Processing math: 47%
Research article Special Issues

mfeeU-Net: A multi-scale feature extraction and enhancement U-Net for automatic liver segmentation from CT Images


  • Medical image segmentation of the liver is an important prerequisite for clinical diagnosis and evaluation of liver cancer. For automatic liver segmentation from Computed Tomography (CT) images, we proposed a Multi-scale Feature Extraction and Enhancement U-Net (mfeeU-Net), incorporating Res2Net blocks, Squeeze-and-Excitation (SE) blocks, and Edge Attention (EA) blocks. The Res2Net blocks which are conducive to extracting multi-scale features of the liver were used as the backbone of the encoder, while the SE blocks were also added to the encoder to enhance channel information. The EA blocks were introduced to skip connections between the encoder and the decoder, to facilitate the detection of blurred liver edges where the intensities of nearby organs are close to the liver. The proposed mfeeU-Net was trained and evaluated using a publicly available CT dataset of LiTS2017. The average dice similarity coefficient, intersection-over-union ratio, and sensitivity of the mfeeU-Net for liver segmentation were 95.32%, 91.67%, and 95.53%, respectively, and all these metrics were better than those of U-Net, Res-U-Net, and Attention U-Net. The experimental results demonstrate that the mfeeU-Net can compete with and even outperform recently proposed convolutional neural networks and effectively overcome challenges, such as discontinuous liver regions and fuzzy liver boundaries.

    Citation: Jun Liu, Zhenhua Yan, Chaochao Zhou, Liren Shao, Yuanyuan Han, Yusheng Song. mfeeU-Net: A multi-scale feature extraction and enhancement U-Net for automatic liver segmentation from CT Images[J]. Mathematical Biosciences and Engineering, 2023, 20(5): 7784-7801. doi: 10.3934/mbe.2023336

    Related Papers:

    [1] Waleed Mohamed Abd-Elhameed, Omar Mazen Alqubori . On generalized Hermite polynomials. AIMS Mathematics, 2024, 9(11): 32463-32490. doi: 10.3934/math.20241556
    [2] Mohra Zayed, Shahid Wani . Exploring the versatile properties and applications of multidimensional degenerate Hermite polynomials. AIMS Mathematics, 2023, 8(12): 30813-30826. doi: 10.3934/math.20231575
    [3] Mohra Zayed, Taghreed Alqurashi, Shahid Ahmad Wani, Cheon Seoung Ryoo, William Ramírez . Several characterizations of bivariate quantum-Hermite-Appell Polynomials and the structure of their zeros. AIMS Mathematics, 2025, 10(5): 11184-11207. doi: 10.3934/math.2025507
    [4] Gyung Won Hwang, Cheon Seoung Ryoo, Jung Yoog Kang . Some properties for 2-variable modified partially degenerate Hermite (MPDH) polynomials derived from differential equations and their zeros distributions. AIMS Mathematics, 2023, 8(12): 30591-30609. doi: 10.3934/math.20231564
    [5] Mohra Zayed, Shahid Ahmad Wani . Properties and applications of generalized 1-parameter 3-variable Hermite-based Appell polynomials. AIMS Mathematics, 2024, 9(9): 25145-25165. doi: 10.3934/math.20241226
    [6] Yunbo Tian, Sheng Chen . Prime decomposition of quadratic matrix polynomials. AIMS Mathematics, 2021, 6(9): 9911-9918. doi: 10.3934/math.2021576
    [7] Mohra Zayed, Shahid Ahmad Wani, William Ramírez, Clemente Cesarano . Advancements in q-Hermite-Appell polynomials: a three-dimensional exploration. AIMS Mathematics, 2024, 9(10): 26799-26824. doi: 10.3934/math.20241303
    [8] Mdi Begum Jeelani . On employing linear algebra approach to hybrid Sheffer polynomials. AIMS Mathematics, 2023, 8(1): 1871-1888. doi: 10.3934/math.2023096
    [9] Young Joon Ahn . An approximation method for convolution curves of regular curves and ellipses. AIMS Mathematics, 2024, 9(12): 34606-34617. doi: 10.3934/math.20241648
    [10] Nusrat Raza, Mohammed Fadel, Kottakkaran Sooppy Nisar, M. Zakarya . On 2-variable q-Hermite polynomials. AIMS Mathematics, 2021, 6(8): 8705-8727. doi: 10.3934/math.2021506
  • Medical image segmentation of the liver is an important prerequisite for clinical diagnosis and evaluation of liver cancer. For automatic liver segmentation from Computed Tomography (CT) images, we proposed a Multi-scale Feature Extraction and Enhancement U-Net (mfeeU-Net), incorporating Res2Net blocks, Squeeze-and-Excitation (SE) blocks, and Edge Attention (EA) blocks. The Res2Net blocks which are conducive to extracting multi-scale features of the liver were used as the backbone of the encoder, while the SE blocks were also added to the encoder to enhance channel information. The EA blocks were introduced to skip connections between the encoder and the decoder, to facilitate the detection of blurred liver edges where the intensities of nearby organs are close to the liver. The proposed mfeeU-Net was trained and evaluated using a publicly available CT dataset of LiTS2017. The average dice similarity coefficient, intersection-over-union ratio, and sensitivity of the mfeeU-Net for liver segmentation were 95.32%, 91.67%, and 95.53%, respectively, and all these metrics were better than those of U-Net, Res-U-Net, and Attention U-Net. The experimental results demonstrate that the mfeeU-Net can compete with and even outperform recently proposed convolutional neural networks and effectively overcome challenges, such as discontinuous liver regions and fuzzy liver boundaries.



    In this paper, we are concerned with the sharp decay rates of solutions to the Cauchy problem for the isentropic Navier-Stokes equations:

    {tρ+div(ρu)=0,(t,x)R+×R3,t(ρu)+div(ρuu)+p(ρ)=divT,(t,x)R+×R3,lim|x|ρ=ˉρ,lim|x|u=0,tR+,(ρ,u)|t=0=(ρ0,u0),xR3, (1.1)

    which governs the motion of a isentropic compressible viscous fluid. The unknown functions ρ and u represent the density and velocity of the fluid respectively. The pressure p=p(ρ) is a smooth function in a neighborhood of a positive constant ˉρ s.t. p(ˉρ)>0. T is the viscosity stress tensor given by T=μ(u+(u)t)+ν(divu)I with I the identity matrix. We assume that the constant viscosity coefficients μ>0 and ν satisfy ν+23μ>0. Throughout this article, by optimal time decay rate, we refer to the best possible decay rate in upper bound as many literatures, and the sharp time decay rate includes the best possible upper and lower bounds.

    Using the classical spectral method, the optimal time decay rate (upper bound) of the linearized equations of the isentropic Navier-Stokes equations are well known. One may then expect that the small solution of the nonlinear equations (1.1) have the same decay rate as the linear one. Our work is devoted to proving the sharp time decay rate (for both upper and lower bound) for the nonlinear system.

    In the case of one space dimension, Zeng [24] and Liu-Zeng [15] offered a detailed analysis of the solution to a class of hyperbolic-parabolic system through point-wise estimate, including the isentropic Navier-Stokes system. For multi-dimensional Navier-Stokes equations (and/or Navier-Stokes-Fourier system), the Hs global existence and time-decay rate of strong solutions with the initial perturbation small in HsL1 are obtained in whole space first by A. Matsumura and T. Nishida [17], [18]. When the small initial perturbation belongs to H3 only, using a weighted energy method, A. Matsumura [16] showed the time-decay rate (1+t)34 of upper bound in L-norm. Since then, there are concrete development on the upper bound time-decay estimates: the optimal Lp (with 2p) upper bound decay rate was proved by G. Ponce [19], combining the spectral analysis on linearized system and the energy method for small initial perturbation in L1. For the isentropic Navier-Stokes equations with artificial viscosity, D. Hoff and K. Zumbrun [6], [7] studied the Green's function and derived the Lp (1p) upper bound time decay rate of diffusive waves for the small initial perturbation belongs to HmL1 with m4. Liu and Wang [14] studied the point-wise estimates of the Green function of the linearized isentropic Navier-Stokes system in 3D and then analyzed the coupling of nonlinear diffusion waves, obtained the optimal (upper bound) decay rate. These results were further extended to the exterior problem [12], [11], or the half space problem [9], [10], [8]. Recently, Guo and Wang in [5] developed a new general energy method for proving the optimal (upper bound) time decay rates of the solutions to the dissipative equations in the whole space, using a family of scaled energy estimates with minimum derivative counts and interpolations among them without linear decay analysis.

    When additional external force is taken into account, the external force does affect the long time behavior of solutions. The upper bound of time decay rates were studied intensively, see for instance [1] and [2] on unbounded domain, [22], [23] on the convergence of the non-stationary flow to the corresponding steady flow when the initial date are small in H3L65, and [4], [3], on the optimal LpLq upper bound decay rates for potential forces.

    The main goal of current paper is to establish the sharp decay rate, on both upper and lower bounds, to the solutions of (1.1) using relatively simple energy method. We remark that similar results had been pursued by M. Schonbek [20], [21] for incompressible Navier-Stokes equations, and by Li, Matsumura-Zhang [13] for isentropic Navier-Stokes-Poisson system. Although they share the same spirit in obtaining the lower bound decay rates, the feature of the spectrum near zero exhibits quite different behaviors, leading to different analysis. For instance, we explored the elegant structure of the higher order nonlinear terms of Navier-Stokes, when choosing conservative variables: density and momentum. The conservative form of the sharp equations provided a natural derivative structure in these terms, leading to the possibility of a faster decay rate estimate. We will make a more detailed comparison later in this paper.

    Define n=ρˉρ, and let m=ρu=(n+ˉρ)u be the momentum. We rewrite (1.1) as

    {tn+divm=0,(t,x)R+×R3,tm+c2nˉμm(ˉμ+ˉν)divm=F,(t,x)R+×R3,lim|x|n=0,lim|x|m=0,tR+,(n,m)|t=0=(ρ0ˉρ,ρ0u0),xR3, (1.2)

    where ˉμ=μˉρ, ˉν=νˉρ, c=p(ˉρ)>0 is the sound speed, and

    F=div{mmn+ˉρ+ˉμ(nmn+ˉρ)}{(ˉμ+ˉν)div(nmn+ˉρ)+(p(n+ˉρ)p(ˉρ)c2n)}.

    It is this structure of F that plays an important role in our analysis.

    Our aim is to obtain a clear picture of the large time behavior of U=(n,m) in L2(R3) when U0=(ρ0ˉρ,ρ0u0) is sufficiently smooth and small. We introduce the following initial value problem of the linearized Navier-Stokes system corresponding to (1.2):

    {t˜n+div˜m=0,(t,x)R+×R3,t˜m+c2˜nˉμ˜m(ˉμ+ˉν)div˜m=0,(t,x)R+×R3,lim|x|˜n=0,lim|x|˜m=0,tR+,(˜n,˜m)|t=0=(ρ0ˉρ,ρ0u0),xR3, (1.3)

    where ˉμ=μˉρ, ˉν=νˉρ, c=p(ˉρ). It is known that the L2-norm of ˜U=(˜n,˜m) decays at the optimal upper bound rate (1+t)34 for generic small initial data, see for instance [18]. A detailed proof on the optimal lower and upper bound rate will be given in the section 3 of this paper. In section 4, we prove that (U˜U)(,t)L2 decays at a faster rate than ˜U(,t)L2, under some reasonable conditions on the initial data. Therefore, U(,t)L2 shares the sharp decay rate of (1+t)34.

    Notation. For ab, we mean that there is a uniform constant C, which may be different on different lines, such that aCb. And ab stands for ab and ba.

    We now state our main result.

    Theorem 1.1. Assume that (n0,m0)L1(R3)H3(R3), δ0=:(n0,m0)L1(R3)H3(R3) is sufficiently small, and

    R3(n0,m0)dx0, (1.4)

    then there is a unique global classical solution ˜U=(˜n,˜m)C([0,);H3(R3)) of the linearized system (1.3) satisfying for some positive constant C

    C1(1+t)34k2k˜n(t)L2(R3)C(1+t)34k2,k=0,1,2,3,C1(1+t)34k2k˜m(t)L2(R3)C(1+t)34k2,k=0,1,2,3,

    and the initial value problem (1.2) has a unique solution U=(n,m)C([0,);H3(R3)). Moreover, let nh=n˜n and mh=m˜m, then it holds that

    k(nh,mh)(t)L2(R3)δ20(1+t)54k2,k=0,1,2,3mh(t)L2(R3)δ20(1+t)114,3nh(t)L2(R3)δ0(1+t)74.

    As a consequence, there exists a positive constant C1 such that

    C11(1+t)34k2kn(t)L2(R3)C1(1+t)34k2,k=0,1,2,C11(1+t)34k2km(t)L2(R3)C1(1+t)34k2,k=0,1,2,3.

    Remark 1.1. We remark that this theorem is valid under the condition (1.4) which is important in the lower bound estimate to the linearized problem. When (1.4) fails, the decay rate of the linearized system (1.3) depends on the order of the degeneracy of moments. Assume (n0,m0)L1H3 and belong to certain appropriate weighted Lp spaces, similar situation happened also in the incompressible Navier-Stokes equations, c.f. [20], [21]. We also note that our condition (1.4) is weaker than those in most of previous results where the differentiability of Fourier transform of initial disturbance is required in general.

    Remark 1.2. In [13], Li, Matsumura-Zhang proved the lower bound decay rate of the linearized isentropic Navier-Stokes-Poisson system, they only require |ˆn0(ξ)|>c0>0 for |ξ|1 with c0 a constant due to the special structure of the spectrum from the help of the Poisson term. This condition is proposed in Fourier space, similar to (1.4) in some sense. In our case, the spectrum is different and the different structure leads to different sharp decay rates.

    In what follows, we will set n=ρˉρ, u=u0. We rewrite (1.1) in the perturbation form as

    {tn+ˉρdivu=ndivuun,tu+γˉρnˉμu(ˉμ+ˉν)divu=uuˉμf(n)u(ˉμ+ˉν)f(n)divug(n)n,lim|x|n=0,lim|x|u=0,(n,u)|t=0=(ρ0ˉρ,u0), (2.1)

    where ˉμ=μˉρ, ˉν=νˉρ, γ=p(ˉρ)ˉρ2, and the nonlinear functions f and g are defined by

    f(n):=nn+ˉρ,g(n):=p(n+ˉρ)n+ˉρp(ˉρ)ˉρ. (2.2)

    We assume that there exist a time of existence T>0 and sufficiently small δ>0, such that a priori estimate

    n(t)H3+u(t)H3δ, (2.3)

    holds for any t[0,T]. First of all, by (2.3) and Sobolev's inequality, we obtain that

    ˉρ2n+ˉρ2ˉρ.

    Hence, we immediately have

    |f(n)|,|g(n)|C|n|,|kf(n)|,|kg(n)|CkN+, (2.4)

    where f(n) and g(n) are nonlinear functions of n defined by (2.2).

    Next, we begin with the energy estimates including n and u themselves. The following results is essentially due to A. Matsumura and T. Nishida [17], [18].

    Theorem 2.1. Assume that (n0,u0)H3(R3), then there exists a constant δ0>0 such that if

    n0H3+u0H3δ0,

    then the problem (2.1) admits a unique global solution (n(t),u(t)) satisfying that for all t0,

    n(t)2H3+u(t)2H3+t0(n(τ)2H2+u(τ)2H3)dτC(n02H3+u02H3),

    where C is a positive constant independent of time.

    The proof of this theorem is divided into several subsections.

    For k=0, multiplying the first equation in (2.1) by γn and the second equation in (2.1) by u, summing up and then integrating the result over R3 by parts. By virtue of Hölder's inequality, Sobolev's inequality and the fact (2.4), we obtain that

    12ddtR3(γ|n|2+|u|2)dx+R3(ˉμ|u|2+(ˉμ+ˉν)|divu|2)dx=R3γ(ndivuun)n(uu+ˉμf(n)u+(ˉμ+ˉν)f(n)divu+g(n)n)udxnL3uL2nL6+(uL3uL2+nL3nL2)uL6+(uLnL2+nLuL2)uL2(nL3+uL3+nL+uL)(n2L2+u2L2). (2.5)

    Now for 1k3, applying k to (2.1) and then multiplying the first equation by γkn and the second equation by ku, summing up and integrating over R3. For k=1 we have

    12ddtR3(γ|n|2+|u|2)dx+R3(ˉμ|2u|2+(ˉμ+ˉν)|divu|2)dx(nL+uL+nL+uL)(n2L2+u2L2+2u2L2). (2.6)

    For k=2 we have

    12ddtR3(γ|2n|2+|2u|2)dx+R3(ˉμ|3u|2+(ˉμ+ˉν)|2divu|2)dx(nL+uL+nL+uL)(2n2L2+2u2L2+3u2L2). (2.7)

    For k=3 we have

    12ddtR3(γ|3n|2+|3u|2)dx+R3(ˉμ|4u|2+(ˉμ+ˉν)|3divu|2)dx(nL+uL+nL+uL)(3n2L2+3u2L2+4u2L2)+nL34uL22nL6+uL34uL22uL6+2nL3(3nL2+4uL2)2uL6. (2.8)

    Summing up the above estimates, noting that δ>0 is small, we obtain that

    ddt0k3(γkn2L2+ku2L2)+C11k4ku2L2C2δ1k3kn2L2. (2.9)

    For 0k2, applying k to the second equation in (2.1) and then multiplying by k+1n. The key idea is to integrate by parts in the t-variable and to use the continuity equation. Thus integrating the results by parts for both the t- and x-variables, we obtain for k=0 that

    ddtR3undx+γˉρR3|n|2dxu2L2+nL22uL2+(nL+uL)(n2L2+u2L2), (2.10)

    for k=1, we get

    ddtR3u2ndx+γˉρR3|2n|2dx2u2L2+2nL23uL2+((n,u)L+(n,u)L)×(n2L2+2n2L2+2u2L2), (2.11)

    and for k=2 we have

    ddtR32u3ndx+γˉρR3|3n|2dx3u2L2+3nL24uL2+((n,u)L+(n,u)L)×(2n2L2+2u2L2+3n2L2+3u2L2). (2.12)

    Plugging the above estimates, using the smallness of δ>0, we obtain that

    ddt0k2R3kuk+1ndx+C31k3kn2L2C41k4ku2L2. (2.13)

    Proof of Theorem 2.1. Multiplying (2.13) by 2C2δC3, adding it with (2.9), with the help of smallness of δ>0, we deduce that there exists a constant C5>0 such that

    ddt{0k3(γkn2L2+ku2L2)+2C2δC30k2R3kuk+1ndx}+C5{1k3kn2L2+1k4ku2L2}0. (2.14)

    Next, we define E(t) to be C15 times the expression under the time derivative in (2.14). Then we may write (2.14) as

    ddtE(t)+n(t)2H2+u(t)2H30. (2.15)

    Observe that since δ is small, then there exists a constant C6>0 such that

    C16(n(t)2H3+u(t)2H3)E(t)C6(n(t)2H3+u(t)2H3).

    Then integrating (2.15) directly in time, we get

    sup0tT(n(t)2H3+u(t)2H3)+C6T0(n(τ)2H2+u(τ)2H3)dτC26(n02H3+u02H3).

    Using a standard continuity argument along with classical local wellposedness theory, this closes the a priori assumption (2.3) if we assume n0H3+u0H3δ0 is sufficiently small. We can then extend the solution globally in time and complete the proof of Theorem 2.1.

    In this section, we consider the initial value problem for the linearized Navier-Stokes system

    {t˜n+div˜m=0,(t,x)R+×R3,t˜m+c2˜nˉμ˜m(ˉμ+ˉν)div˜m=0,(t,x)R+×R3,lim|x|˜n=0,lim|x|˜m=0,tR+,(˜n,˜m)|t=0=(ρ0ˉρ,ρ0u0),xR3, (3.1)

    where ˉμ=μˉρ, ˉν=νˉρ, c=p(ˉρ).

    In terms of the semigroup theory for evolutionary equations, the solution (˜n,˜m) of the linearized Navier-Stokes problem (3.1) can be expressed for ˜U=(˜n,˜m)t as

    ˜Ut=B˜U,t0,˜U(0)=˜U0,

    which gives rise to

    ˜U(t)=S(t)˜U0=etB˜U0,t0,

    where B is defined as

    B=(0divc2ˉμangle+(ˉμ+ˉν)div).

    What left is to analyze the differential operator B in terms of its Fourier expression A(ξ) and show the long time properties of the semigroup S(t). Applying the Fourier transform to system (3.1), we have

    tˆ˜U(t,ξ)=A(ξ)ˆ˜U(t,ξ),t0,ˆ˜U(0,ξ)=ˆ˜U0(ξ),

    where ξ=(ξ1,ξ2,ξ3)t, and A(ξ) is defined as

    A(ξ)=(0iξtc2iξˉμ|ξ|2I3×3(ˉμ+ˉν)ξξ).

    The eigenvalues of the matrix A can be computed by

    det(A(ξ)λI)=(λ+ˉμ|ξ|2)2(λ2+(2ˉμ+ˉν)|ξ|2λ+c2|ξ|2)=0,

    which implies

    λ0=ˉμ|ξ|2(double),λ1=λ1(|ξ|),λ2=λ2(|ξ|).

    The semigroup etA is expressed as

    etA=eλ0tP0+eλ1tP1+eλ2tP2,

    where the project operators Pi can be computed as

    Pi=ijA(ξ)λjIλiλj.

    By a direct computation, we can verify the exact expression for the Fourier transform ˆG(t,ξ) of Green's function G(t,x)=etB as

    ˆG(t,ξ)=etA=(λ1eλ2tλ2eλ1tλ1λ2iξt(eλ1teλ2t)λ1λ2c2iξ(eλ1teλ2t)λ1λ2eλ0t(Iξξ|ξ|2)+ξξ|ξ|2λ1eλ1tλ2eλ2tλ1λ2)=(ˆNˆM).

    Indeed, we can make the following decomposition for (˜n,˜m)=G˜U0 as

    ˆ˜n=ˆNˆ˜U0=(ˆN+ˆN)ˆ˜U0,ˆ˜m=ˆMˆ˜U0=(ˆM+ˆM)ˆ˜U0,

    where

    ˆN=(λ1eλ2tλ2eλ1tλ1λ20),ˆN=(0iξt(eλ1teλ2t)λ1λ2),ˆM=(c2iξ(eλ1teλ2t)λ1λ20),ˆM=(0eλ0t(Iξξ|ξ|2)+ξξ|ξ|2λ1eλ1tλ2eλ2tλ1λ2).

    We further decompose the Fourier transform ˆN, ˆM into low frequency term and high frequency term below.

    Define

    ˆN=ˆN1+ˆN2,ˆN=ˆN1+ˆN2,ˆM=ˆM1+ˆM2,ˆM=ˆM1+ˆM2,

    where ()1=χ(ξ)(), ()2=(1χ(ξ))(), and χ(ξ) is a smooth cut off function such that

    χ(ξ)={1,|ξ|R,0,|ξ|R+1.

    Then we have the following decomposition for (˜n,˜m)=G˜U0 as

    ˆ˜n=ˆNˆ˜U0=ˆN1ˆ˜U0+ˆN2ˆ˜U0=(ˆN1+ˆN1)ˆ˜U0+(ˆN2+ˆN2)ˆ˜U0,ˆ˜m=ˆMˆ˜U0=ˆM1ˆU0+ˆM2ˆ˜U0=(ˆM1+ˆM1)ˆ˜U0+(ˆM2+ˆM2)ˆ˜U0. (3.2)

    To derive the long time decay rate of solution, we need to use accurate approximation to the Fourier transform ˆG(t,x) of Green's function for both lower frequency and high frequency. In terms of the definition of the eigenvalues, we are able to obtain that it holds for |ξ|η for some small positive constant η that

    λ1=2ˉμ+ˉν2|ξ|2+i24c2|ξ|2(2ˉμ+ˉν)2|ξ|4=a+bi,λ2=2ˉμ+ˉν2|ξ|2i24c2|ξ|2(2ˉμ+ˉν)2|ξ|4=abi, (3.3)

    and we have

    λ1eλ2tλ2eλ1tλ1λ2=e12(2ˉμ+ˉν)|ξ|2t[cos(bt)+12(2ˉμ+ˉν)|ξ|2sin(bt)b]O(1)e12(2ˉμ+ˉν)|ξ|2t,|ξ|η,
    λ1eλ1tλ2eλ2tλ1λ2=e12(2ˉμ+ˉν)|ξ|2t[cos(bt)12(2ˉμ+ˉν)|ξ|2sin(bt)b]O(1)e12(2ˉμ+ˉν)|ξ|2t,|ξ|η,
    eλ1teλ2tλ1λ2=e12(2ˉμ+ˉν)|ξ|2tsin(bt)bO(1)1|ξ|e12(2ˉμ+ˉν)|ξ|2t,|ξ|η,

    where

    b=124c2|ξ|2(2ˉμ+ˉν)2|ξ|4c|ξ|+O(|ξ|3),|ξ|η.

    For the high frequency |ξ|η, we are also able to obtain that it holds for |ξ|η that

    λ1=2ˉμ+ˉν2|ξ|212(2ˉμ+ˉν)2|ξ|44c2|ξ|2=ab,λ2=2ˉμ+ˉν2|ξ|2+12(2ˉμ+ˉν)2|ξ|44c2|ξ|2=a+b, (3.4)

    and we have

    λ1eλ2tλ2eλ1tλ1λ2=12e(a+b)t[1+e2bt]a2be(a+b)t[1e2bt]O(1)eR0t,|ξ|η,
    λ1eλ1tλ2eλ2tλ1λ2=a+b2be(a+b)t[1e2bt]+e(ab)tO(1)eR0t,|ξ|η,
    eλ1teλ2tλ1λ2=12be(a+b)t[1e2bt]O(1)1|ξ|2eR0t,|ξ|η,

    where

    b=12(2ˉμ+ˉν)2|ξ|44c2|ξ|212(2ˉμ+ˉν)|ξ|22c22ˉμ+ˉν+O(|ξ|2),|ξ|η.

    Here R0, η are some fixed positive constants.

    In this section, we apply the spectral analysis to the semigroup for the linearized Navier-Stokes system. We will establish the L2 and Lp (2p) time decay rate of the global solutions for the linearized Navier-Stokes system.

    With the help of the formula for Green's function in Fourier space and the asymptotic analysis on its elements, we are able to establish the L2 time decay rate. Indeed, we have the L2-time decay rate of the global strong solution to the problem for the linearized Navier-Stokes system as follows.

    Proposition 4.1. Let U0=(n0,m0)L1(R3)Hl(R3) with l3, then (˜n,˜m) solves the linearized Navier-Stokes system (3.1) and satisfies for 0kl that

    k(˜n,˜m)(t)L2(R3)C(1+t)34k2(U0L1(R3)+kU0L2(R3)),

    where C is a positive constant independent of time.

    Proof. A straightforward computation together with the formula of the Green's function ˆG(t,ξ) gives

    ˆ˜n(t,ξ)=λ1eλ2tλ2eλ1tλ1λ2ˆn0iξˆm0(eλ1teλ2t)λ1λ2{O(1)e12(2ˉμ+ˉν)|ξ|2t(|ˆn0|+|ˆm0|),|ξ|η,O(1)eR0t(|ˆn0|+|ˆm0|),|ξ|η,ˆ˜m(t,ξ)=c2iξ(eλ1teλ2t)λ1λ2ˆn0+eλ0tˆm0+(λ1eλ1tλ2eλ2tλ1λ2eλ0t)ξ(ξˆm0)|ξ|2{O(1)eˉμ|ξ|2t(|ˆn0|+|ˆm0|),|ξ|η,O(1)eR0t(|ˆn0|+|ˆm0|),|ξ|η,

    here and below, R0, η are some fixed positive constants. Therefore, we have the L2-decay rate for (˜n,˜m) as

    (ˆ˜n,ˆ˜m)(t)2L2(R3)=|ξ|η|(ˆ˜n,ˆ˜m)(t,ξ)|2dξ+|ξ|η|(ˆ˜n,ˆ˜m)(t,ξ)|2dξ|ξ|ηe2ˉμ|ξ|2t(|ˆn0|2+|ˆm0|2)dξ+|ξ|ηe2R0t(|ˆn0|2+|ˆm0|2)dξ(1+t)32(n0,m0)2L1(R3)L2(R3).

    And the L2-decay rate on the derivatives of (˜n,˜m) as

    (^k˜n,^k˜m)(t)2L2(R3)=|ξ|η|ξ|2k|(ˆ˜n,ˆ˜m)(t,ξ)|2dξ+|ξ|η|ξ|2k|(ˆ˜n,ˆ˜m)(t,ξ)|2dξ|ξ|ηe2ˉμ|ξ|2t|ξ|2k(|ˆn0|2+|ˆm0|2)dξ+|ξ|ηe2R0t|ξ|2k(|ˆn0|2+|ˆm0|2)dξ(1+t)32k((n0,m0)2L1(R3)+(kn0,km0)2L2(R3)).

    The proof of the Proposition 4.1 is completed.

    It should be noted that the L^2 -time decay rates derived above are optimal.

    Proposition 4.2. Let U_0 = ( n_0, m_0)\in L^1({\mathop{\mathbb R\kern 0pt}\nolimits}^3)\cap H^l({\mathop{\mathbb R\kern 0pt}\nolimits}^3) with l\geq3 , assume that M_n = \int_{{\mathop{\mathbb R\kern 0pt}\nolimits}^3}n_0(x) d x and M_m = \int_{{\mathop{\mathbb R\kern 0pt}\nolimits}^3} m_0(x) dx satisfies that M_n , M_m are at least not all zeros, then the solution (\widetilde n,\widetilde m) of the linearized Navier-Stokes system (3.1) given by Proposition 4.1 satisfies for 0\leq k\leq l

    \begin{eqnarray*} \begin{split} C^{-1}(1+t)^{-\frac34-\frac{k}2}&\leq \|{\nabla}^{k} \widetilde n(t)\|_{L^2({\mathop{\mathbb R\kern 0pt}\nolimits}^3)} \leq C(1+t)^{-\frac34-\frac{k}2},\\ C^{-1}(1+t)^{-\frac34-\frac{k}2}&\leq \|{\nabla}^{k} \widetilde m(t)\|_{L^2({\mathop{\mathbb R\kern 0pt}\nolimits}^3)} \leq C(1+t)^{-\frac34-\frac{k}2}, \end{split} \end{eqnarray*}

    where C is a positive constant independent of time.

    Proof. We only show the case of k = 0 for simplicity, the argument applies to the other orders of derivatives. From the formula of the Green's function \widehat G(t,\xi) , we deduce that

    \begin{eqnarray*} \begin{split} &\quad\widehat {\widetilde n}(t,\xi)\\& = \frac{\lambda_1e^{\lambda_2 t}-\lambda_2e^{\lambda_1 t}}{\lambda_1-\lambda_2}\widehat{n}_0-\frac{i\xi\cdot\widehat{m}_0(e^{\lambda_1 t}-e^{\lambda_2 t})}{\lambda_1-\lambda_2} \\ & = e^{-\frac12 (2\bar\mu+\bar\nu)|\xi|^2t}\left[\cos(b t)\widehat{n}_0-i\xi\cdot\widehat{m}_0\frac{\sin(b t)}{b}\right]+e^{-\frac12(2\bar\mu+\bar\nu)|\xi|^2t}\left[\frac12 (2\bar\mu+\bar\nu)|\xi|^2\frac{\sin(b t)}{b}\widehat{n}_0\right]\\ & = T_1+T_2, \quad \text{for}\quad|\xi|\leq \eta,\\ \end{split} \end{eqnarray*}
    \begin{eqnarray*} \begin{split} &\quad\widehat {\widetilde m}(t,\xi)\\ & = -\frac{c^2 i\xi(e^{\lambda_1 t}-e^{\lambda_2 t})}{\lambda_1-\lambda_2}\widehat{n}_0+e^{-\lambda_0 t}\widehat{m}_0+\left(\frac{\lambda_1e^{\lambda_1 t}-\lambda_2e^{\lambda_2 t}}{\lambda_1-\lambda_2}-e^{-\lambda_0 t}\right)\frac{\xi (\xi\cdot\widehat{m}_0)} {|\xi|^2} \\ & = \bigg[e^{-\frac12 (2\bar\mu+\bar\nu)|\xi|^2t}\left[\cos(b t)\frac{\xi (\xi\cdot\widehat{m}_0)} {|\xi|^2}-c^2i\xi\frac{\sin(b t)}{b}\widehat{n}_0\right]+e^{- \bar\mu|\xi|^2t}\left[\widehat{m}_0-\frac{\xi (\xi\cdot\widehat{m}_0)} {|\xi|^2}\right]\bigg]\\ &\quad\quad-e^{-\frac12 (2\bar\mu+\bar\nu)|\xi|^2t}\left[\frac12 (2\bar\mu+\bar\nu)|\xi|^2\frac{\sin(b t)}{b}\frac{\xi (\xi\cdot\widehat{m}_0)} {|\xi|^2}\right]\\ & = S_1+S_2, \quad \text{for}\quad|\xi|\leq \eta,\\ \end{split} \end{eqnarray*}

    here and below, \eta is a sufficiently small but fixed constant.

    It is easy to check that

    \begin{equation} \begin{split} \|\widehat {\widetilde n}(t,\xi)\|^2_{L^2}& = \int_{|\xi|\leq \eta}|\widehat {\widetilde n}(t,\xi)|^2 d\xi+\int_{|\xi|\geq \eta}|\widehat {\widetilde n}(t,\xi)|^2 d\xi\\ &\geq \int_{|\xi|\leq \eta}|T_1+T_2|^2 d\xi \geq \int_{|\xi|\leq \eta}\frac12|T_1|^2-|T_2|^2 d\xi. \end{split} \end{equation} (4.1)

    We then calculate that

    \begin{equation} \begin{split} \int_{|\xi|\leq \eta}|T_2|^2 d\xi &\lesssim \|\widehat{n}_0\|_{L^\infty}^2\int_{|\xi|\leq \eta}e^{-(2\bar\mu+\bar\nu)|\xi|^2t}|\xi|^4\left(\frac{\sin(b t)}{b}\right)^2 d\xi \\ &\lesssim\|\widehat{n}_0\|_{L^\infty}^2\int_{|\xi|\leq \eta}e^{-(2\bar\mu+\bar\nu)|\xi|^2t}|\xi|^2 d\xi\lesssim(1+t)^{-\frac52}\|{n}_0\|_{L^1}^2. \end{split} \end{equation} (4.2)

    Since n_0(x) \in L^1 implies \widehat{n}_0(\xi) \in C({{\mathop{\mathbb R\kern 0pt}\nolimits}^3}) . If \widehat{n}_0(0) = \int_{{\mathop{\mathbb R\kern 0pt}\nolimits}^3}n_0(x) d x\neq 0 , we deduce that \widehat{n}_0(\xi)\neq 0 for |\xi|\leq\eta when \eta is sufficiently small. One finds that, when M_n\neq 0 ,

    \begin{eqnarray*} \begin{split} |\widehat{n}_0(\xi)|^2\geq \frac1C\left|\int_{{\mathop{\mathbb R\kern 0pt}\nolimits}^3}n_0(x) d x\right|^2 \geq \frac{M_n^2}C,\quad \text{for}\quad |\xi|\leq\eta. \end{split} \end{eqnarray*}

    For \widehat{m}_0 , a similar argument yields that, when M_m\neq 0 , we have

    \begin{eqnarray*} \begin{split} \frac{\left|\xi\cdot\widehat{m}_0(\xi)\right|^2}{|\xi|^2} \geq \frac{\left|\xi\cdot M_m\right|^2}{C|\xi|^2},\quad \text{for}\quad |\xi|\leq\eta. \end{split} \end{eqnarray*}

    When M_n\neq 0 , M_m\neq 0 , with the help of the above analysis, using b\sim c|\xi|+O(|\xi|^3) for |\xi|\leq \eta , we obtain that

    \begin{equation} \label{optimal3} \begin{split} &\quad\int_{|\xi|\leq \eta}|T_1|^2 d\xi\nonumber\\ &\geq \frac{M_n^2}C\int_{|\xi|\leq \eta}e^{-(2\bar\mu+\bar\nu)|\xi|^2t}\cos^2(b t)d\xi+\frac1C\int_{|\xi|\leq \eta}\frac{\left|\xi\cdot M_m\right|^2}{b^2}e^{-(2\bar\mu+\bar\nu)|\xi|^2t}\sin^2(b t)d\xi \nonumber \end{split} \end{equation}
    \begin{equation} \begin{split}&\geq \frac{\min\{{M_n^2},\frac{M_m^2}{3c^2}\}}C\int_{|\xi|\leq \eta}e^{-(2\bar\mu+\bar\nu)|\xi|^2t}\big(\cos^2(b t)+\sin^2(b t)\big)d\xi\\ &\geq C_1\int_{|\xi|\leq \eta}e^{-(2\bar\mu+\bar\nu)|\xi|^2t}d\xi\\ &\geq C^{-1}(1+t)^{-\frac32}. \end{split} \end{equation} (4.3)

    If M_n\neq 0 , M_m = 0 , and by the conituinity of \widehat{m}_0 near \xi = 0 , there exists a small enough constant \epsilon such that \epsilon\to0 as \xi\to 0 , and

    \begin{eqnarray*} |\widehat{m}_0(\xi)|^2 < \epsilon,\quad \text{for}\quad |\xi|\leq\eta. \end{eqnarray*}

    We thus use the help of spherical coordinates and the change of variables r = |\xi|\sqrt{t} to obtain that

    \begin{equation} \begin{split} &\quad\int_{|\xi|\leq \eta}|T_1|^2 d\xi\\ &\geq \frac{M_n^2}C\int_{|\xi|\leq \eta}e^{-(2\bar\mu+\bar\nu)|\xi|^2t}\cos^2(b t)d\xi-\frac{\epsilon}{Cc^2}\int_{|\xi|\leq \eta}e^{-(2\bar\mu+\bar\nu)|\xi|^2t}\sin^2(b t)d\xi\\ &\geq \frac{M_n^2}Ct^{-\frac32}\int_0^{\eta\sqrt t}e^{-(2\bar\mu+\bar\nu)r^2}\cos^2(cr\sqrt t)r^2dr-\frac{\epsilon}{Cc^2}t^{-\frac32}\int_0^{\eta\sqrt t}e^{-(2\bar\mu+\bar\nu)r^2}\sin^2(cr\sqrt t)r^2dr\\ &\geq \frac{M_n^2}Ct^{-\frac32}\sum\limits_{k = 0}^{[\frac{c\eta t}\pi]-1}\int_{\frac{k\pi}{c\sqrt t}}^{\frac{k\pi+\frac\pi4}{c\sqrt t}}e^{-(2\bar\mu+\bar\nu)r^2}\cos^2(cr\sqrt t)r^2dr-\frac{\epsilon}{Cc^2}(1+t)^{-\frac32}\\ &\geq \frac{M_n^2}{2C}t^{-\frac32}\sum\limits_{k = 0}^{[\frac{c\eta t}\pi]-1}\int_{\frac{k\pi}{c\sqrt t}}^{\frac{k\pi+\frac\pi4}{c\sqrt t}}e^{-(2\bar\mu+\bar\nu)r^2}r^2dr-\frac{\epsilon}{Cc^2}(1+t)^{-\frac32}\\ &\geq C_1^{-1}(1+t)^{-\frac32}-C_2^{-1}\epsilon(1+t)^{-\frac32}\\ .&\geq C^{-1}(1+t)^{-\frac32} \end{split} \end{equation} (4.4)

    In the case of M_n = 0 , M_m\neq0 , we can use a similar argument to obtain that

    \begin{equation} \begin{split} &\quad\int_{|\xi|\leq \eta}|T_1|^2 d\xi\\ &\geq -\frac\epsilon C\int_{|\xi|\leq \eta}e^{-(2\bar\mu+\bar\nu)|\xi|^2t}\cos^2(b t)d\xi+\frac{M_m^2}{3Cc^2}\int_{|\xi|\leq \eta}e^{-(2\bar\mu+\bar\nu)|\xi|^2t}\sin^2(b t)d\xi\\ &\geq C^{-1}(1+t)^{-\frac32}. \end{split} \end{equation} (4.5)

    Combining the above estimates (4.1), (4.2), (4.3), (4.4) and (4.5), we obtain the lower bound of the time decay rate for {\widetilde n}(t,x) as

    \begin{eqnarray*} \| {\widetilde n}(t,x)\|^2_{L^2} = \|\widehat {\widetilde n}(t,\xi)\|^2_{L^2}\geq C^{-1}(1+t)^{-\frac32}. \end{eqnarray*}

    The lower bound of the time decay rate for {\widetilde m}(t,x) can be shown in a similar fashion. It is not difficult to derive that

    \begin{equation} \begin{split} \|\widehat {\widetilde m}(t,\xi)\|^2_{L^2}\geq \int_{|\xi|\leq \eta}\frac12|S_1|^2-|S_2|^2 d\xi, \end{split} \end{equation} (4.6)

    then we find that

    \begin{equation} \begin{split} \int_{|\xi|\leq \eta}|S_2|^2 d\xi \lesssim(1+t)^{-\frac52}\|{m}_0\|_{L^1}^2. \end{split} \end{equation} (4.7)

    We then calculate that

    \begin{eqnarray*} \begin{split} &\quad\int_{|\xi|\leq \eta}|S_1|^2 d\xi\\ &\geq \bigg\{\frac{c^4M_n^2}C\int_{|\xi|\leq \eta}\frac{|\xi|^2}{b^2}e^{-(2\bar\mu+\bar\nu)|\xi|^2t}\sin^2(b t)d\xi\\ &\qquad+\frac{1}C\int_{|\xi|\leq \eta}\frac{\left|\xi\cdot M_m\right|^2}{|\xi|^2}e^{-(2\bar\mu+\bar\nu)|\xi|^2t}\cos^2(bt)d\xi\bigg\}\\ &\qquad +\bigg\{\int_{|\xi|\leq \eta}e^{-\frac12(4\bar\mu+\bar\nu)|\xi|^2t}\cos(bt)\frac{\xi(\xi\cdot{\widehat m}_0)}{|\xi|^2}\left({\widehat m}_0-\frac{\xi(\xi\cdot{\widehat m}_0)}{|\xi|^2}\right)d\xi\bigg\}\\ & = J_1+J_2. \end{split} \end{eqnarray*}

    A direct computation gives rise to

    \begin{equation} J_1\geq C^{-1}(1+t)^{-\frac32},\qquad J_2 = 0. \end{equation} (4.8)

    Combining the above estimates (4.6), (4.7) and (4.8), we obtain the lower bound of the time decay rate for {\widetilde m}(t,x) as

    \begin{eqnarray*} \| {\widetilde m}(t,x)\|^2_{L^2} = \|\widehat {\widetilde m}(t,\xi)\|^2_{L^2}\geq C^{-1}(1+t)^{-\frac32}. \end{eqnarray*}

    Then the proof of Proposition 4.2 is completed.

    In this subsection, we establish the following L^p -time decay rate of the global strong solution to the linearized Navier-Stokes system with p \in [2,+\infty] .

    Proposition 4.3. Let U_0 = (n_0,m_0)\in L^1({\mathop{\mathbb R\kern 0pt}\nolimits}^3)\cap W^{l,p}({\mathop{\mathbb R\kern 0pt}\nolimits}^3) with l\geq 3 , then (\widetilde n,\widetilde m) solves the linearized Navier-Stokes system (3.1) and satisfies for 0\leq k\leq l and p \in [2,+\infty] that

    \begin{eqnarray*} \|{\nabla}^{k} (\widetilde n,\widetilde m)(t)\|_{L^p({\mathop{\mathbb R\kern 0pt}\nolimits}^3)} \leq C(1+t)^{-\frac32(1-\frac1p)-\frac{k}2}\big(\|U_0\|_{L^1({\mathop{\mathbb R\kern 0pt}\nolimits}^3)}+\|{\nabla}^k U_0\|_{L^p({\mathop{\mathbb R\kern 0pt}\nolimits}^3)}\big), \end{eqnarray*}

    where C is a positive constant independent of time.

    To prove Proposition 4.3, the following two lemmas in [6] are helpful.

    Lemma 4.1. Let n\geq 1 and assume that \hat f(\xi) \in L^\infty \cap C^{n+1}({\mathop{\mathbb R\kern 0pt}\nolimits}^n/\{0\}) , with

    \begin{eqnarray*} \begin{split} |{\nabla}^\alpha _\xi \hat f(\xi)| \leq C'{ \left\{\begin{array}{l} |\xi|^{-|\alpha|+\sigma_1},\quad |\xi|\leq R, |\alpha| = n,\\ |\xi|^{-|\alpha|-\sigma_2},\quad |\xi|\geq R, |\alpha| = n-1,n,n+1, \end{array}\right.} \end{split} \end{eqnarray*}

    where \sigma_1, \sigma_2>0 and n>2-2\sigma_2 . Then \hat f(\xi) is continuous at 0 and \infty , and

    \begin{eqnarray*} f = m_1+m_2\delta, \end{eqnarray*}

    where m_1\in L^1({\mathop{\mathbb R\kern 0pt}\nolimits}^n) satisfies \|m_1\|_{L^1({\mathop{\mathbb R\kern 0pt}\nolimits}^n)} \leq C(C') , m_2 is the constant

    \begin{eqnarray*} m_2 = (2\pi)^{-\frac n 2}\lim\limits_{|\xi| \to \infty} \hat f(\xi), \end{eqnarray*}

    and \delta is the Dirac distribution. In particular, \hat f(\xi) is a strong L^p multiplier, 1\leq p \leq \infty , in the sense that, for any g\in L^p ,

    \begin{eqnarray*} \|f\ast g\|_{L^p} \leq C\| g\|_{L^p},\quad 1\leq p \leq \infty, \end{eqnarray*}

    where C depends only on |m_2|\leq \| \hat f\|_{L^\infty} and the constant C' above.

    Lemma 4.2. Let \hat g(t,\xi) = \hat K(t,\xi)\hat f(\xi) , where \hat K(t,\xi) = e^{-\vartheta|\xi|^2 t} , \hat f(\xi)\in L^\infty\cap C^{n+1}({\mathop{\mathbb R\kern 0pt}\nolimits}^n) , and

    \begin{eqnarray*} |{\nabla}_\xi ^\beta \hat f(\xi)|\leq C'|\xi|^{-|\beta|},\quad |\beta|\leq n+1. \end{eqnarray*}

    Then {\nabla}_x ^\alpha g(t,\cdot)\in L^p for t>0 , and for all \alpha , 1\leq p \leq \infty , we have

    \begin{eqnarray*} \|{\nabla}_x ^\alpha g(t,\cdot)\|_{L^p}\leq C(|\alpha|)t^{-\frac n 2(1-\frac 1 p)-\frac{|\alpha|}{2}}. \end{eqnarray*}

    In particular, \widehat{{\nabla}_x ^\alpha g(t,x)} = (i\xi)^\alpha \hat g(t,\xi) is a strong L^p multiplier, with norm bounded by C(|\alpha|,\vartheta)C't^{-\frac{|\alpha|}2} , where the constant C(|\alpha|,\vartheta) depends only on |\alpha| and \vartheta .

    Now let us turn to the proof of Proposition 4.3.

    Proof of Proposition 4.3. We first analyze above higher frequency terms denoted by \widehat {(\cdot)}_2 . Recall that

    \begin{eqnarray*} \begin{split} \lambda_1 = -(2\bar\mu+\bar\nu) |\xi|^2+\frac{2c^2}{2\bar\mu+\bar\nu}+O(|\xi|^{-2}),\quad \lambda_2 = -\frac{2c^2}{2\bar\mu+\bar\nu}+O(|\xi|^{-2}),\quad |\xi|\geq \eta. \end{split} \end{eqnarray*}

    We shall prove that the higher frequency terms are L^p Fourier multipliers with an exponential time decay coefficient C e^{-c_1t} for some constants c_1>0 . For simplicity, we only show that \widehat {\mathcal N}_2 is an L^p Fourier multiplier at higher frequency as follows. It holds

    \begin{eqnarray*} \frac{\lambda_1e^{\lambda_2 t}-\lambda_2e^{\lambda_1 t}}{\lambda_1-\lambda_2} = e^{\lambda_2 t}+\frac{\lambda_2 e^{\lambda_2 t}}{\lambda_1-\lambda_2}-\frac{\lambda_2 e^{\lambda_1 t}}{\lambda_1-\lambda_2}. \end{eqnarray*}

    By a direct computation, it is easy to verify

    \begin{eqnarray*} |{\nabla}_\xi^k \lambda_2|\lesssim|\xi|^{-2-k},\quad |\xi|\geq\eta, \end{eqnarray*}

    which gives rise to

    \begin{eqnarray*} \begin{split} \bigg|{\nabla}_\xi^k \Big[(1-\chi(\cdot))e^{\lambda_2 t}\Big]\bigg|, \left|{\nabla}_\xi^k \Big[(1-\chi(\cdot))\frac{\lambda_2 e^{\lambda_2 t}}{\lambda_1-\lambda_2}\Big]\right|\lesssim{ \left\{\begin{array}{l} 0, \quad |\xi|\leq R,\\ e^{-c_1t}|\xi|^{-2-k},\quad |\xi|\geq R, \end{array}\right.} \end{split} \end{eqnarray*}

    here and below, R>0 is a given constant. Thus, from Lemma 4.1 it follows that the inverse Fourier transform of the term (1-\chi(\cdot))\left(e^{\lambda_2 t}+\frac{\lambda_2 e^{\lambda_2 t}}{\lambda_1-\lambda_2}\right) is an L^p multiplier with the coefficient Ce^{-c_1t} . The other part of \widehat {\mathcal N}_2 at higher frequency can be written as

    \begin{eqnarray*} (1-\chi(\cdot))\frac{\lambda_2 e^{\lambda_1 t}}{\lambda_1-\lambda_2} \sim e^{-\frac12(2\bar\mu+\bar\nu) |\xi|^2 t}\Big[(1-\chi(\cdot))\frac{e^{(-\lambda_2-\frac12(2\bar\mu+\bar\nu) |\xi|^2)t}}{\lambda_1-\lambda_2}\Big]. \end{eqnarray*}

    We can regard e^{-\frac12 (2\bar\mu+\bar\nu) |\xi|^2 t} as the function K(t,\xi) of Lemma 4.2, and the rest term satisfies the condition. Thus, the inverse Fourier transform of (1-\chi(\cdot))\frac{\lambda_2 e^{\lambda_1 t}}{\lambda_1-\lambda_2} is also an L^p multiplier with the coefficient Ce^{-c_1t} . These facts imply that \widehat {\mathcal N}_2 at higher frequency is an L^p multiplier with the coefficient Ce^{-c_1t} . Applying the similar analysis to the terms \widehat {\mathfrak N}_2 , \widehat {\mathcal M}_2 , and \widehat {\mathfrak M}_2 , we can show that their inverse Fourier transform are all L^p multiplier with the constant coefficient Ce^{-c_1t} . Then

    \begin{equation} \|({\nabla}_x^k({\mathcal N}_2 \ast f),{\nabla}_x^k({\mathfrak N}_2 \ast f),{\nabla}_x^k( {\mathcal M}_2 \ast f),{\nabla}_x^k( {\mathfrak M}_2 \ast f))(t)\|_{L^p} \leq Ce^{-c_1t}\|{\nabla}_x^k f\|_{L^p}, \end{equation} (4.9)

    for all integer k\geq 0 , and p\in[2,\infty] .

    We also need to deal with the corresponding lower frequency terms denoted by \widehat {(\cdot)}_1 . Recall that

    \begin{eqnarray*} \begin{split} &\frac{\lambda_1e^{\lambda_2 t}-\lambda_2e^{\lambda_1 t}}{\lambda_1-\lambda_2}, \frac{\lambda_1e^{\lambda_1 t}-\lambda_2e^{\lambda_2 t}}{\lambda_1-\lambda_2},\frac{|\xi|(e^{\lambda_1 t}-e^{\lambda_2 t})}{\lambda_1-\lambda_2}\sim O(1)e^{-\frac12 (2\bar\mu+\bar\nu)|\xi|^2t},\quad |\xi|\leq\eta, \end{split} \end{eqnarray*}

    which imply that for |\xi|\leq\eta that

    \begin{eqnarray*} |\widehat{\mathcal N}_1|\sim O(1)e^{-c_2|\xi|^2t},\quad |\widehat{\mathfrak N}_1|\sim O(1)e^{-c_2|\xi|^2t},\\ |\widehat{\mathcal M}_1|\sim O(1)e^{-c_2|\xi|^2t},\quad |\widehat{\mathfrak M}_1|\sim O(1)e^{-c_2|\xi|^2t}, \end{eqnarray*}

    for some constants c_2>0 . Thus, by Hausdroff-Young's inequality with p\in[2, +\infty] , we can obtain

    \begin{equation} \begin{split} \|({\nabla}^k{\mathcal N}_1,{\nabla}^k{\mathfrak N}_1,{\nabla}^k {\mathcal M}_1,{\nabla}^k {\mathfrak M}_1)(t)\|_{L^p} \leq& C\left(\int_{|\xi|\leq \eta}\big||\xi|^k e^{-c_2|\xi|^2t}\big|^q d\xi\right)^{\frac 1q}\\ \leq& C(1+t)^{-\frac32(1-\frac1p)-\frac k2}. \end{split} \end{equation} (4.10)

    Combining (4.9) and (4.10), we finally have for t>0 that

    \begin{eqnarray*} \begin{split} \|({\nabla}^k(N \ast f),{\nabla}^k(M \ast f))(t)\|_{L^p}& = \|({\nabla}^k((N_1+N_2) \ast f),{\nabla}^k((M_1+M_2) \ast f))(t)\|_{L^p}\\ &\leq C(1+t)^{-\frac32(1-\frac1p)-\frac k2}\|f\|_{L^1}+Ce^{-c_1t}\|{\nabla}^k f\|_{L^p}\\ &\leq C(1+t)^{-\frac32(1-\frac1p)-\frac k2}(\|f\|_{L^1}+\|{\nabla}^k f\|_{L^p}). \end{split} \end{eqnarray*}

    The proof of Proposition 4.3 is completed.

    We are ready to prove Theorem 1.1 on the sharp time decay rate of the global solution to the initial value problem for the nonlinear Navier-Stokes system.

    In what follows, we will set n_h = n-\widetilde n and m_h = m-\widetilde m , then we have

    \begin{equation} \left\{\begin{array}{l} \partial_t n_h + {\mathop{{\rm{div}}}\nolimits} m_h = 0,\qquad (t,x)\in{\mathop{\mathbb R\kern 0pt}\nolimits}^+\times{\mathop{\mathbb R\kern 0pt}\nolimits}^3,\\ \partial_t m_h + c^2 {\nabla} n_h- \bar\mu △ m_h- (\bar\mu+\bar\nu){\nabla}{\mathop{{\rm{div}}}\nolimits} m_h = F,\qquad (t,x)\in{\mathop{\mathbb R\kern 0pt}\nolimits}^+\times{\mathop{\mathbb R\kern 0pt}\nolimits}^3,\\ \lim\limits_{|x|\to\infty}n_h = 0, \quad\lim\limits_{|x|\to\infty} m_h = 0,\qquad t\in{\mathop{\mathbb R\kern 0pt}\nolimits}^+,\\ (n_h,m_h)\big|_{t = 0} = (0,0),\qquad x\in{\mathop{\mathbb R\kern 0pt}\nolimits}^3, \end{array}\right. \end{equation} (5.1)

    where \bar\mu = \frac{\mu}{\bar\rho} , \bar\nu = \frac{\nu}{\bar\rho} , c = \sqrt{p'(\bar \rho)} , and

    \begin{eqnarray*} \begin{split} F = &-{\mathop{{\rm{div}}}\nolimits}\Big\{ \frac{(m_h+\widetilde m)\otimes (m_h+\widetilde m)}{n_h+\widetilde n+\bar\rho}+\bar\mu{\nabla}\big(\frac{(n_h+\widetilde n)(m_h+\widetilde m)}{n_h+\widetilde n+\bar\rho}\big)\Big\}\\ & - {\nabla}\Big\{(\bar\mu+\bar\nu){\mathop{{\rm{div}}}\nolimits}(\frac{(n_h+\widetilde n)(m_h+\widetilde m)}{n_h+\widetilde n+\bar\rho})+\big(p(n_h+\widetilde n+\bar\rho)-p(\bar\rho)-c^2(n_h+\widetilde n)\big)\Big\}. \end{split} \end{eqnarray*}

    Denote U_h = (n_h, m_h)^t , we have the equivalent form of system (5.1) in vector form

    \begin{eqnarray*} \partial_t U_h = BU_h+H,\quad t\geq0,\qquad U_h(0) = 0, \end{eqnarray*}

    where the nonlinear term H(\widetilde U, U_h) = (0, F(\widetilde U, U_h))^t . Thus, we can represent the solution in term of the semigroup

    \begin{eqnarray*} U_h(t) = S(t)\ast U_{h}(0)+\int_0^t S(t-\tau)\ast H(\widetilde U, U_h)(\tau)d\tau, \end{eqnarray*}

    which (n_h, m_h) can be decomposed as

    \begin{equation} n_h = N\ast U_{h}(0)+\int_0^t \mathfrak N (t-\tau)\ast H(\tau)d\tau, \end{equation} (5.2)
    \begin{equation} m_h = M\ast U_{h}(0)+\int_0^t \mathfrak M (t-\tau)\ast H(\tau)d\tau. \end{equation} (5.3)

    Furthermore, in view of the above definition for \widehat{\mathfrak N}(\xi) and \widehat{\mathfrak M}(\xi) , it is easy to verify for some constants c_3>0 , c_4>0 , R_0>0 , we discover that

    \begin{eqnarray*} |\widehat{\mathfrak N}(\xi)|\sim O(1)e^{-c_3|\xi|^2t}, \quad |\widehat{\mathfrak M}(\xi)|\sim O(1)e^{-c_3|\xi|^2t}, \quad|\xi|\leq\eta, \end{eqnarray*}
    \begin{eqnarray*} |\widehat{\mathfrak N}(\xi)|\sim O(1)\frac1{|\xi|}e^{-R_0t}, \quad |\widehat{\mathfrak M}(\xi)|\sim O(1)\frac1{|\xi|^2}e^{-R_0t}+O(1)e^{-c_4|\xi|^2t}, \quad |\xi|\geq\eta. \end{eqnarray*}

    Thus, applying a similar argument as in the proof of Proposition 4.1, we have

    \begin{equation} \|({\nabla}^k {\mathfrak N}\ast H, {\nabla}^k {\mathfrak M}\ast H)(t)\|_{L^2} \leq C(1+t)^{-\frac32(\frac1q-\frac12)-\frac12-\frac k 2}\big(\|Q\|_{L^q}+\|{\nabla}^{k+1} Q\|_{L^2}\big),\quad q = 1,2, \end{equation} (5.4)
    \begin{equation} \|({\nabla}^k {\mathfrak N}\ast H, {\nabla}^k {\mathfrak M}\ast H)(t)\|_{L^2} \leq C(1+t)^{-\frac32(\frac1q-\frac12)-\frac12-\frac k 2}\big(\|Q\|_{L^q}+\|{\nabla}^{k} Q\|_{L^2}\big),\quad q = 1,2, \end{equation} (5.5)
    \begin{equation} \|{\nabla}^k {\mathfrak M}\ast H(t)\|_{L^2} \leq C(1+t)^{-\frac32(\frac1q-\frac12)-\frac12-\frac k 2}\big(\|Q\|_{L^q}+\|{\nabla}^{k-1} Q\|_{L^2}\big),\quad q = 1,2, \end{equation} (5.6)

    for any non-negative integer k and

    \begin{equation} \begin{split} Q = &\Big|\frac{(m_h+\widetilde m)\otimes (m_h+\widetilde m)}{n_h+\widetilde n+\bar\rho}+\bar\mu{\nabla}\big(\frac{(n_h+\widetilde n)(m_h+\widetilde m)}{n_h+\widetilde n+\bar\rho}\big)\Big|\\ &+\Big|(\bar\mu+\bar\nu){\mathop{{\rm{div}}}\nolimits}(\frac{(n_h+\widetilde n)(m_h+\widetilde m)}{n_h+\widetilde n+\bar\rho})+\big(p(n_h+\widetilde n+\bar\rho)-p(\bar\rho)-c^2(n_h+\widetilde n)\big)\Big|. \end{split} \end{equation} (5.7)

    For readers' convenience, we show how to estimate \|{\nabla}^k {\mathfrak M}\ast H(t)\|_{L^2} as an example. The other two estimates can be obtained by the similar argument. Indeed,

    \begin{eqnarray*} \begin{split} &\quad\|{\nabla}^k {\mathfrak M}\ast H(t)\|_{L^2}^2\\ &\lesssim\int_{|\xi|\leq \eta}e^{-2c_3|\xi|^2t}|\xi|^{2k}|\widehat H|^2d\xi +\int_{|\xi|\geq \eta}e^{-2R_0 t}|\xi|^{2k-4}|\widehat H|^2d\xi\\ &\quad+\int_{|\xi|\geq \eta}e^{-2c_4|\xi|^2 t}|\xi|^{2k}|\widehat H|^2d\xi\\ &\lesssim\int_{|\xi|\leq \eta}e^{-2c_3|\xi|^2t}|\xi|^{2k+2}|\widehat Q|^2d\xi +\int_{|\xi|\geq \eta}e^{-2R_0 t}|\xi|^{2k-2}|\widehat Q|^2d\xi\\ &\quad+\int_{|\xi|\geq \eta}e^{-2c_4|\xi|^2t}|\xi|^{2k+2}|\widehat Q|^2d\xi\\ &\lesssim(1+t)^{-3(\frac1q-\frac12)-1-k}\big(\|Q\|^2_{L^q({\mathop{\mathbb R\kern 0pt}\nolimits}^3)}+\|{\nabla}^{\tilde k} Q\|^2_{L^2({\mathop{\mathbb R\kern 0pt}\nolimits}^3)}\big),\quad q = 1,2,\quad k-1\leq\tilde k\in{\mathop{\mathbb N\kern 0pt}\nolimits}^+. \end{split} \end{eqnarray*}

    In this subsection, we establish the faster decay rate for (n_h, m_h) . We will start with an a priori assumption on a carefully chosen quantity \Lambda(t) defined in (5.8), and then later prove a better estimate with the help of the smallness of initial data.

    We begin with following Lemma.

    Lemma 5.1. Let r_1, r_2>0 be real, one has

    \begin{eqnarray*} \begin{split} \int_0^{\frac t 2}(1+t-\tau)^{-r_1}(1+\tau)^{-r_2} d\tau = &\int_0^{\frac t 2}(1+\frac t 2+\tau)^{-r_1}(1+\frac t 2-\tau)^{-r_2} d\tau\\ \lesssim&{ \left\{\begin{array}{l} (1+t)^{-r_1}, \quad \mathit{\text{for}} \quad r_2 > 1,\\ (1+t)^{-(r_1-\epsilon)},\quad \mathit{\text{for}} \quad r_2 = 1,\\ (1+t)^{-(r_1+r_2-1)},\quad \mathit{\text{for}} \quad r_2 < 1, \end{array}\right.} \end{split} \end{eqnarray*}

    and

    \begin{eqnarray*} \begin{split} \int_{\frac t 2}^t(1+t-\tau)^{-r_1}(1+\tau)^{-r_2} d\tau = &\int_0^{\frac t 2}(1+t-\tau)^{-r_2}(1+\tau)^{-r_1} d\tau\\ \lesssim&{ \left\{\begin{array}{l} (1+t)^{-r_2}, \quad \mathit{\text{for}} \quad r_1 > 1,\\ (1+t)^{-(r_2-\epsilon)},\quad \mathit{\text{for}} \quad r_1 = 1,\\ (1+t)^{-(r_1+r_2-1)},\quad \mathit{\text{for}} \quad r_1 < 1, \end{array}\right.} \end{split} \end{eqnarray*}

    where \epsilon>0 is a small but fixed constant.

    Proposition 5.1. Under the assumptions of Theorem 1.1, the solution (n_h, m_h) of the nonlinear system (5.1) satisfies for k = 0,1,2 that

    \begin{eqnarray*} \begin{split} &\|({\nabla} ^k n_h,{\nabla}^{k} m_h)\|_{L^2}\leq C\delta_0^2(1+t)^{-\frac54-\frac {k} 2},\\ &\|{\nabla} ^3 m_h\|_{L^2}\leq C\delta_0^2(1+t)^{-\frac{11}4},\quad \|{\nabla} ^3 n_h\|_{L^2}\leq C\delta_0(1+t)^{-\frac74}, \end{split} \end{eqnarray*}

    where C is a positive constant independent of time.

    From (5.7), we deduce

    \begin{eqnarray*} Q(\widetilde U, U_h) = Q_1+Q_2+Q_3+Q_4, \end{eqnarray*}

    which implies for a smooth solution (n,m) satisfying \|(n,m)\|_{H^3}<\infty that

    \begin{eqnarray*} \begin{split} &Q_1 = Q_1(\widetilde U, U_h)\sim O(1)\left(n_h^2+m_h\otimes m_h+\widetilde n^2+\widetilde m\otimes\widetilde m \right),\\ &Q_2 = Q_2(\widetilde U, U_h)\sim O(1)\left(\widetilde n n_h+\widetilde m\otimes m_h\right),\\ &Q_3 = Q_3(\widetilde U, U_h)\sim O(1)\left({\nabla}(n_h\cdot m_h)+{\nabla}(\widetilde n\cdot\widetilde m)\right),\\ &Q_4 = Q_4(\widetilde U, U_h)\sim O(1)\left({\nabla}(\widetilde n\cdot m_h)+{\nabla}( n_h\cdot\widetilde m) \right). \end{split} \end{eqnarray*}

    Define

    \begin{equation} \begin{split} \Lambda(t) = :&\sup\limits_{0\leq s \leq t}\bigg\{\sum\limits_{k = 0}^2(1+s)^{\frac54+\frac k 2}{\delta_0}^{-\frac34}\|({\nabla} ^k n_h,{\nabla}^{k}m_h)(s)\|_{L^2}\\ &\quad+(1+s)^{\frac74}\|({\nabla}^3 n_h, {\nabla}^3 m_h)(s)\|_{L^2}\bigg \}. \end{split} \end{equation} (5.8)

    Proposition 5.2. Under the assumptions of Theorem 1.1, if for some T>0 , \Lambda(t) \leq \delta_0^{\frac12} for any t\in[0,T] , then it holds that

    \begin{eqnarray*} \begin{split} \Lambda(t)\leq C\delta_0^{\frac34},\quad t\in[0,T], \end{split} \end{eqnarray*}

    where C is a positive constant independent of time.

    The proof of this Proposition 5.2 consists of following three steps.

    Starting with (5.4), (5.5), (5.6) and (5.8), we have after a complicate but straightforward computation that

    \begin{equation} \begin{split} \|(n_h, m_h)\|_{L^2}&\lesssim\int_0^t \|(\mathfrak N (t-\tau)\ast H(\tau), \mathfrak M (t-\tau)\ast H(\tau))\|_{L^2}d\tau\\ &\lesssim\int_0^{t} (1+t-\tau)^{-\frac54}\big(\|Q(\tau)\|_{L^1}+\| Q(\tau)\|_{L^2}\big)d\tau\\ &\lesssim\left(\delta_0^2+\delta_0^{\frac32}\Lambda^2(t)\right)\int_0^{t} (1+t-\tau)^{-\frac54}(1+\tau)^{-\frac32}d\tau\\ &\lesssim(1+t)^{-\frac54}\left(\delta_0^2+\delta_0^{\frac32}\Lambda^2(t)\right). \end{split} \end{equation} (5.9)

    It is easy to verify that

    \begin{eqnarray*} \begin{split} \|Q(t)\|_{L^1}\lesssim&\|Q_1\|_{L^1}+\|Q_2\|_{L^1}+\|Q_3\|_{L^1}+\|Q_4\|_{L^1}\\ \lesssim &\|(\widetilde n,\widetilde m)\|_{L^2}^2+\|( n_h, m_h)\|_{L^2}^2+ \|( n_h, m_h)\|_{L^2}\big(\|({\nabla}\widetilde n,{\nabla}\widetilde m)\|_{L^2}\\ &+\|({\nabla} n_h,{\nabla} m_h)\|_{L^2}\big)+\|(\widetilde n,\widetilde m)\|_{L^2}\left(\|({\nabla}\widetilde n,{\nabla}\widetilde m)\|_{L^2} +\| ({\nabla} n_h,{\nabla} m_h)\|_{L^2}\right)\\ \lesssim & (1+t)^{-\frac32}\left(\delta_0^2+\delta_0^{\frac32}\Lambda^2(t)\right). \end{split} \end{eqnarray*}

    Indeed, by virtue of Hölder's inequality and Gagliardo-Nirenberg's inequality, we obtain that

    \begin{eqnarray*} \|u\|_{L^\infty}\lesssim \|{\nabla} u\|_{L^2}^{\frac12}\|{\nabla} ^2 u\|_{L^2}^{\frac12}, \end{eqnarray*}

    which implies that

    \begin{eqnarray*} \begin{split} &\|Q(t)\|_{L^2}\\ \lesssim &\|(\widetilde n,\widetilde m)\|_{L^\infty}\big(\|(\widetilde n,\widetilde m)\|_{L^2}+\|({\nabla}\widetilde n,{\nabla}\widetilde m)\|_{L^2}+\|(n_h,m_h)\|_{L^2}\\ &+\|({\nabla} n_h,{\nabla} m_h)\|_{L^2}\big)+\|( n_h,m_h)\|_{L^\infty}\left(\|( n_h,m_h)\|_{L^2}+\|( {\nabla} n_h,{\nabla} m_h)\|_{L^2}\right)\\ &+\|({\nabla}\widetilde n,{\nabla}\widetilde m)\|_{L^\infty}\|(n_h,m_h)\|_{L^2}\\ \lesssim &(1+t)^{-\frac94}\left(\delta_0^2+\delta_0^{\frac32}\Lambda^2(t)\right). \end{split} \end{eqnarray*}

    Furthermore, exactly as in the estimate of the high order derivatives, we have

    \begin{equation} \begin{split} &\|({\nabla} n_h, {\nabla} m_h)\|_{L^2}\\ \lesssim &\int_0^{\frac t 2} \|({\nabla}\mathfrak N , {\nabla}\mathfrak M)(t-\tau)\ast H(\tau)\|_{L^2}d\tau+\int_{\frac t 2}^t \|(\mathfrak N, \mathfrak M )(t-\tau)\ast {\nabla} H(\tau)\|_{L^2}d\tau\\ \lesssim &\int_0^{\frac t 2} (1+t-\tau)^{-\frac74}\big(\| Q(\tau)\|_{L^1}+\|{\nabla} Q(\tau)\|_{L^2}\big)d\tau +\int_{\frac t 2}^t (1+t-\tau)^{-\frac12}\|{\nabla} Q(\tau)\|_{L^2}d\tau\\ \lesssim&\left(\delta_0^2+\delta_0^{\frac98}\Lambda^2(t)\right)\Bigg(\int_0^{\frac t 2} (1+t-\tau)^{-\frac74}(1+\tau)^{-\frac32}d\tau+\int_{\frac t 2}^t (1+t-\tau)^{-\frac12}(1+\tau)^{-\frac{11}4}d\tau\Big)\\ \lesssim&(1+t)^{-\frac74}\left(\delta_0^2+\delta_0^{\frac98}\Lambda^2(t)\right), \end{split} \end{equation} (5.10)

    Similarly, it holds that

    \begin{eqnarray*} \begin{split} &\|{\nabla} Q(t)\|_{L^2}\nonumber\\ \lesssim &\|(\widetilde n,\widetilde m)\|_{L^\infty}\big(\|({\nabla}\widetilde n,{\nabla}\widetilde m)\|_{L^2}+\|({\nabla}^2\widetilde n,{\nabla}^2\widetilde m)\|_{L^2}+\|( {\nabla} n_h,{\nabla} m_h)\|_{L^2}\nonumber \end{split} \end{eqnarray*}
    \begin{eqnarray*} \begin{split}&\quad+\|( {\nabla}^2 n_h,{\nabla}^2 m_h)\|_{L^2}\big)+\|({\nabla}\widetilde n,{\nabla}\widetilde m)\|_{L^\infty}\big(\|({\nabla}\widetilde n,{\nabla}\widetilde m)\|_{L^2}+\|(n_h,m_h)\|_{L^2}\\ &\quad+\|( {\nabla} n_h,{\nabla} m_h)\|_{L^2}\big)+\|( n_h,m_h)\|_{L^\infty}\big(\|( {\nabla}^2\widetilde n,{\nabla}^2\widetilde m)\|_{L^2}+\|( {\nabla} n_h,{\nabla} m_h)\|_{L^2}\\ &\quad+\|({\nabla}^2 n_h,{\nabla}^2 m_h)\|_{L^2}\big)+\|( {\nabla} n_h,{\nabla} m_h)\|_{L^\infty}\|( {\nabla} n_h,{\nabla} m_h)\|_{L^2}\\ \lesssim & (1+t)^{-\frac{11}4}\left(\delta_0^2+\delta_0^{\frac98}\Lambda^2(t)\right). \end{split} \end{eqnarray*}

    Thus, we also get that

    \begin{equation} \begin{split} &\|({\nabla}^2 n_h, {\nabla}^2 m_h)(t)\|_{L^2}\\ \lesssim& \int_0^{\frac t 2} \|({\nabla}^2 \mathfrak N, {\nabla}^2 \mathfrak M) (t-\tau)\ast H(\tau)\|_{L^2}d\tau\\ &\quad+\int_{\frac t 2}^t \|(\mathfrak N, \mathfrak M) (t-\tau)\ast {\nabla}^2 H(\tau)\|_{L^2}d\tau\\ \lesssim &\int_0^{\frac t 2}(1+t-\tau)^{-\frac94}\big(\|Q(\tau)\|_{L^1}+\|{\nabla}^2 Q(\tau)\|_{L^2}\big)d\tau\\ &\quad+\int_{\frac t 2}^t (1+t-\tau)^{-\frac12}\|{\nabla} ^2Q(\tau)\|_{L^2}d\tau\\ \lesssim& \left(\delta_0^2+\delta_0\Lambda(t)+\delta_0^{\frac34}\Lambda^2(t)\right)\bigg(\int_0^{\frac t 2} (1+t-\tau)^{-\frac94}(1+\tau)^{-\frac32}d\tau\\ &\quad+\int_{\frac t 2}^t (1+t-\tau)^{-\frac12}(1+\tau)^{-\frac{13}4}d\tau\bigg)\\ \lesssim&(1+t)^{-\frac94}\left(\delta_0^2+\delta_0\Lambda(t)+\delta_0^{\frac34}\Lambda^2(t)\right). \end{split} \end{equation} (5.11)

    Finally, we have

    \begin{eqnarray*} \begin{split} &\|{\nabla}^2 Q(t)\|_{L^2}\\ \lesssim &(\|(\widetilde n,\widetilde m)\|_{L^\infty}+\|(n_h,m_h)\|_{L^\infty})(\|({\nabla}^3\widetilde n,{\nabla}^3 \widetilde m)\|_{L^2}+\|({\nabla}^3n_h,{\nabla}^3 m_h)\|_{L^2})\\ &\quad+(\|({\nabla} \widetilde n,{\nabla}\widetilde m)\|_{L^\infty}+\|({\nabla} n_h,{\nabla} m_h)\|_{L^\infty}) (\|({\nabla}\widetilde n,{\nabla} \widetilde m)\|_{L^2}+\|({\nabla} n_h,{\nabla} m_h)\|_{L^2})\\ &\quad+(\|(\widetilde n,\widetilde m)\|_{L^\infty}+\|(n_h,m_h)\|_{L^\infty}+\|({\nabla} \widetilde n,{\nabla}\widetilde m)\|_{L^\infty}+\|({\nabla} n_h,{\nabla} m_h)\|_{L^\infty})\\ &\quad\times(\|({\nabla}^2\widetilde n,{\nabla}^2 \widetilde m)\|_{L^2}+\|({\nabla}^2 n_h,{\nabla}^2 m_h)\|_{L^2})\\ \lesssim&(1+t)^{-\frac{13}4}\left(\delta_0^2+\delta_0\Lambda(t)+\delta_0^{\frac34}\Lambda^2(t)\right). \end{split} \end{eqnarray*}

    In this subsection, we will close the a priori estimates and complete the proof of Proposition 5.2. For this purpose, we need to derive the time decay rate of higher order derivatives of (n_h,m_h) . We will establish the following lemma.

    Lemma 5.2. Under the assumption of Theorem 1.1, one has

    \begin{eqnarray*} \|{\nabla}^2 n(t)\|_{H^1}+\|{\nabla}^2 u(t)\|_{H^1}\lesssim (1+t)^{-\frac74}\left(\delta_0+\delta_0^{\frac34}\Lambda(t)\right). \end{eqnarray*}

    In particular, it holds that

    \begin{eqnarray*} \|{\nabla}^3 (n_h, m_h)(t)\|_{L^2}\lesssim (1+t)^{-\frac74}\left(\delta_0+\delta_0^{\frac34}\Lambda(t)\right). \end{eqnarray*}

    Proof. First of all, in view of (2.12), recovering the dissipation estimate for n , we see that

    \begin{equation} \begin{split} &\frac{d}{dt}\int_{{\mathop{\mathbb R\kern 0pt}\nolimits}^3} {\nabla}^2 u\cdot {\nabla}^3 n dx +C_1\|{\nabla}^3 n\|_{L^2}^2 dx\\ \leq &C_2\left(\|{\nabla}^3 u\|_{L^2}^2+\|{\nabla}^4 u\|_{L^2}^2\right)+C(1+t)^{-\frac{3}2}\left(\delta_0+\delta_0^{\frac38}\Lambda(t)\right)\\ &\quad\times\left(\|{\nabla}^2 n\|_{L^2}^2+\|{\nabla}^2 u\|_{L^2}^2+\|{\nabla}^3 u\|_{L^2}^2\right). \end{split} \end{equation} (5.12)

    Summing up (2.7) and (2.8) in the energy estimate for (n,u) , we can directly derive

    \begin{equation} \begin{split} &\frac{d}{dt}\int_{{\mathop{\mathbb R\kern 0pt}\nolimits}^3} \left(\gamma |{\nabla}^2 n|^2+|{\nabla}^2 u|^2 + \gamma|{\nabla}^3 n|^2 +|{\nabla}^3 u|^2 \right)dx + C_3\left(\|{\nabla}^3 u|^2 _{L^2}+\|{\nabla}^4 u\|^2 _{L^2}\right) \\ \leq &C(1+t)^{-\frac{3}2}\left(\delta_0+\delta_0^{\frac38}\Lambda(t)\right)\left(\|{\nabla}^2 n\|_{L^2}^2+\|{\nabla}^2 u\|_{L^2}^2+\|{\nabla}^3 n\|_{L^2}^2\right). \end{split} \end{equation} (5.13)

    Multiplying (5.12) by \epsilon_1\frac{C_3}{C_2} with \epsilon_1>0 a small but fixed constant, adding it with (5.13), we deduce that there exists a constant C_4>0 such that

    \begin{eqnarray*} \begin{split} &\frac{d}{dt}\bigg\{\sum\limits_{2\leq k\leq3}\left(\gamma \|{\nabla}^k n\|^2_{L^2}+\|{\nabla}^k u\|^2_{L^2} \right)+\epsilon_1\frac{C_3}{C_2}\int_{{\mathop{\mathbb R\kern 0pt}\nolimits}^3} {\nabla}^2 u\cdot {\nabla}^{3} n dx \bigg\}\\ &\quad+ C_4\Big(\|{\nabla}^{3} n\|_{L^2}^2+\sum\limits_{3\leq k\leq4}\|{\nabla}^{k} u\|^2_{L^2}\Big)\\ \leq &C(1+t)^{-\frac{3}2}\left(\delta_0+\delta_0^{\frac38}\Lambda(t)\right)\left(\|{\nabla}^2 n\|_{L^2}^2+\|{\nabla}^2 u\|_{L^2}^2\right). \end{split} \end{eqnarray*}

    Next, we define

    \mathcal E_1(t) = \bigg\{\sum\limits_{2\leq k\leq3}\left(\gamma \|{\nabla}^k n\|^2_{L^2}+\|{\nabla}^k u\|^2_{L^2} \right)+\epsilon_1\frac{C_3}{C_2}\int_{{\mathop{\mathbb R\kern 0pt}\nolimits}^3} {\nabla}^2 u\cdot {\nabla}^{3} n dx \bigg\}.

    Observe that since \epsilon_1\frac{C_3}{C_2} is small, then there exists a constant C_5>0 such that

    \begin{eqnarray*} C_5^{-1}\left(\|{\nabla}^2 n(t)\|^2_{H^1}+\|{\nabla}^2 u(t)\|^2_{H^1}\right) \leq\mathcal E_1(t)\leq C_5\left(\|{\nabla}^2 n(t)\|^2_{H^1}+\|{\nabla}^2 u(t)\|^2_{H^1}\right). \end{eqnarray*}

    Then we arrive at

    \begin{eqnarray*} \frac{d}{dt}\mathcal E_1(t)+C_4\Big(\|{\nabla}^{3} n(t)\|_{L^2}^2+\|{\nabla}^3 u(t)\|^2_{H^1}\Big) \leq C(1+t)^{-5}\left(\delta_0+\delta_0^{\frac38}\Lambda(t)\right)\left(\delta_0^2+\delta_0^{\frac32}\Lambda^2(t)\right). \end{eqnarray*}

    Denote S(t) = \Big\{\xi\big| |\xi| \leq \sqrt{\frac{3(1+\gamma)}{C_4}}(1+t)^{-\frac12}\Big\} the time-dependent n -dimensional sphere. This decomposition allows us to estimate L^2 time decay depend on (\widehat {n}, \widehat {u}) for frequency values \xi \in S(t) , then we obtain that

    \begin{eqnarray*} \begin{split} &\frac{C_4}{3}\|{\nabla}^{3} (n, u)(x)\|_{L^2}^2 \geq\frac{C_4}{3}\int_{S(t)^c} |\xi|^6|(\widehat{n}, \widehat{u})(\xi)|^2d\xi\\ \geq&(1+\gamma)(1+t)^{-1}\int_{{\mathop{\mathbb R\kern 0pt}\nolimits}^3} |\xi|^4|(\widehat{n}, \widehat{u})(\xi)|^2d\xi-(1+\gamma)(1+t)^{-1}\int_{S(t)} |\xi|^4|(\widehat{n}, \widehat{u})(\xi)|^2d\xi. \end{split} \end{eqnarray*}

    Hence we have

    \begin{eqnarray*} \begin{split} &\frac{d}{dt}\mathcal E_1(t)+(1+t)^{-1}\mathcal E_1(t)+\|{\nabla}^{3} n\|_{L^2}^2+\|{\nabla}^3 u\|^2_{H^1}\\ \lesssim&(1+t)^{-5}\left(\delta_0+\delta_0^{\frac38}\Lambda(t)\right)\left(\delta_0^2+\delta_0^{\frac32}\Lambda^2(t)\right)+(1+t)^{-1}\int_{S(t)} |\xi|^4|(\widehat{n}, \widehat{u})(\xi)|^2d\xi\\ &\quad+(1+t)^{-1}\int_{{\mathop{\mathbb R\kern 0pt}\nolimits}^3} {\nabla}^2 u\cdot {\nabla}^{3} n dx. \end{split} \end{eqnarray*}

    Multiplying the above equation by (1+t)^5 , we obtain that

    \begin{eqnarray*} \begin{split} &\frac{d}{dt}\Big\{(1+t)^5\mathcal E_1(t)\Big\}+(1+t)^5\Big(\|{\nabla}^{3} n\|_{L^2}^2+\|{\nabla}^3 u\|^2_{H^1}\Big) \lesssim(1+t)^{\frac12}\left(\delta_0^2+\delta_0^{\frac32}\Lambda^2(t)\right). \end{split} \end{eqnarray*}

    Integrating it with respect to time from 0 to T , then we have

    \begin{eqnarray*} \begin{split} &(1+t)^5\mathcal E_1(t)+\int_0^T(1+t)^5\Big(\|{\nabla}^{3} n\|_{L^2}^2+\|{\nabla}^3 u\|^2_{H^1}\Big)dt\\ \lesssim& \mathcal E_1(0)+(1+t)^{\frac32}\left(\delta_0^2+\delta_0^{\frac32}\Lambda^2(t)\right), \end{split} \end{eqnarray*}

    which implies that

    \begin{eqnarray*} \|{\nabla}^3 n\|^2_{L^2}+\|{\nabla}^3 u\|^2_{L^2}\lesssim\mathcal E_1(t)\lesssim (1+t)^{-5}\delta_0^2+(1+t)^{-\frac72}\left(\delta_0^2+\delta_0^{\frac32}\Lambda^2(t)\right). \end{eqnarray*}

    Finally, we have

    \begin{eqnarray*} \|{\nabla}^3 n_h\|_{L^2}+\|{\nabla}^3 m_h\|_{L^2}\lesssim (1+t)^{-\frac74}\left(\delta_0+\delta_0^{\frac34}\Lambda(t)\right). \end{eqnarray*}

    This completes the proof of this Lemma.

    In this subsection, we first combine the above a priori estimates of (5.8), (5.9), (5.10), (5.11) and Lemma 5.2 together to give the proof of the Proposition 5.2. In deed, for any t\in[0,T] , we have shown that

    \begin{equation} \Lambda(t)\leq C\left(\delta_0+\delta_0^{\frac14}\Lambda(t)+\Lambda^2(t)\right) \leq C\delta_0^{\frac34}. \end{equation} (5.14)

    With the help of standard continuity argument, Proposition 5.2 and the smallness of \delta_0>0 , implies that \Lambda(t)\leq C\delta_0^{\frac34} for any t>0 . Moreover, we deduce the time decay estimate for (n_h, m_h) from (5.9), (5.10), (5.11), Lemma 5.2 and (5.14) that

    \begin{eqnarray*} \begin{split} &\|({\nabla}^k n_h, {\nabla}^k m_h)\|_{L^2}\lesssim \delta_0^2(1+t)^{-\frac54-\frac k2},\quad k = 0,1,\\ &\|{\nabla}^2 (n_h, m_h)\|_{L^2}\lesssim\delta_0^{\frac74}(1+t)^{-\frac94},\quad \|{\nabla}^3 (n_h, m_h)\|_{L^2}\lesssim\delta_0(1+t)^{-\frac74}. \end{split} \end{eqnarray*}

    Consequently, for any t\in[0,T] we have

    \begin{equation} \Lambda(t)\leq C\delta_0. \end{equation} (5.15)

    From (5.11) and (5.15), thus we also get that

    \begin{eqnarray*} \|{\nabla}^2 (n_h, m_h)\|_{L^2}\lesssim\delta_0^2(1+t)^{-\frac94}. \end{eqnarray*}

    For {\nabla}^3 m_h , in view of the (5.6), we see that

    \begin{eqnarray*} \begin{split} &\|{\nabla}^3 m_h(t)\|_{L^2} \\\lesssim &\int_0^{\frac t 2}(1+t-\tau)^{-\frac{11}4}\big(\|Q(\tau)\|_{L^1}+\|{\nabla}^2 Q(\tau)\|_{L^2}\big)d\tau\\ &\quad+\int_{\frac t 2}^t (1+t-\tau)^{-\frac12}\|{\nabla} ^2Q(\tau)\|_{L^2}d\tau\\ \lesssim& \delta_0^2\bigg(\int_0^{\frac t 2} (1+t-\tau)^{-\frac{11}4}(1+\tau)^{-\frac32}d\tau+\int_{\frac t 2}^t (1+t-\tau)^{-\frac12}(1+\tau)^{-\frac{13}4}d\tau\bigg)\\ \lesssim&\delta_0^2(1+t)^{-\frac{11}4}. \end{split} \end{eqnarray*}

    Hence, we finish the proof of the Proposition 5.1. Theorem 1.1 follows.

    Y. Chen is partially supported by the China Postdoctoral Science Foundation under grant 2019M663198, Guangdong Basic and Applied Basic Research Foundation under grant 2019A1515110733, NNSF of China under grants 11801586, 11971496 and China Scholarship Council. The research of R. Pan is partially supported by National Science Foundation under grants DMS-1516415 and DMS-1813603, and by National Natural Science Foundation of China under grant 11628103. L. Tong's research is partially supported by China Scholarship Council.



    [1] R. L. Siegel, K. D. Miller, H. E. Fuchs, A. Jemal, Cancer statistics, 2021, CA Cancer J. Clin., 71 (2021), 7–33. https://doi.org/10.3322/caac.21654
    [2] S. Gul, M. S. Khan, A. Bibi, A. Khandakar, M. A. Ayari, M. E. H. Chowdhury, Deep learning techniques for liver and liver tumor segmentation: A review, Comput. Biol. Med., 147 (2022), 105620. https://doi.org/10.1016/j.compbiomed.2022.105620 doi: 10.1016/j.compbiomed.2022.105620
    [3] X. Shu, Y. Yang, B. Wu, Adaptive segmentation model for liver CT images based on neural network and level set method, Neurocomputing, 453 (2021), 438–452. https://doi.org/10.1016/j.neucom.2021.01.081 doi: 10.1016/j.neucom.2021.01.081
    [4] L. Soler, H. Delingette, G. Malandain, J. Montagnat, N. Ayache, C. Koehl, et al., Fully automatic anatomical, pathological, and functional segmentation from CT scans for hepatic surgery, Comput. Aided Surg., 6 (2010) 131–142. https://doi.org/10.3109/10929080109145999 doi: 10.3109/10929080109145999
    [5] X. Lu, J. Wu, X. Ren, B. Zhang, Y. Li, The study and application of the improved region growing algorithm for liver segmentation, Optik, 125 (2014), 2142–2147. https://doi.org/10.1016/j.ijleo.2013.10.049 doi: 10.1016/j.ijleo.2013.10.049
    [6] J. Wang, Y. Cheng, C. Guo, Y. Wang, S. Tamura, Shape-intensity prior level set combining probabilistic atlas and probability map constrains for automatic liver segmentation from abdominal CT images, Int. J. Comput. Assist. Radiol. Surg., 11 (2016), 817–826. https://doi.org/10.1007/s11548-015-1332-9 doi: 10.1007/s11548-015-1332-9
    [7] J. Long, E. Shelhamer, T. Darrell, Fully convolutional networks for semantic segmentation, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, (2015), 3431–3440. https://doi.org/10.1109/TPAMI.2016.2572683
    [8] O. Ronneberger, P. Fischer, T. Brox, U-Net: Convolutional networks for biomedical image segmentation, in Medical Image Computing and Computer-Assisted Intervention—MICCAI 2015, (2015), 234–241. https://doi.org/10.1007/978-3-319-24574-4_28
    [9] Z. Zhang, Q. Liu, Y. Wang, Road extraction by deep residual u-net, IEEE Geosci. Remote Sens. Lett., 15 (2018), 749–753. https://doi.org/10.1109/LGRS.2018.2802944 doi: 10.1109/LGRS.2018.2802944
    [10] K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, (2016), 770–778. https://doi.org/10.1109/CVPR.2016.90
    [11] O. Oktay, J. Schlemper, L. L. Folgoc, M. Lee, M. Heinrich, K. Misawa, et al., Attention U-net: Learning where to look for the pancreas, preprint, arXiv: 1804.03999. https://doi.org/10.48550/arXiv.1804.03999
    [12] C. Li, Y. Tan, W. Chen, X. Luo, Y. Gao, X. Jia, et al., Attention Unet++: A nested attention-aware U-net for liver CT image segmentation, in 2020 IEEE International Conference on Image Processing (ICIP), (2020), 345–349. https://doi.org/10.1109/ICIP40778.2020.9190761
    [13] Z. Zhou, M. M. Rahman Siddiquee, N. Tajbakhsh, J. Liang, UNet++: A Nested U-net architecture for medical image segmentation, in Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support, (2018), 3–11.
    [14] D. T. Kushnure, S. N. Talbar, MS-UNet: A multi-scale UNet with feature recalibration approach for automatic liver and tumor segmentation in CT images, Comput. Med. Imaging Graph., 89 (2021), 101885. https://doi.org/10.1016/j.compmedimag.2021.101885 doi: 10.1016/j.compmedimag.2021.101885
    [15] J. Wang, X. Zhang, P. Lv, L. Zhou, H. Wang, EAR-U-Net: EfficientNet and attention-based residual U-Net for automatic liver segmentation in CT, preprint, arXiv: 2110.01014. https://doi.org/10.48550/arXiv.2110.01014
    [16] P. Lv, J. Wang, H. Wang, 2.5D lightweight RIU-Net for automatic liver and tumor segmentation from CT, Biomed. Signal Process. Control, 75 (2022), 103567. https://doi.org/10.1016/j.bspc.2022.103567 doi: 10.1016/j.bspc.2022.103567
    [17] J. Wang, X. Zhang, L. Guo, C. Shi, S. Tamura, Multi-scale attention and deep supervision-based 3D UNet for automatic liver segmentation from CT, Math. Biosci. Eng., 20 (2023), 1297–1316. https://doi.org/10.3934/mbe.2023059 doi: 10.3934/mbe.2023059
    [18] T. Fan, G. Wang, X. Wang, Y. Li, H. Wang, MSN-Net: A multi-scale context nested U-Net for liver segmentation, Signal Image Video P, 15 (2021), 1089–1097. https://doi.org/10.1007/s11760-020-01835-9 doi: 10.1007/s11760-020-01835-9
    [19] P. Lv, J. Wang, X. Zhang, C. Ji, L. Zhou, H. Wang, An improved residual U-Net with morphological-based loss function for automatic liver segmentation in computed tomography, Math. Biosci. Eng., 19 (2022), 1426–1447. https://doi.org/10.3934/mbe.2022066 doi: 10.3934/mbe.2022066
    [20] J. D. L. Araújo, L. B. da Cruz, J. O. B. Diniz, J. L. Ferreira, A. C. Silva, A. C. de Paiva, et al., Liver segmentation from computed tomography images using cascade deep learning, Comput. Biol. Med., 140 (2022), 105095. https://doi.org/10.1016/j.compbiomed.2021.105095 doi: 10.1016/j.compbiomed.2021.105095
    [21] H. Huang, L. Lin, R. Tong, H. Hu, Q. Zhang, Y. Iwamoto, et al., Unet 3+: A full-scale connected Unet for medical image segmentation, in ICASSP 2020—2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), (2020), 1055–1059. https://doi.org/10.1109/ICASSP40776.2020.9053405
    [22] S. Sun, Z. Cao, D. Liao, R. Lv, A magnified adaptive feature pyramid network for automatic microaneurysms detection, Comput. Biol. Med., 139 (2021), 105000. https://doi.org/10.1016/j.compbiomed.2021.105000 doi: 10.1016/j.compbiomed.2021.105000
    [23] S. H. Gao, M. M. Cheng, K. Zhao, X. Y. Zhang, M. H. Yang, P. Torr, Res2Net: A new multi-scale backbone architecture, IEEE Trans. Pattern Anal. Mach. Intell., 43 (2019), 652–662. https://doi.org/10.1109/TPAMI.2019.2938758 doi: 10.1109/TPAMI.2019.2938758
    [24] J. Hu, L. Shen, G. Sun, Squeeze-and-excitation networks, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, (2018), 7132–7141. https://doi.org/10.1109/CVPR.2018.00745
    [25] A. Krizhevsky, I. Sutskever, G. E. Hinton, Imagenet classification with deep convolutional neural networks, Commun. ACM, 60 (2017), 84–90. https://doi.org/10.1145/3065386 doi: 10.1145/3065386
    [26] C. Szegedy, S. Ioffe, V. Vanhoucke, A. Alemi, Inception-v4, inception-resnet and the impact of residual connections on learning, in Proceedings of the AAAI Conference on Artificial Intelligence, 31 (2017). https://doi.org/10.1609/aaai.v31i1.11231
    [27] J. Wang, P. Lv, H. Wang, C. Shi, SAR-U-Net: Squeeze-and-excitation block and atrous spatial pyramid pooling based residual U-Net for automatic liver segmentation in computed tomography, Comput. Methods Programs Biomed., 208 (2021), 106268. https://doi.org/10.1016/j.cmpb.2021.106268 doi: 10.1016/j.cmpb.2021.106268
    [28] D. T. Kushnure, S. Tyagi, S. N. Talbar, LiM-Net: Lightweight multi-level multiscale network with deep residual learning for automatic liver segmentation in CT images, Biomed. Signal Process. Control, 80 (2023), 104305. https://doi.org/10.1016/j.bspc.2022.104305 doi: 10.1016/j.bspc.2022.104305
    [29] T. Fan, G. Wang, Y. Li, H. Wang, MA-Net: A multi-scale attention network for liver and tumor segmentation, IEEE Access, 8 (2020), 179656–179665. https://doi.org/10.1109/ACCESS.2020.3025372 doi: 10.1109/ACCESS.2020.3025372
    [30] D. P. Fan, G. P. Ji, T. Zhou, G. Chen, H. Fu, J. Shen, et al., PraNet: Parallel reverse attention network for polyp segmentation, in Medical Image Computing and Computer Assisted Intervention—MICCAI 2020, (2020), 263–273. https://doi.org/10.1007/978-3-030-59725-2_26
    [31] C. Y. Lee, S. Xie, P. Gallagher, Z. Zhang, Z. Tu, Deeply-supervised nets, in Proceedings of the Eighteenth International Conference on Artificial Intelligence and Statistics, 38 (2015), 562–570. https://doi.org/10.48550/arXiv.1409.5185
    [32] P. Bilic, P. F. Christ, E. Vorontsov, G. Chlebus, H. Chen, Q. Dou, et al., The liver tumor segmentation benchmark (lits), Med. Image Anal., 84 (2023), 102680. https://doi.org/10.1016/j.media.2022.102680 doi: 10.1016/j.media.2022.102680
    [33] C. Zhang, Q. Hua, Y. Chu, P. Wang, Liver tumor segmentation using 2.5D UV-Net with multi-scale convolution, Comput. Biol. Med., 133 (2021), 104–424. https://doi.org/10.1016/j.compbiomed.2021.104424 doi: 10.1016/j.compbiomed.2021.104424
    [34] Q. Jin, Z. Meng, C. Sun, H. Cui, R. Su, RA-UNet: A hybrid deep attention-aware network to extract liver and tumor in CT scans, Front. Bioeng. Biotechnol., 8 (2020), 1471. https://doi.org/10.3389/fbioe.2020.605132 doi: 10.3389/fbioe.2020.605132
    [35] J. Li, X. Ou, N. Shen, J. Sun, J. Ding, J. Zhang, et al., Study on strategy of CT image sequence segmentation for liver and tumor based on U-Net and Bi-ConvLSTM, Expert Syst. Appl., 180 (2021), 115008. https://doi.org/10.1016/j.eswa.2021.115008 doi: 10.1016/j.eswa.2021.115008
    [36] R. Bi, C. Ji, Z. Yang, M. Qiao, P. Lv, H. Wang, Residual based attention-Unet combing DAC and RMP modules for automatic liver tumor segmentation in CT, Math. Biosci. Eng., 19 (2022), 4703–4718. https://doi.org/10.3934/mbe.2022219 doi: 10.3934/mbe.2022219
    [37] S. Shao, X. Zhang, R. Cheng, C. Deng, Semantic segmentation method of 3D liver image based on contextual attention model, in 2021 IEEE International Conference on Systems, Man, and Cybernetics (SMC), (2021), 3042–3049. https://doi.org/10.1109/SMC52423.2021.9659018
  • Reader Comments
  • © 2023 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(3172) PDF downloads(205) Cited by(7)

Figures and Tables

Figures(12)  /  Tables(1)

/

DownLoad:  Full-Size Img  PowerPoint
Return
Return

Catalog