have already remarked in the Introduction, the inequalities on some of the reaction rate constants are not completely explicit. Instead, they point out "directions" in which these parameters have to be scaled.

    Theorem 1 in [4] states that for any positive values of Stot, Etot and Ftot, there exists ε0>0 such that if EtotStot<ε0, then there exists a choice of rate constants such that the system admits 2n2+1 positive steady states. Also, in the recent paper [12], the authors prove the existence of parameters for which there are 2n2+1 positive steady states and nearly half of them are stable and the other half unstable, using a different degeneration of the full network. However, the maximum expected number of stoichiometrically compatible steady states for the n-site system is equal to 2n1 (this is an upper bound by [4]), which has been verified for n=3 and 4 in [8]. Probably, it is not possible to find a region in parameter space with this number of positive steady states using degeneration techniques.

    In [7] and [6] it is proved that if kcat0/cat0<kcat1/cat1 then there exist constants Etot,Ftot,Stot such that the distributive sequential 2-site phosphorylation presents multistationarity. In [24] the author gives new regions of multistationarity for n=2, more precisely it is shown in Theorem 1 that given parameters K1,L0,L1,kcat0,kcat1,cat0,cat1, for any K0 small enough there exist constants Etot,Ftot,Stot such that the distributive sequential 2-site phosphorylation presents multistationarity. How small K0 has to be taken is explicitly given in terms of the other parameters by a rather involved equation (a similar result is obtained interchanging K0 and L1). Our Proposition 17 gives a similar result.

    Proposition 17 is in agreement with [10,Corollary 4.11] which establishes that in order for the distributive sequential 2-site phosphorylation to present multistationarity the total concentration of the substrate needs to be larger than either the concentration of the kinase or the phosphatase (Stot>Ftot or Stot>Etot). In the regions of multistationarity we found for n=3,4 and 5 this is the case as well. Propositions 17, 18, 19, 20, 21 and 22 are of the same flavor as Theorems 2 and 4, in the sense that all of them give sufficient conditions on the rate constants and total conservation constants such that after rescaling of the kon's and on's, the distributive sequential n-site phosphorylation system is multistationary. Note that most of the conditions for multistationarity we found can be given a biochemical interpretation in terms of the different Michaelis-Menten maximal velocities Vimax(E)=kcatiEtot,Vimax(F)=catiFtot for each independent phosphorylation/dephosphorylation step [25]. For instance, inequalities (43) in Proposition 17 can be written in classical biochemical terms as:

    V0max(E)<V0max(F),V1max(F)<V1max(E).

    The computational approach described in Section 5 can be used for other networks of moderate size.

    In conclusion, to the best of our knowledge other results (for instance [6], [7] and [24]) only give precise regions of multistationarity for the distributive sequential 2-site phosphorylation system. The present work gives new regions of multistationarity for the distributive sequential n-site phosphorylation system for any n, providing lower bounds in the number of positive steady states in such regions which are greater than three when n>3. For small n we give several distinct regions of multistationarity computed via a Computer Algebra System basic implementation. Moreover, our approach is easily applicable to other networks, both analytically and computationally.

    We are very thankful to Eugenia Ellis and Andrea Solotar for organizing the excellent project "Matemáticas en el Cono Sur", which lead to this work. We also thank Eugenia Ellis for the warm hospitality in Montevideo, Uruguay, on December 3–7, 2018, where we devised our main results. We thank the referees and Jeremy Gunawardena for their useful comments that helped improve the manuscript. AD, MG and MPM were partially supported by UBACYT 20020170100048BA, CONICET PIP 11220150100473 and 11220150100483, and ANPCyT PICT 2016-0398, Argentina.

    The authors declare no conflict of interest in this paper.



    [1] R. Singh, S. V. S. Deo, E. Dhamija, S. Mathur, S. Thulkar, To Evaluate the Accuracy of Axillary Staging Using Ultrasound and Ultrasound-Guided Fine-Needle Aspiration Cytology (USG-FNAC) in Early Breast Cancer Patients-a Prospective Study, Indian J. Surg. Oncol., 11 (2020), 726-734. doi: 10.1007/s13193-020-01222-3
    [2] I. González-Huebra, A. Elizalde, A. García-Baizán, M. Calvo, A. Ezponda, F. Martínez-Regueira, et al., Is it worth to perform preoperative MRI for breast cancer after mammography, tomosynthesis and ultrasound?, Magn. Reson. Imag., 57 (2019), 317-322. doi: 10.1016/j.mri.2018.12.005
    [3] Y. Huang, F. Li, J. Han, C. Peng, Q. Li, L. Cao, et al., Shear wave elastography of breast lesions: quantitative analysis of elastic heterogeneity improves diagnostic performance, Ultrasound Med. Biol., 45 (2019), 1909-1917. doi: 10.1016/j.ultrasmedbio.2019.04.019
    [4] M. Di Segni, V. de Soccio, V. Cantisani, G. Bonito, A. Rubini, G. Di Segni, et al., Automated classification of focal breast lesions according to S-detect: validation and role as a clinical and teaching tool, J. Ultrasound, 21 (2018), 105-118. doi: 10.1007/s40477-018-0297-2
    [5] T. V. Bartolotta, A. A. M. Orlando, L. Spatafora, M. Dimarco, C. Gagliardo, A. Taibbi, S-Detect characterization of focal breast lesions according to the US BI RADS lexicon: a pictorial essay, J. Ultrasound, 23 (2020), 207-215. doi: 10.1007/s40477-020-00447-w
    [6] T. V. Bartolotta, A. A. M. Orlando, M. L. Di Vittorio, F. Amato, M. Dimarco, D. Matranga, et al., S-Detect characterization of focal solid breast lesions: a prospective analysis of inter-reader agreement for US BI-RADS descriptors, J. Ultrasound, 5 (2020), 1-8.
    [7] R. M. Jales, M. T. Dória, K. P. Serra, M. M. Miranda, C. A. Menossi, K. Schumacher, et al., Power Doppler Ultrasonography and Shear Wave Elastography as Complementary Imaging Methods for Suspected Local Breast Cancer Recurrence, J. Ultrasound Med., 37 (2018), 1493-1501. doi: 10.1002/jum.14493
    [8] R. Guo, G. Lu, B. Qin, B. Fei, Ultrasound Imaging Technologies for Breast Cancer Detection and Management: A Review, Ultrasound Med. Biol., 44 (2018), 37-70. doi: 10.1016/j.ultrasmedbio.2017.09.012
    [9] J. Geisel, M. Raghu, R. Hooley, The role of ultrasound in breast cancer screening: the case for and against ultrasound, Semin Ultrasound CT MR, 39 (2018), 25-34. doi: 10.1053/j.sult.2017.09.006
    [10] J. Y. Wu, Z. Z. Zhao, W. Y. Zhang, M. Liang, B. Ou, H. Y. Yang, et al., Computer-aided diagnosis of solid breast lesions with ultrasound: factors associated wth false-negative and false-positive results, J. Ultrasound Med., 38 (2019), 3193-3202. doi: 10.1002/jum.15020
    [11] A. S. Tagliafico, M. Piana, D. Schenone, R. Lai, A. M. Massone, N. Houssami, Overview of radiomics in breast cancer diagnosis and prognostication, Breast, 49 (2020), 74-80. doi: 10.1016/j.breast.2019.10.018
    [12] K. Kim, M. K. Song, E. K. Kim, J. H. Yoon, Clinical application of S-Detect to breast masses on ultrasonography: a study evaluating the diagnostic performance and agreement with a dedicated breast radiologist, Ultrasonography, 36 (2017), 3-9. doi: 10.14366/usg.16012
    [13] E. Cho, E. K. Kim, M. K. Song, J. H. Yoon, Application of computer-aided diagnosis on breast ultrasonography: evaluation of diagnostic performances and agreement of radiologists according to different levels of experience, J. Ultrasound Med., 37 (2018), 209-216. doi: 10.1002/jum.14332
    [14] J. H. Choi, B. J. Kang, J. E. Baek, H. S. Lee, S. H. Kim, Application of computer-aided diagnosis in breast ultrasound interpretation: improvements in diagnostic performance according to reader experience, Ultrasonography, 37 (2018): 217-225. doi: 10.14366/usg.17046
    [15] W. H. Yuan, A. F. Li, Y. H. Chou, H. C. Hsu, Y. Y. Chen, Clinical and ultrasonographic features of male breast tumors: A retrospective analysis, PLoS One, 13 (2018), e0194651. doi: 10.1371/journal.pone.0194651
    [16] S. H. Yeo, G. R. Kim, S. H. Lee, W. K. Moon, Comparison of ultrasound elastography and color doppler ultrasonography for distinguishing small triple-negative nreast cancer from fibroadenoma, J. Ultrasound Med., 37 (2018), 2135-2146. doi: 10.1002/jum.14564
    [17] A. Vourtsis, Three-dimensional automated breast ultrasound: Technical aspects and first results, Diagn. Interv. Imaging, 100 (2019), 579-592. doi: 10.1016/j.diii.2019.03.012
    [18] D. A. Spak, J. S. Plaxco, L. Santiago, M. J. Dryden, B. E. Dogan, BI-RADS(®) fifth edition: A summary of changes, Diagn. Interv. Imaging, 98 (2017), 179-190. doi: 10.1016/j.diii.2017.01.001
    [19] X. X. Qu, Y. Song, Y. H. Zhang, H. M. Qing, Value of ultrasonic elastography and conventional ultrasonography in the differential diagnosis of non-mass-like breast lesions, Ultrasound Med. Biol., 45 (2019), 1358-1366. doi: 10.1016/j.ultrasmedbio.2019.01.020
    [20] B. L. Niell, P. E. Freer, R. J. Weinfurtner, E. K. Arleo, J. S. Drukteinis, Screening for breast cancer, Radiol. Clin. North Am., 55 (2017), 1145-1162. doi: 10.1016/j.rcl.2017.06.004
    [21] F. Valdora, N. Houssami, F. Rossi, M. Calabrese, A. S. Tagliafico, Rapid review: radiomics and breast cancer, Breast Cancer Res. Treat., 169 (2018), 217-229. doi: 10.1007/s10549-018-4675-4
    [22] F. Taskin, Y. Durum, A. Soyder, A. Unsal, Review and management of breast lesions detected with breast tomosynthesis but not visible on mammography and ultrasonography, Acta. Radiol., 58 (2017), 1442-1447. doi: 10.1177/0284185117710681
  • This article has been cited by:

    1. Peng Xue, Mingyu Si, Dongxu Qin, Bingrui Wei, Samuel Seery, Zichen Ye, Mingyang Chen, Sumeng Wang, Cheng Song, Bo Zhang, Ming Ding, Wenling Zhang, Anying Bai, Huijiao Yan, Le Dang, Yuqian Zhao, Remila Rezhake, Shaokai Zhang, Youlin Qiao, Yimin Qu, Yu Jiang, Unassisted Clinicians Versus Deep Learning–Assisted Clinicians in Image-Based Cancer Diagnostics: Systematic Review With Meta-analysis, 2023, 25, 1438-8871, e43832, 10.2196/43832
    2. Anastasia-Maria Leventi-Peetz, Kai Weber, Probabilistic machine learning for breast cancer classification, 2022, 20, 1551-0018, 624, 10.3934/mbe.2023029
    3. Peng-fei Lyu, Yu Wang, Qing-Xiang Meng, Ping-ming Fan, Ke Ma, Sha Xiao, Xun-chen Cao, Guang-Xun Lin, Si-yuan Dong, Mapping intellectual structures and research hotspots in the application of artificial intelligence in cancer: A bibliometric analysis, 2022, 12, 2234-943X, 10.3389/fonc.2022.955668
    4. Xiaolei Wang, Shuang Meng, Diagnostic accuracy of S-Detect to breast cancer on ultrasonography: A meta-analysis (PRISMA), 2022, 101, 1536-5964, e30359, 10.1097/MD.0000000000030359
    5. Samara Acosta-Jiménez, Javier Camarillo-Cisneros, Abimael Guzmán-Pando, Susana Aideé González-Chávez, Jorge Issac Galván-Tejada, Graciela Ramírez-Alonso, César Francisco Pacheco-Tena, Rosa Elena Ochoa-Albiztegui, 2023, Chapter 8, 978-3-031-18255-6, 83, 10.1007/978-3-031-18256-3_8
    6. Yuqun Wang, Lei Tang, Pingping Chen, Man Chen, The Role of a Deep Learning-Based Computer-Aided Diagnosis System and Elastography in Reducing Unnecessary Breast Lesion Biopsies, 2022, 15268209, 10.1016/j.clbc.2022.12.016
    7. Peizhen Huang, Bin Zheng, Mengyi Li, Lin Xu, Sajjad Rabbani, Abdulilah Mohammad Mayet, Chengchun Chen, Beishu Zhan, He Jun, Ateeq Ur Rehman, The Diagnostic Value of Artificial Intelligence Ultrasound S-Detect Technology for Thyroid Nodules, 2022, 2022, 1687-5273, 1, 10.1155/2022/3656572
    8. Qiyu Liu, Meijing Qu, Lipeng Sun, Hui Wang, Accuracy of ultrasonic artificial intelligence in diagnosing benign and malignant breast diseases, 2021, 100, 0025-7974, e28289, 10.1097/MD.0000000000028289
    9. N.N. Bayandina, E.N. Slavnova, A cytological method in the early diagnosis of cervical cancer: evolution, principles, technologies, prospects, 2023, 12, 2305-218X, 49, 10.17116/onkolog20231202149
    10. Peng Sun, 2023, Chapter 35, 978-981-19-9372-5, 323, 10.1007/978-981-19-9373-2_35
    11. Hee Jeong Kim, Hak Hee Kim, Ki Hwan Kim, Ji Sung Lee, Woo Jung Choi, Eun Young Chae, Hee Jung Shin, Joo Hee Cha, Woo Hyun Shim, Use of a commercial artificial intelligence-based mammography analysis software for improving breast ultrasound interpretations, 2024, 34, 1432-1084, 6320, 10.1007/s00330-024-10718-3
    12. E. L. Teodozova, E. Yu. Khomutova, Artificial intelligence in radial diagnostics of breast cancer, 2023, 3, 2782-3024, 26, 10.61634/2782-3024-2023-12-26-35
    13. Xiao Zou, Jintao Zhai, Shengyou Qian, Ang Li, Feng Tian, Xiaofei Cao, Runmin Wang, Improved breast ultrasound tumor classification using dual-input CNN with GAP-guided attention loss, 2023, 20, 1551-0018, 15244, 10.3934/mbe.2023682
    14. Lu Li, Hongyan Deng, Xinhua Ye, Yong Li, Jie Wang, Comparison of the diagnostic efficacy of mathematical models in distinguishing ultrasound imaging of breast nodules, 2023, 13, 2045-2322, 10.1038/s41598-023-42937-x
    15. Na Li, Wanling Liu, Yunyun Zhan, Yu Bi, Xiabi Wu, Mei Peng, Value of S-Detect combined with multimodal ultrasound in differentiating malignant from benign breast masses, 2024, 55, 2090-4762, 10.1186/s43055-023-01183-x
    16. Pengjie Song, Li Zhang, Longmei Bai, Qing Wang, Yanlei Wang, Diagnostic performance of ultrasound with computer-aided diagnostic system in detecting breast cancer, 2023, 9, 24058440, e20712, 10.1016/j.heliyon.2023.e20712
    17. Qing Dan, Ziting Xu, Hannah Burrows, Jennifer Bissram, Jeffrey S. A. Stringer, Yingjia Li, Diagnostic performance of deep learning in ultrasound diagnosis of breast cancer: a systematic review, 2024, 8, 2397-768X, 10.1038/s41698-024-00514-z
    18. Panpan Zhang, Min Zhang, Menglin Lu, Chaoying Jin, Gang Wang, Xianfang Lin, Comparative Analysis of the Diagnostic Value of S-Detect Technology in Different Planes Versus the BI-RADS Classification for Breast Lesions, 2024, 10766332, 10.1016/j.acra.2024.08.005
    19. Zhuohua Lin, Ligang Cui, Yan Xu, Qiang Fu, Youjing Sun, Feasibility and potential of intraoperative ultrasound in arthroscopy of femoroacetabular impingement, 2024, 2054-8397, 10.1093/jhps/hnad050
    20. Manisha Bahl, Jung Min Chang, Lisa A. Mullen, Wendie A. Berg, Artificial Intelligence for Breast Ultrasound: AJR Expert Panel Narrative Review, 2024, 0361-803X, 1, 10.2214/AJR.23.30645
    21. Jie He, Nan Liu, Li Zhao, New progress in imaging diagnosis and immunotherapy of breast cancer, 2025, 16, 1664-3224, 10.3389/fimmu.2025.1560257
  • Reader Comments
  • © 2021 the Author(s), licensee AIMS Press. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0)
通讯作者: 陈斌, bchen63@163.com
  • 1. 

    沈阳化工大学材料科学与工程学院 沈阳 110142

  1. 本站搜索
  2. 百度学术搜索
  3. 万方数据库搜索
  4. CNKI搜索

Metrics

Article views(6294) PDF downloads(445) Cited by(21)

Figures and Tables

Figures(5)  /  Tables(3)

/

DownLoad:  Full-Size Img  PowerPoint
Return
Return

Catalog