Anti-vascular endothelial growth factor (Anti-VEGF) therapy has become a standard way for choroidal neovascularization (CNV) and cystoid macular edema (CME) treatment. However, anti-VEGF injection is a long-term therapy with expensive cost and may be not effective for some patients. Therefore, predicting the effectiveness of anti-VEGF injection before the therapy is necessary. In this study, a new optical coherence tomography (OCT) images based self-supervised learning (OCT-SSL) model for predicting the effectiveness of anti-VEGF injection is developed. In OCT-SSL, we pre-train a deep encoder-decoder network through self-supervised learning to learn the general features using a public OCT image dataset. Then, model fine-tuning is performed on our own OCT dataset to learn the discriminative features to predict the effectiveness of anti-VEGF. Finally, classifier trained by the features from fine-tuned encoder as a feature extractor is built to predict the response. Experimental results on our private OCT dataset demonstrated that the proposed OCT-SSL can achieve an average accuracy, area under the curve (AUC), sensitivity and specificity of 0.93, 0.98, 0.94 and 0.91, respectively. Meanwhile, it is found that not only the lesion region but also the normal region in OCT image is related to the effectiveness of anti-VEGF.
Citation: Dehua Feng, Xi Chen, Xiaoyu Wang, Xuanqin Mou, Ling Bai, Shu Zhang, Zhiguo Zhou. Predicting effectiveness of anti-VEGF injection through self-supervised learning in OCT images[J]. Mathematical Biosciences and Engineering, 2023, 20(2): 2439-2458. doi: 10.3934/mbe.2023114
[1] | Dan-Ni Chen, Jing Cheng, Xiao Shen, Pan Zhang . Semi-stable quiver bundles over Gauduchon manifolds. AIMS Mathematics, 2023, 8(5): 11546-11556. doi: 10.3934/math.2023584 |
[2] | Fengxia Zhang, Ying Li, Jianli Zhao . The semi-tensor product method for special least squares solutions of the complex generalized Sylvester matrix equation. AIMS Mathematics, 2023, 8(3): 5200-5215. doi: 10.3934/math.2023261 |
[3] | Shu-Xin Miao, Jing Zhang . On Uzawa-SSI method for non-Hermitian saddle point problems. AIMS Mathematics, 2020, 5(6): 7301-7315. doi: 10.3934/math.2020467 |
[4] | Jiao Xu, Hairui Zhang, Lina Liu, Huiting Zhang, Yongxin Yuan . A unified treatment for the restricted solutions of the matrix equation AXB=C. AIMS Mathematics, 2020, 5(6): 6594-6608. doi: 10.3934/math.2020424 |
[5] | Yuezhen Ren, Ruihu Li, Guanmin Guo . New entanglement-assisted quantum codes constructed from Hermitian LCD codes. AIMS Mathematics, 2023, 8(12): 30875-30881. doi: 10.3934/math.20231578 |
[6] | Shahroud Azami . Monotonicity of eigenvalues of Witten-Laplace operator along the Ricci-Bourguignon flow. AIMS Mathematics, 2017, 2(2): 230-243. doi: 10.3934/Math.2017.2.230 |
[7] | Lv Zhang, Qingbiao Wu . Modified Newton-EHS method for solving nonlinear problems with complex symmetric Jacobian matrices. AIMS Mathematics, 2023, 8(10): 24233-24253. doi: 10.3934/math.20231236 |
[8] | Reena Jain, Hemant Kumar Nashine, Jung Rye Lee, Choonkil Park . Unified relational-theoretic approach in metric-like spaces with an application. AIMS Mathematics, 2021, 6(8): 8959-8977. doi: 10.3934/math.2021520 |
[9] | Ghodratallah Fasihi-Ramandi . Hamilton’s gradient estimate for fast diffusion equations under geometric flow. AIMS Mathematics, 2019, 4(3): 497-505. doi: 10.3934/math.2019.3.497 |
[10] | Sourav Shil, Hemant Kumar Nashine . Positive definite solution of non-linear matrix equations through fixed point technique. AIMS Mathematics, 2022, 7(4): 6259-6281. doi: 10.3934/math.2022348 |
Anti-vascular endothelial growth factor (Anti-VEGF) therapy has become a standard way for choroidal neovascularization (CNV) and cystoid macular edema (CME) treatment. However, anti-VEGF injection is a long-term therapy with expensive cost and may be not effective for some patients. Therefore, predicting the effectiveness of anti-VEGF injection before the therapy is necessary. In this study, a new optical coherence tomography (OCT) images based self-supervised learning (OCT-SSL) model for predicting the effectiveness of anti-VEGF injection is developed. In OCT-SSL, we pre-train a deep encoder-decoder network through self-supervised learning to learn the general features using a public OCT image dataset. Then, model fine-tuning is performed on our own OCT dataset to learn the discriminative features to predict the effectiveness of anti-VEGF. Finally, classifier trained by the features from fine-tuned encoder as a feature extractor is built to predict the response. Experimental results on our private OCT dataset demonstrated that the proposed OCT-SSL can achieve an average accuracy, area under the curve (AUC), sensitivity and specificity of 0.93, 0.98, 0.94 and 0.91, respectively. Meanwhile, it is found that not only the lesion region but also the normal region in OCT image is related to the effectiveness of anti-VEGF.
Let (M,ω) be a complex n-dimensional compact Hermitian manifold and χ be a smooth real (1, 1)-form on (M,ω). Γωk is the set of all real (1, 1)-forms whose eigenvalues belong to the k-positive cone Γk. For any u∈C2(M), we can get a new (1, 1)-form
χu:=χ+√−1∂¯∂u. |
In any local coordinate chart, χu can be expressed as
χu=√−1(χi¯j+ui¯j)dzi∧d¯zj. |
In this article, we study the following form of parabolic Hessian quotient equations
{∂u(x,t)∂t=logCknχku∧ωn−kClnχlu∧ωn−l−logϕ(x,u), (x,t)∈M×[0,T),u(x,0)=u0(x), x∈M, | (1.1) |
where 0≤l<k≤n, [0,T) is the maximum time interval in which the solution exists and ϕ(x,z)∈C∞(M×R) is a given strictly positive function.
The study of the parabolic flows is motivated by complex equations
χku∧ωn−k=ClnCknϕ(x,u)χlu∧ωn−l, χu∈Γωk. | (1.2) |
Equation (1.2) include some important geometry equations, for example, complex Monge-Ampère equation and Donaldson equation [6], which have attracted extensive attention in mathematics and physics since Yau's breakthrough in the Calabi conjecture [28]. Since Eq (1.2) are fully nonlinear elliptic, a classical way to solve them is the continuity method. Using this method, the complex Monge-Ampère equation
χnu=ϕ(x)ωn, χu∈Γωn |
was solved by Yau [28]. Donaldson equation
χnu=∫Mχn∫Mχ∧ωn−1χu∧ωn−1, χu∈Γωn |
was independently solved by Li-Shi-Yao [11], Collins-Szèkelyhidi [3] and Sun [17]. Equation (1.2) also include the complex k-Hessian equation and complex Hessian quotient equation, which, respectively, correspond to
Cknχku∧ωn−k=ϕ(x)ωn, χu∈Γωk, |
χku∧ωn−k=ClnCknϕ(x)χlu∧ωn−l, χu∈Γωk. |
Dinew and Kolodziej [7] proved a Liouville type theorem for m-subharmonic functions in Cn, and combining with the estimate of Hou-Ma-Wu [10], solved the complex k-Hessian equation by using the continuity method. Under the cone condition, Sun [16] solved the complex Hessian quotient equation by using the continuity method. There have been many extensive studies for complex Monge-Ampère equation, Donaldson equation, the complex k-Hessian equation and the complex Hessian quotient equation on closed complex manifolds, see, e.g., [4,12,20,22,23,29,30]. When the right hand side function ϕ in Eq (1.2) depends on u, that is ϕ=ϕ(x,u), it is interesting to ask whether we can solve them. We intend to solve (1.2) by the parabolic flow method.
Equation (1.1) covers some of the important geometric flows in complex geometry. If k=n and l=0, (1.1) is known as the complex Monge-Ampère flow
∂u(x,t)∂t=logχnuωn−logϕ(x), (x,t)∈M×[0,T), |
which is equivalent to the Kähler-Ricci flow. The result of Yau [28] was reproduced by Cao [2] through Kähler-Ricci flow. Using the complex Monge-Ampère flow, similar results on a compact Hermitian manifold and a compact almost Hermitian manifold were proved by Gill [9] and Chu [5], respectively. To study the normalized twisted Chern-Ricci flow
∂ωt∂t=−Ric(ωt)−ωt+η, |
which is equivalent to the following Mong-Ampère flow
∂φ∂t=log(θt+ddcφ)nΩ−φ, |
Tô [25,26] considered the following complex Monge-Ampère flow
∂φ∂t=log(θt+ddcφ)nΩ−F(t,x,φ), |
where Ω is a smooth volume form on M. From this, we can see that the given function ϕ depends on u in some geometric flows. If l=0, (1.1) is called as the complex k-Hessian flow
∂u(x,t)∂t=logCknχku∧ωn−kωn−logϕ(x,u), (x,t)∈M×[0,T). |
The solvability of complex k-Hessian flow was showed by Sheng-Wang [21].
In this paper, our research can be viewed as a generalization of Tô's work in [26] and Sheng-Wang's work in [21]. To solve the complex Hessian quotient flow, the condition of the parabolic C-subsolution is needed. According to Phong and Tô [14], we can give the definition of the parabolic C-subsolution to Eq (1.1).
Definition 1.1. Let u_(x,t)∈C2,1(M×[0,T)) and χu_∈Γωk, if there exist constants δ, R>0, such that for any (x,t)∈M×[0,T),
logσk(λ)σl(λ)−∂tu_≤logϕ(x,u_), λ−λ(u_)+δI∈Γn, |
implies that
|λ|<R, |
then u_ is said to be a parabolic C-subsolution of (1.1), where λ(u_) denotes eigenvalue set of χu_.
Obviously, we can give the equivalent definition of parabolic C-subsolution of (1.1).
Definition 1.2. Let u_(x,t)∈C2,1(M×[0,T)) and χu_∈Γωk, if there exist constant ˜δ>0, for any (x,t)∈M×[0,T), such that
limμ→∞logσk(λ(u_)+μei)σl(λ(u_)+μei)>∂u_∂t+˜δ+logϕ(x,u_), 1≤i≤n, | (1.3) |
then u_ is said to be a parabolic C-subsolution of (1.1).
Our main result is
Theorem 1.3. Let (M,g) a compact Hermitian manifold and χ be a smooth real (1,1)-form on M. Assume there exists a parabolic C-subsolutionu_ for Eq (1.1) and
∂tu_≥max{supM(logσk(λ(u0))σl(λ(u0))−logϕ(x,u0)),0}, | (1.4) |
ϕz(x,z)ϕ>cϕ>0, | (1.5) |
where cϕ is a constant.Then there exits a unique smooth solution u(x,t) to (1.1) all timewith
supx∈M(u0(x)−u_(x,0))=0. | (1.6) |
Moreover, u(x,t) isC∞ convergent to a smooth function u∞, which solves Eq (1.2).
The rest of this paper is organized as follows. In Section 2, we give some important lemmas and estimate on |ut(x,t)|. In Section 3, we prove C0 estimates of Eq (1.1) by the parabolic C-subsolution condition and the Alexandroff-Bakelman-Pucci maximum principle. In Section 4, using the parabolic C-subsolution condition, we establish the C2 estimate for Eq (1.1) by the method of Hou-Ma-Wu [10]. In Section 5, we adapt the blowup method of Dinew and Kolodziej [7] to obtain the gradient estimate. In Section 6, we give the proof of the long-time existence of the solution to the parabolic equation and its convergence, that is Theorem 1.3.
In this section, we give some notations and lemmas. In holomorphic coordinates, we can set
ω=√−1gi¯jdzi∧d¯zj=√−1δijdzi∧d¯zj, χ=√−1χi¯jdzi∧d¯zj, |
χu=√−1(χi¯j+ui¯j)dzi∧d¯zj=√−1Xi¯jdzi∧d¯zj, |
χu_=√−1(χi¯j+u_i¯j)dzi∧d¯zj=√−1X_i¯jdzi∧d¯zj. |
λ(u) and λ(u_) denote the eigenvalue set of {Xi¯j} and {X_i¯j} with respect to {gi¯j}, respectively. In local coordinates, (1.1) can be written as
∂tu=logσk(λ(u))σl(λ(u))−logϕ(x,u). | (2.1) |
For simplicity, set
F(λ(u))=logσk(λ(u))σl(λ(u)), |
then (2.1) is abbreviated as
∂tu=F(λ(u))−logϕ(x,u). | (2.2) |
We use the following notation
Fi¯j=∂F∂Xi¯j, F=∑iFi¯i, Fi¯j,p¯q=∂2F∂Xi¯j∂Xp¯q. |
For any x0∈M, we can choose a local holomorphic coordinates such that the matrix {Xi¯j} is diagonal and X1¯1≥⋯≥Xn¯n, then we have, at x0∈M,
λ(u)=(λ1,⋯,λn)=(X1¯1,⋯,Xn¯n), |
Fi¯j=Fi¯iδij=(σk−1(λ|i)σk−σl−1(λ|i)σl)δij, F1¯1≤⋯≤Fn¯n. |
To prove a priori C0-estimate for solution to Eq (1.1), we need the following variant of the Alexandroff-Bakelman-Pucci maximum principle, which is Proposition 10 in [20].
Lemma 2.1. [20] Let v:B(1)→R be a smooth function, which meets the conditionv(0)+ϵ≤inf∂B(1)v, whereB(1) denotes the unit ball in Rn.Define the set
Ω={x∈B(1):|Dv(x)|<ϵ2, and v(y)≥v(x)+Dv(x)⋅(y−x),∀y∈B(1)}. |
Then there exists a costant c0>0 such that
c0ϵn≤∫Ωdet(D2v). |
Next, we give an estimate on |ut(x,t)|.
Lemma 2.2. Under the assumption of Theorem 1.3, let u(x,t) be a solution to (1.1). Then for any (x,t)∈M×[0,T), we have
min{infMut(x,0),0}≤ut(x,t)≤max{supMut(x,0),0}. | (2.3) |
Furthermore, there is a constant C>0 such that
supM×[0,T)|∂tu(x,t)|≤supM|∂tu(x,0)|≤C, |
where C depends on H=|u0|C2(M) and |ϕ|C0(M×[−H,H]).
Proof. Differentiating (2.2) on both sides simultaneously at t, we obtain
(ut)t=Fi¯jXi¯jt−ϕzϕut=Fi¯j(ut)i¯j−ϕzϕut. | (2.4) |
Set uεt=ut−εt, ε>0. For any T′∈(0,T), suppose uεt achieves its maximum Mt at (x0,t0)∈M×[0,T′]. Without loss of generality, we may suppose Mt≥0. If t0>0, From the parabolic maximum principle and (2.4), we get
0≤(uεt)t−Fi¯j(uεt)i¯j+ϕzϕuεt≤(ut)t−ε−Fi¯j(ut)i¯j+ϕzϕut−εϕzϕt0≤−ε−εϕzϕt0. |
This is obviously a contradiction, so t0=0 and
supM×[0,T′]uεt(x,t)=supMut(x,0), |
that is
supM×[0,T′]ut(x,t)=supM×[0,T′](uεt(x,t)+εt)≤supMut(x,0)+εT′. |
Letting ε→0, we obtain
supM×[0,T′]ut(x,t)≤supMut(x,0). |
Since T′∈(0,T) is arbitrary, we have
supM×[0,T)ut(x,t)≤supMut(x,0). | (2.5) |
Similarly, setting uεt=ut+εt, ε>0, we obtain
infM×[0,T)ut(x,t)≥infMut(x,0). | (2.6) |
(2.1) yields
|ut(x,0)|=|logσk(λ(u0))σl(λ(u0))−logϕ(x,u0)|≤C. | (2.7) |
Combining (2.5)–(2.7), we complete the proof of Proposition 2.2.
From the concavity of F(λ(u)) and the condition of the parabolic C-subsolution, we give the following lemma, which plays an important role in the estimation of C2.
Lemma 2.3. Under the assumption of Theorem 1.3 and assuming that X1¯1≥⋯≥Xn¯n, there exists two positive constantsN and θ such that we have either
Fi¯i(u_i¯i−ui¯i)−∂t(u_−u)≥θ(1+F) | (2.8) |
or
F1¯1≥θN(1+F). | (2.9) |
Proof. Since u_ is a parabolic C-subsolution to Eq (1.1), from Definition 1.2, there are uniform constants ˜δ>0 and N>0, such that
logσk(λ(u_)+Ne1)σl(λ(u_)+Nue1)>∂u_∂t+˜δ+logϕ(x,u_). | (2.10) |
If ϵ>0 is sufficiently small, it can be obtained from (2.10)
logσk(λ(u_)−ϵI+Ne1)σl(λ(u_)−ϵI+Nue1)≥∂u_∂t+˜δ+logϕ(x,u_). |
Set λ′=λ(u_)−ϵI+Ne1, then
F(λ′)≥∂u_∂t+˜δ+logϕ(x,u_). | (2.11) |
Using the concavity of F(λ(u)) gives
Fi¯i(u_i¯i−ui¯i)=Fi¯i({X_i¯i−Xi¯i)=Fi¯i({X_i¯i−ϵδii+Nδi1−Xi¯i)+ϵF−NF1¯1≥F(λ′)−F(λ(u))+ϵF−NF1¯1. | (2.12) |
From Lemma 2.2 and (1.4), we obtain
u_t(x,t)≥ut(x,t), ∀(x,t)∈M×[0,T). | (2.13) |
In addition, it can be obtained from the condition (1.6)
u_(x,0)≥u(x,0), ∀x∈M×[0,T). | (2.14) |
(2.13) and (2.14) deduce that
u_(x,t)≥u(x,t), ∀(x,t)∈M×[0,T). | (2.15) |
It follows from this that
ϕ(x,u_)≥ϕ(x,u). | (2.16) |
Combining (2.2), (2.11) and (2.16) gives that
F(λ′)−F(λ(u))≥u_t(x,t)−ut(x,t)+˜δ+logϕ(x,u_)−logϕ(x,u)≥u_t(x,t)−ut(x,t)+˜δ. | (2.17) |
Put (2.17) into (2.12)
Fi¯i(u_i¯i−ui¯i)≥u_t(x,t)−ut(x,t)+˜δ+ϵF−NF1¯1≥˜δ+ϵF−NF1¯1. |
Let
θ=min{˜δ2, ϵ2}. |
If F1¯1N≤θ(1+F), Inequality (2.8) is obtained, otherwise Inequality (2.9) must be true.
In this section, we prove the C0 estimates by the existence of the parabolic C-subsolution and the Alexandroff-Bakelman-Pucci maximum principle.
Proposition 3.1. Under the assumption of Theorem1.3, let u(x,t) be a solution to (1.1). Then there exists a constant C>0 such that
|u(x,t)|C0(M×[0,T))≤C, |
where C depends on |u0|C2(M) and |u_|C2(M×[0,T)).
Proof. Combining (2.13), (2.14) and ∂ϕ(x,z)∂z≥0 yields
u_t(x,t)+logϕ(x,u_)≥ut(x,t)+logϕ(x,u). | (3.1) |
Let's rewrite Eq (2.2) as
F(λ(u))=∂tu+logϕ(x,u). | (3.2) |
when fix t∈[0,T), Eq (3.2) is elliptic. From (3.1), we see that the parabolic C-subsolution u_(x,t) is a C-subsolution to Eq (3.2) in the elliptic sense. From (2.15), we have
supM×[0,T)(u−u_)=0. |
Our goal is to obtain a lower bound for L=infM×t(u−u_). Note that λ(u)∈Γk, which implies that λ(u)∈Γ1, then Δ(u−u_)≥−˜C, where Δ is the complex Laplacian with respect to ω. According to Tosatti-Weinkove's method [22], we can prove that ‖u−u_‖L1(M) is bounded uniformly. Let G:M×M→R be the associated Green's function, then, by Yau [28], there is a uniform constant K such that
G(x,y)+K≥0, ∀(x,y)∈M×M, and ∫y∈MG(x,y)ωn(y)=0. |
Since
supM×[0,T)(u−u_)=0, |
then for fixed t∈[0,T) there exists a point x0∈M such that (u−u_)(x0,t)=0. Thus
(u−u_)(x0,t)=∫M(u−u_)dμ−∫y∈MG(x0,y)Δ(u−u_)(y)ωn(y)=∫M(u−u_)dμ−∫y∈M(G(x0,y)+K)Δ(u−u_)(y)ωn(y)≤∫M(u−u_)dμ+˜CK∫Mωn, |
that is
∫M(u_−u)dμ=∫M|(u−u_)|dμ≤˜CK∫Mωn. |
Let us work in local coordinates, for which the infimum L is achieved at the origin, that is L=u(0,t)−u_(0,t). We write B(1)={z:|z|<1}. Let v=u−u_+ϵ|z|2, for a small ϵ>0. We have infv=L=v(0), and v(z)≥L+ϵ for z∈∂B(1). From Lemma 2.1, we get
c0ϵ2n≤∫Ωdet(D2v). | (3.3) |
At the same time, if x∈Ω, then D2v(x)≥0 implies that
ui¯j(x)−u_i¯j(x)+ϵδij≥0. |
If ϵ is sufficiently small, then
λ(u)∈λ(u_)−δI+Γn. |
Set μ=λ(u)−λ(u_). Since λ(u) satisfies Eq (3.2), then
F(λ(u_)+μ)=∂tu+logϕ(x,u), μ+δI∈Γn. | (3.4) |
u_ is a C-subsolution to Eq (3.2) in the elliptic sense, so there is a uniform constant R>0, such that
|μ|≤R, |
which means |vi¯j|≤C, for any x∈Ω. As in Blocki [1], for x∈Ω, we have D2v(x)≥0 and so
D2v(x)≤22ndet(vi¯j)2≤C′. |
From this and (3.3), we obtain
c0ϵ2n≤∫Ωdet(D2v)≤C′⋅vol(Ω). | (3.5) |
On the other hand, by the definition of Ω in Lemma 2.1, for x∈Ω, we get
v(0)≥v(x)−Dv(x)⋅x>v(x)−ϵ2, |
and so
|v(x)|>|L+ϵ2|. |
It follows that
∫M|v(x)|≥∫Ω|v(x)|≥|L+ϵ2|⋅vol(Ω). | (3.6) |
Since ‖u−u_‖L1(M) is bounded uniformly, ∫M|v(x)| is also bounded uniformly. If L is very large, Inequality (3.6) contradicts (3.5), which means that L has a lower bound. For any t∈[0,T), Inequality (3.1) holds, thus
|u(x,t)|C0(M×[0,T))≤|L|+supM×[0,T)|u_|≤C. |
In this section, we prove that the second-order estimates are controlled by the square of the gradient estimate linearly. Our calculation is a parabolic version of that in Hou-Ma-Wu [10].
Proposition 4.1. Under the assumption of Theorem1.3, let u(x,t) be a solution to (1.1). Then there exists a constant ˜C such that
supM×[0,T)|√−1∂¯∂u|≤˜C(supM×[0,T)|∇u|2+1), |
where ˜C depends χ, ω, |ϕ|C2(M×[−C,C]), |u_|C2(M×[0,T)), |∂tu_|C0(M×[0,T)) and |u0|C2(M).
Proof. Let λ(u)=(λ1,…,λn) and λ1 is the maximum eigenvalue. For any T′<T, we consider the following function
W(x,t)=logλ1+φ(|∇u(x,t)|2)+ψ(u(x,t)−u_(x,t)), (x,t)∈M×[0,T′], | (4.1) |
where φ and ψ are determined later. We want to apply the maximum principle to the function W. Since the eigenvalues of the matrix {Xi¯j} with respect to ω need not be distinct at the point where W achieves its maximum, we will perturb {Xi¯j} following the technique of [20]. Let W achieve its maximum at (x0,t0)∈M×[0,T′]. Near (x0,t0), we can choose local coordinates such that {Xi¯j} is diagonal with X1¯1≥⋯≥Xn¯n, and λ(u)=(X1¯1,⋯,Xn¯n). Let D be a diagonal matrix such that D11=0 and 0<D22<⋯<Dnn are small, satisfying Dnn<2D22. Define the matrix ˜X=X−D. At (x0,t0), ˜X has eigenvalues
˜λ1=λ1, ˜λi=λi−Dii, n≥i≥2. |
Since all the eigenvalues of ˜X are distinct, we can define near (x0,t0) the following smooth function
˜W=log˜λ1+φ(|∇u|2)+ψ(u−u_), | (4.2) |
where
φ(s)=−12log(1−s2K), 0≤s≤K−1, |
ψ(s)=−Elog(1+s2L), −L+1≤s≤L−1, |
K=supM×[0,T′]|∇u|2+1, |
L=supM×[0,T′]|u|+supM×[0,T′]|u_|+1, |
E=2L(C1+1), |
and C1>0 is to be chosen later. Direct calculation yields
0<14K≤φ′≤12K, φ″=2(φ′)2>0, | (4.3) |
and
C1+1≤−ψ′≤2(C1+1), ψ″≥4ϵ1−ϵ(ψ′)2,∀ ϵ≤14E+1. | (4.4) |
Without loss of generality, we can assume that λ1>1. From here on, all calculations are done at (x0,t0). From the maximum principle, calculating the first and second derivatives of the function ˜W gives
0=˜Wi=˜λ1,iλ1+φ′(|∇u|2)i+ψ′(u−u_)i,1≤i≤n, | (4.5) |
0≥˜Wi¯i=˜λ1,i¯iλ1−˜λ1,i˜λ1,¯iλ21+φ′(|∇u|2)i¯i+φ″|(|∇u|2)i|2+ψ′(u−u_)i¯i+ψ″|(u−u_)i|2. | (4.6) |
0≤˜Wt=˜λ1,tλ1+φ′(|∇u|2)t+ψ′(u−u_)t. | (4.7) |
Define
L:=Fi¯j∇∂∂¯zj∇∂∂zi−∂t. |
Obviously,
0≥L˜W=Llog˜λ1+Lφ(|∇u|2)+Lψ(u−u_). | (4.8) |
Next, we will estimate the terms in (4.8). Direct calculation shows that
Llog˜λ1=Fi¯i˜λ1,i¯iλ1−Fi¯i|˜λ1,i|2λ21−˜λ1,tλ1. | (4.9) |
According to Inequality (78) in [20], we have
˜λ1,i¯i≥Xi¯i1¯1−2Re(Xi¯11¯T1i1)−C0λ1, | (4.10) |
where C0 depending χ, ω, |ϕ|C2(M×[−C,C]), |u_|C2(M×[0,T)), |∂tu_|C0(M×[0,T)) and |u0|C2(M)). From here on, C0 can always absorb the constant it represents before, and can change from one line to the next, but it does not depend on the parameter we choose later. By calculating the covariant derivatives of (4.7) in the direction ∂∂z1 and ∂∂¯z1, we obtain
ut1=Fi¯iXi¯i1−(logϕ)1−(logϕ)uu1, | (4.11) |
and
ut1¯1=Fi¯j,p¯qXi¯j1Xp¯q¯1+Fi¯iXi¯i1¯1−(logϕ)1¯1−(logϕ)1uu¯1−(logϕ)u¯1u1−(logϕ)uu|u1|2−(logϕ)uu1¯1. | (4.12) |
Notice that
X1¯1i=χ1¯1i+u1¯1i=(χ11i−χi11+Tpi1χp¯1)+Xi¯11−T1i1λ1, | (4.13) |
therefore
|X1¯1i|2≤|Xi¯11|2−2λ1Re(Xi¯11¯T1i1)+C0(λ21+|X1¯1i|). | (4.14) |
Combining (4.14) with
˜λ1,i=X1¯1i−(D11)i, |
gives
−Fi¯i|˜λ1,i|2λ21=−Fi¯i|X1¯1i|2λ21+2λ21Fi¯iRe(X1¯1i(D11)¯i)−Fi¯i|(D11)i|2λ2i≥−Fi¯i|X1¯1i|2λ21−C0λ21Fi¯i|X1¯1i|−C0F≥−Fi¯i|Xi¯11|2λ21+2λ1Fi¯iRe(Xi¯11¯T1i1)−C0λ21Fi¯i|X1¯1i|−C0F. | (4.15) |
Let's set λ1≥K, otherwise the proof is completed. Putting 4.10–4.12 and (4.15) into (4.9) yields
Llog˜λ1≥−Fi¯j,p¯qXi¯j1Xp¯q¯1λ1−Fi¯i|Xi¯11|2λ21−C0λ1Fi¯i|X1¯1i|λ1−C0F+(logϕ)1¯1+(logϕ)1uu¯1+(logϕ)u¯1u1+(logϕ)uu|u1|2−(logϕ)uχ1¯1λ1≥−Fi¯j,p¯qXi¯j1Xp¯q¯1λ1−Fi¯i|Xi¯11|2λ21−C0λ1Fi¯i|X1¯1i|λ1−C0F−C0 | (4.16) |
A simple computation gives
Lφ(|∇u|2)=φ′Fi¯i(|∇u|2)i¯i+φ″Fi¯i|(|∇u|2)i|2−φ′(|∇u|2)t. | (4.17) |
Next, we estimate the formula (4.17). Differentiating Eq (2.2), we have
(ut)p=Fi¯iXi¯ip−(logϕ)p−(logϕ)uup, | (4.18) |
and
(ut)¯p=Fi¯iXi¯i¯p−(logϕ)¯p−(logϕ)uu¯p. | (4.19) |
It follows from (4.18) and (4.19) that
∂t|∇u|2=∑putpu¯p+∑pupu¯t¯p=Fi¯i∑p(Xi¯ipu¯p+Xi¯i¯pup)−∑p(logϕ)pu¯p−∑p(logϕ)¯pup−2∑p(logϕ)u|∇u|2. | (4.20) |
By commuting derivatives, we have the identity
Fi¯iXi¯ip=Fi¯iui¯ip+Fi¯iχi¯ip=Fi¯iupi¯i−Fi¯iTqpiuq¯i−Fi¯iuqR qi¯ip+Fi¯iχi¯ip. | (4.21) |
Direct calculation gives
Fi¯i(|∇u|2)i¯i=∑pFi¯i(upi¯iu¯p+u¯pi¯iup)+∑pFi¯i(upiu¯p¯i+up¯iu¯pi). | (4.22) |
It follows from (4.21) that
Fi¯iupi¯iu¯p−Fi¯iXi¯ipu¯p=Fi¯iTqpiuq¯iu¯p+Fi¯iu¯puqR qi¯ip−Fi¯iχi¯ipu¯p≥−C0K12Fi¯iXi¯i−C0K12F−C0KF. | (4.23) |
Noticing that
Fi¯iXi¯i=σk−1(λ|i)σkλi−σl−1(λ|i)σlλi=k−l, | (4.24) |
from this and (4.23), we obtain
Fi¯iupi¯iu¯p−Fi¯iXi¯ipu¯p≥−C0K12−C0K12F−C0KF. | (4.25) |
In the same way, we can get
Fi¯iu¯pi¯iup−Fi¯iXi¯i¯pup≥−C0K12−C0K12F−C0KF. | (4.26) |
Using (4.20)–(4.26) in (4.17), we have
Lφ(|∇u|2)≥φ″Fi¯i|(|∇u|2)i|2+φ′(−C0K12−C0K12F−C0KF)+φ′∑p((logϕ)pu¯p+(logϕ)¯pup+2(logϕ)u|∇u|2)+∑pFi¯i(|upi|2+|u¯pi|2)≥φ″Fi¯i|(|∇u|2)i|2+∑pFi¯i(|upi|2+|u¯pi|2)−C0−C0F. | (4.27) |
A simple calculation gives
Lψ(u−u_)=ψ″Fi¯i|(u−u_)i|2+ψ′[Fi¯i(u−u_)i¯i−∂t(u−u_)]. | (4.28) |
Substituting (4.27), (4.28) and (4.16) into (4.8),
0≥−Fi¯j,p¯qXi¯j1Xp¯q¯1λ1−Fi¯i|Xi¯11|2λ21−C0λ1Fi¯i|X1¯1i|λ1+φ″Fi¯i|(|∇u|2)i|2+φ′∑pFi¯i(|upi|2+|u¯pi|2)−C0−C0F+ψ″Fi¯i|(u−u_)i|2+ψ′[Fi¯i(u−u_)i¯i−∂t(u−u_)]. | (4.29) |
Let δ>0 be a sufficiently small constant to be chosen later and satisfy
δ≤min{11+4E,12}. | (4.30) |
We separate the rest of the calculations into two cases.
Case 1: λn<−δλ1.
Using(4.5), we find that
−Fi¯i|X1¯1i|2λ21=−Fi¯i|φ′(|∇u|2)i+ψ′(u−u_)i−(D11)iλ1|2≥−2(φ′)2Fi¯i|(|∇u|2)i|2−2Fi¯i|ψ′(u−u_)i−(D11)iλ1|2≥−2(φ′)2Fi¯i|(|∇u|2)i|2−C0|ψ′|2KF−C0F. | (4.31) |
From (4.13), we have
|Xi¯11|2λ21≤|X1¯1i|2λ21+C0(1+|X1¯1i|λ1). | (4.32) |
Combining (4.31) with (4.32), we conclude that
−Fi¯i|Xi¯11|2λ21≥−2(φ′)2Fi¯i|(|∇u|2)i|2−C0|ψ′|2KF−C0F−C0Fi¯i|X1¯1i|λ1. | (4.33) |
Note that the operator F is concave, which implies that
−Fi¯j,p¯qXi¯j1Xp¯q¯1λ1≥0. | (4.34) |
Applying (4.33) and (4.34) to (4.29) and using φ″=2(φ′)2 yield that
0≥−C0Fi¯i|X1¯1i|λ1−C0λ1Fi¯i|X1¯1i|λ1+φ′∑pFi¯i(|upi|2+|u¯pi|2)−C0|ψ′|2KF−C0F−C0+ψ″Fi¯i|(u−u_)i|2+ψ′[Fi¯i(u−u_)i¯i−∂t(u−u_)]. | (4.35) |
Note that the fact
|X1¯1i|λ1=|−φ′(upiu¯p+upu¯pi)−ψ′(u−u_)i+(D11)iλ1|, |
It follows that
−C0Fi¯i|X1¯1i|λ1≥−C0φ′K−12Fi¯i(|upi|+|u¯pi|)+C0ψ′K12F−C0F. | (4.36) |
Using the following inequality
K−12(|upi|+|u¯pi|)≤14C0(|upi|2+|u¯pi|2)+C0K. |
deduces
−C0Fi¯i|X1¯1i|λ1≥−14φ′Fi¯i(|upi|2+|u¯pi|2)+C0ψ′K12F−C0F. | (4.37) |
Note that λ1>1, we have
−C0λ1Fi¯i|X1¯1i|λ1≥−14φ′Fi¯i(|upi|2+|u¯pi|2)+C0ψ′K12F−C0F. | (4.38) |
Since ψ″>0, which implies that
ψ″Fi¯i|(u−u_)i|2≥0. | (4.39) |
According to Lemma 2.3, there are at most two possibilities:
(1) If (2.8) holds true, then
ψ′[Fi¯i(u−u_)i¯i−∂t(u−u_)]≥θ(1+F)|ψ′|. | (4.40) |
Substituting (4.37)—(4.40) into (4.35) and using φ′≥14K yield that
0≥18K∑pFi¯i(|upi|2+|u¯pi|2)−C0|ψ′|2KF−C0(F+1)+θ(1+F)|ψ′|≥18KFi¯iλ2i−C0(C1+1)2KF+θ(C1+1)(F+1)−C0(F+1)≥δ2λ218nKF−C0(C1+1)2KF+θ(C1+1)(F+1)−C0(F+1). | (4.41) |
We may set θC1≥C0. It follows from (4.41) that λ1≤˜CK.
(2) If (2.9) holds true,
F1¯1>θN(1+F). | (4.42) |
According to ψ′<0 and the concavity of the operator F, we have
ψ′[Fi¯i(u−u_)i¯i−∂t(u−u_)]=ψ′[Fi¯i(Xi¯i−X_i¯i)−∂t(u−u_)]≥ψ′[F(χu)−F(χu_)−∂tu+∂tu_]=ψ′[ϕ(x,u)+∂tu_−F(χu_)]≥C0ψ′. | (4.43) |
Using (4.37)—(4.39) and (4.43) in (4.35), together with (4.42), we find that
0≥18KFi¯iλ2i−C0|ψ′|2KF+C0ψ′−C0(F+1)≥θλ218NK(1+F)+δ2λ218nKF−C0(C1+1)2KF−C0(C1+1)−C0(F+1). | (4.44) |
Let λ1 be sufficiently large, so that
θλ218NK(1+F)−C0(C1+1)−C0(F+1)≥0, |
It follows from (4.44)that λ1≤˜CK.
Case 2: λn≥−δλ1.
Let
I={i∈{1,⋯,n}|Fi¯i>δ−1F1¯1}. |
Let us first treat those indices which are not in I. Similar to (4.31), we obtain
−∑i∉IFi¯i|X1¯1i|2λ21≥−2(φ′)2∑i∉IFi¯i|(|∇u|2)i|2−C0Kδ|ψ′|2F1¯1−C0F. | (4.45) |
Using (4.32) yields that
−∑i∉IFi¯i|Xi¯11|2λ21≥−2(φ′)2∑i∉IFi¯i|(|∇u|2)i|2−C0Kδ|ψ′|2F1¯1−C0∑i∉IFi¯i|X1¯1i|λ1−C0F. | (4.46) |
Since
−Fi¯1,1¯i=Fi¯i−F1¯1X1¯1−Xi¯i andλi≥λn≥−δλ1, |
which implies that
−∑i∈IFi¯1,1¯i≥1−δ1+δ1λ1∑i∈IFi¯i, |
It follows that
−Fi¯1,1¯i|Xi¯11|2λ1≥1−δ1+δ∑i∈IFi¯i|Xi¯11|2λ21. | (4.47) |
Recalling φ″=2(φ′)2 and 0<δ≤12, we obtain from (4.5) that
∑i∈Iφ″Fi¯i|(|∇u|2)i|2=2∑i∈IFi¯i|Xi¯11λ1+ψ′(u−u_)i+χ11i−χi11+Tpi1χp¯1−(D11)iλ1|2≥2∑i∈IFi¯i(δ|Xi¯11λ1|2−2δ1−δ(ψ′)2|(u−u_)i|2−C0)≥2δ∑i∈IFi¯i|Xi¯11λ1|2−4δ1−δ(ψ′)2Fi¯i|(u−u_)i|2−C0F. | (4.48) |
Notice that ψ″≥4ϵ1−ϵ(ψ′)2 if ϵ=14E+1. Since 14E+1≥δ, we get that
ψ″Fi¯i|(u−u_)i|2−4δ1−δ(ψ′)2Fi¯i|(u−u_)i|2≥0. | (4.49) |
Take (4.46)–(4.49) into (4.29),
0≥−C0∑i∉IFi¯i|X1¯1i|λ1−C0λ1Fi¯i|X1¯1i|λ1+φ′∑pFi¯i(|upi|2+|u¯pi|2)−C0F−C0−C0Kδ|ψ′|2F1¯1+ψ′[Fi¯i(u−u_)i¯i−∂t(u−u_)]. | (4.50) |
Similar to (4.37) and (4.38), by using the third term of (4.50) to absorb the first two terms of it, We get that
0≥18K∑pFi¯i(|upi|2+|u¯pi|2)−C0Kδ|ψ′|2F1¯1−C0F−C0+ψ′[Fi¯i(u−u_)i¯i−∂t(u−u_)]≥18K∑pFi¯iλ2i−C0Kδ|ψ′|2F1¯1−C0(F+1)+ψ′[Fi¯i(u−u_)i¯i−∂t(u−u_)]. | (4.51) |
According to Lemma2.3, there are at most two possibilities:
(1) If (2.8) holds true, then
ψ′[Fi¯i(u−u_)i¯i−∂t(u−u_)]≥θ(1+F)|ψ′|. | (4.52) |
Put (4.52) into (4.51)
0≥18K∑pFi¯iλ2i−C0Kδ|ψ′|2F1¯1−C0(F+1)+θ(1+F)|ψ′|≥18K∑pF1¯1λ21−C0Kδ(1+C1)2F1¯1−C0(F+1)+θ(1+F)(1+C1). | (4.53) |
Here, C1 is determined finally, such that
θC1≥C0. |
It follows from (4.53) that
λ1≤˜CK. |
(2) If (2.9) holds true,
F1¯1>θN(1+F). | (4.54) |
Substituting (4.43) into (4.51) and using (4.54) give that
0≥18Kλ21−C0Kδ|(1+C1)|2−NθC0(1+C1)−NθC0 | (4.55) |
It follows that
λ1≤˜CK. |
To obtain the gradient estimates, we adapt the blow-up method of Dinew and Kolodziej [7] and reduce the problem to a Liouville type theorem which is proved in [7].
Proposition 5.1. Under the assumption of Theorem 1.3, let u(x,t) be a solution to (1.1). Then there exists a uniform constant ˜C such that
supM×[0,T)|∇u|≤˜C. | (5.1) |
Proof. Suppose that the gradient estimate (5.1) does not hold. Then there exists a sequence (xm,tm)∈M×[0,T) with tm→T such that
supM×[0,tm]|∇u(x,t)|=|∇u(xm,tm)| andlimm→∞|∇u(xm,tm)|=∞. |
After passing to a subsequence, we may assume that limm→∞xm=x0∈M. We choose a coordinate chart {U,(z1,⋯,zn)} at x0, which we identify with an open set in Cn, and such that ω(0)=β=√−1∑idzi∧d¯zi. We may assume that the open set contains ¯B1(0) and m is sufficiently large so that zm=z(xm)∈B1(0). Define
|∇u(xm,tm)|=Cm, ˜um(z)=u(zCm,tm). |
From this and Proposition 4.1, we have
supM|∇˜um|=1, supM|√−1 ¯∂∂˜um|≤˜C. |
This yields that ˜um is contained in the Hölder space C1,γ(Cn) with a uniform. Along with a standard application of Azela-Ascoli theorem, we may suppose ˜um has a limit ˜u∈C1,γ(Cn) with
|˜u|+|∇˜u|<C and |∇˜u(0)|≠0, | (5.2) |
in particular ˜u is not constant. On the other hand, similar to the method of Dinew and Kolodziej [7], we have
[χu(zCm)]k∧[ω(zCm)]n−k=e∂tuϕm(zCm,u)[χu(zCm)]l∧[ω(zCm)]n−l. |
Fixing z, we obtain
C2(k−l)m[O(1C2m)β+√−1 ¯∂∂˜um(z)]k∧[(1+O(|z|2C2m))β]n−k=e∂tumϕm(zCm,um)[O(1C2m)β+√−1 ¯∂∂˜um(z)]l∧[(1+O(|z|2C2m))β]n−l. | (5.3) |
Lemma 2.2 gives ∂tu is bounded uniformly. Since
ϕm(zCm,um)≤supM×[−C,C]ϕ, |
which implies that ϕm(zCm,um) is bounded uniformly. Taking the limits on both sides of 5.3 by m→∞ yields that
(√−1∂¯∂˜u)k∧βn−k=0. | (5.4) |
which is in the pluripotential sense. Moreover, a similar reasoning tells us that for any 1≤p≤k,
(√−1∂¯∂˜u)p∧βn−p≥0. | (5.5) |
Then, (5.4) and (5.5) imply ˜u is a k-subharmonic. By a result of Blocki [1], ˜u is a maximal k-subharmonic function in Cn. Applying the Liouville theorem in [7], we find that ˜u is a constant, which contradicts with (5.2).
In this section, we shall give a proof of the long-time existence to the flow and its convergence, that is Theorem 1.3.
From Lemma 2.2, Proposition 3.1, Propositions 4.1 and 5.1, we conclude that Eq 1.1 is uniformly parabolic. Therefore by Evans-Krylov's regularity theory [8,13,19,27] for uniformly parabolic equation, we obtain higher order derivative estimates. By the a priori estimates which don't depend on time, one can prove that the short time existence on [0,T) extends to [0,∞), that is the smooth solution exists at all time t>0. After proving C∞ estimates on [0,∞), we are able to show the convergence of the solution flow.
Let v=eγtut, where 0<γ<cϕ. Commuting derivative of v with respect to t and using (2.4), we obtain
vt=eγtutt+γv=γv+eγt(Fi¯juti¯j−ϕzϕut)=Fi¯jvi¯j+(γ−ϕzϕ)v. |
Using the condition (1.5) yields γ−ϕzϕ<0, According to the parabolic maximum principle, it follows that
supM×[0,∞)|v(x,t)|≤supM|ut(x,0)|≤supM|F(λ(u0))−ϕ(x,u0)|≤C, |
which means that |ut| decreases exponentially, in particular
∂t(u+Cγe−γt)≤0. |
According to Proposition 3.1, it follows that u+Cγe−γt is bounded uniformly and decreasing in t. Thus it converges to a smooth function u∞. From the higher order prior estimates, we can see that the function u(x,t) converges smoothly to u∞. Letting t→∞ in Eq (2.1),
σk(λ(u∞))σl(λ(u∞))=ϕ(x,u∞). |
In this paper, we have considered the parabolic Hessian quotient equation (1.1), in which the right hand side function ϕ depends on u. Firstly, we prove C0 estimates of Eq (1.1) by the parabolic C-subsolution condition and the Alexandroff-Bakelman-Pucci maximum principle. Secondly, we establish the C2 estimate for Eq (1.1) by using the parabolic C-subsolution condition. Thirdly, we obtain the gradient estimate by adapting the blowup method. Finally we give the proof of the long-time existence of the solution to the parabolic equation and its convergence. As an application, we show the solvability of a class of complex Hessian quotient equations, which generalizes the relevant results.
This work was supported by the Natural Science Foundation of Anhui Province Education Department (Nos. KJ2021A0659, gxgnfx2018017); Quality Enginering Project of Anhui Province Education Department (Nos. 2018jyxm0491, 2019mooc205, 2020szsfkc0686); Science Research Project of Fuyang Normal University (No. 2021KYQD0011)
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
[1] |
W. L. Wong, X. Su, X. Li, C. M. G. Cheung, R. Klein, C. Y. Cheng, et al., Global prevalence of age-related macular degeneration and disease burden projection for 2020 and 2040: A systematic review and meta-analysis, Lancet Glob. Health, 2 (2014), e106–e116, https://doi.org/10.1016/S2214-109X(13)70145-1. doi: 10.1016/S2214-109X(13)70145-1
![]() |
[2] |
H. E. Grossniklaus, W. Green, Choroidal neovascularization, Am. J. Ophthalmol., 137 (2004), 496–503. https://doi.org/10.1016/j.ajo.2003.09.042 doi: 10.1016/j.ajo.2003.09.042
![]() |
[3] |
M. R. Hee, C. R. Baumal, C. A. Puliafito, J. S. Duker, E. Reichel, J. R. Wilkins, et al., Optical coherence tomography of age-related macular degeneration and choroidal neovascularization, Ophthalmology, 103 (1996), 1260–1270. https://doi.org/10.1016/S0161-6420(96)30512-5 doi: 10.1016/S0161-6420(96)30512-5
![]() |
[4] |
T. D. Ting, M. Oh, T. A. Cox, C. H. Meyer, C. A. Toth, Decreased visual acuity associated with cystoid macular mdema in meovascular age-related macular degeneration, Arch. Ophthalmol., 120 (2002), 731–737. https://doi.org/10.1001/archopht.120.6.731 doi: 10.1001/archopht.120.6.731
![]() |
[5] |
N. Shah, M. G. Maguire, D. F. Martin, J. Shaffer, G. S. Ying, J. E. Grunwald, et al., Angiographic cystoid macular edema and outcomes in the comparison of age-related macular degeneration treatments trials, Ophthalmology, 123 (2016), 858–864. https://doi.org/10.1016/j.ophtha.2015.11.030 doi: 10.1016/j.ophtha.2015.11.030
![]() |
[6] | A. Loewenstein, D. Zur, Postsurgical cystoid macular edema, in Macular Edema, (2010), 148–159. https://doi.org/10.1159/000320078 |
[7] |
M. Rajappa, P. Saxena, J. Kaur, Chapter 6-ocular angiogenesis: Mechanisms and recent advances in therapy, Adv. Clin. Chem., 50 (2010), 103–121. https://doi.org/10.1016/S0065-2423(10)50006-4 doi: 10.1016/S0065-2423(10)50006-4
![]() |
[8] |
V. Tah, H. O. Orlans, J. Hyer, E. Casswell, N. Din, V. Sri Shanmuganathan, et al, Anti-VEGF therapy and the retina: An update, J. Ophthalmol., 2015 (2015). https://doi.org/10.1155/2015/627674 doi: 10.1155/2015/627674
![]() |
[9] |
H. Gerding, J. Monexs, R. Tadayoni, F. Boscia, I. Pearce, S. Priglinger, Ranibizumab in retinal vein occlusion: Treatment recommendations by an expert panel, Brit. J. Ophthalmol., 99 (2015), 297–304. https://doi.org/10.1136/bjophthalmol-2014-305041 doi: 10.1136/bjophthalmol-2014-305041
![]() |
[10] | D. Yorston, Anti-VEGF drugs in the prevention of blindness, Community Eye Health, 27 (2014), 44–46. |
[11] |
R. Marano, I. Toth, N. Wimmer, M. Brankov, P. Rakoczy, Dendrimer delivery of an anti-VEGF oligonucleotide into the eye: A long-term study into inhibition of laser-induced CNV, distribution, uptake and toxicity, Gene Ther., 12 (2005), 1544–1550. https://doi.org/10.1038/sj.gt.3302579 doi: 10.1038/sj.gt.3302579
![]() |
[12] |
J. H. Chang, N. K. Garg, E. Lunde, K. Y. Han, S. Jain, D. T. Azar, Corneal neovascularization: An anti-VEGF therapy review, Surv. Ophthalmol., 57 (2012), 415–429. https://doi.org/10.1016/j.survophthal.2012.01.007 doi: 10.1016/j.survophthal.2012.01.007
![]() |
[13] |
J. M. Schmitt, Optical coherence tomography (OCT): A review, IEEE J. Sel. Top. Quant. Electron., 5 (1999), 1205–1215. https://doi.org/10.1109/2944.796348 doi: 10.1109/2944.796348
![]() |
[14] |
K. Gnep, A. Fargeas, R. E. Gutiexrrez-Carvajal, F. Commandeur, R. Mathieu, J. D. Ospina, et al., Haralick textural features on T2-weighted MRI are associated with biochemical recurrence following radiotherapy for peripheral zone prostate cancer, J. Magn. Reson. Imaging, 45 (2017), 103–117. https://doi.org/10.1002/jmri.25335 doi: 10.1002/jmri.25335
![]() |
[15] |
C. Shen, Z. Liu, Z. Wang, J. Guo, H. Zhang, Y. Wang, et al., Building CT radiomics based nomogram for preoperative esophageal cancer patients lymph node metastasis prediction, Transl. Oncol., 11 (2018), 815–824. https://doi.org/10.1016/j.tranon.2018.04.005 doi: 10.1016/j.tranon.2018.04.005
![]() |
[16] |
R. Paul, S. H. Hawkins, Y. Balagurunathan, M. Schabath, R. J. Gillies, L. O. Hall, et al., Deep feature transfer learning in combination with traditional features predicts survival among patients with lung adenocarcinoma, Tomography, 2 (2016), 388–395. https://doi.org/10.18383/j.tom.2016.00211 doi: 10.18383/j.tom.2016.00211
![]() |
[17] | W. Han, L. Qin, C. Bay, X. Chen, K. H. Yu, A. Li, et al., Integrating deep transfer learning and radiomics features in glioblastoma multiforme patient survival prediction, in Medical Imaging 2020: Image Processing, International Society for Optics and Photonics, (2020), 113132S. https://doi.org/10.1117/12.2549325 |
[18] |
W. Han, L. Qin, C. Bay, X. Chen, K. H. Yu, N. Miskin, et al., Deep transfer learning and radiomics feature prediction of survival of patients with high-grade gliomas, Am. J. Neuroradiology, 41 (2020), 40–48. https://doi.org/10.3174/ajnr.A6365 doi: 10.3174/ajnr.A6365
![]() |
[19] |
N. Papandrianos, E. I. Papageorgiou, A. Anagnostis, Development of convolutional neural networks to identify bone metastasis for prostate cancer patients in bone scintigraphy, Ann. Nucl. Med., 34 (2020), 824–832. https://doi.org/10.1007/s12149-020-01510-6 doi: 10.1007/s12149-020-01510-6
![]() |
[20] | C. Fourcade, L. Ferrer, G. Santini, N. Moreau, C. Rousseau, M. Lacombe, et al., Combining superpixels and deep learning approaches to segment active organs in metastatic breast cancer PET images, in 2020 42nd Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC), IEEE, (2020), 1536–1539. https://doi.org/10.1109/EMBC44109.2020.9175683 |
[21] |
K. H. Cha, L. Hadjiiski, H. P. Chan, A. Z. Weizer, A. Alva, R. H. Cohan, et al., Bladder cancer treatment response assessment in CT using radiomics with deep-learning, Sci. Rep., 7 (2017), 1–12. https://doi.org/10.1038/s41598-017-09315-w doi: 10.1038/s41598-017-09315-w
![]() |
[22] |
M. Byra, K. Dobruch-Sobczak, Z. Klimonda, H. Piotrzkowska-Wroblewska, J. Litniewski, Early prediction of response to neoadjuvant chemotherapy in breast cancer sonography using siamese convolutional neural networks, IEEE J. Biomed. Health Inf., 25 (2020), 797–805. https://doi.org/10.1109/JBHI.2020.3008040 doi: 10.1109/JBHI.2020.3008040
![]() |
[23] |
D. Romo-Bucheli, U. Erfurth, H. Bogunović, End-to-end deep learning model for predicting treatment requirements in neovascular AMD from longitudinal retinal OCT imaging, IEEE J. Biomed. Health Inf., 24 (2020), 3456–3465. https://doi.org/10.1109/JBHI.2020.3000136 doi: 10.1109/JBHI.2020.3000136
![]() |
[24] |
S. Starke, S. Leger, A. Zwanenburg, K. Leger, F. Lohaus, A. Linge, et al., 2D and 3D convolutional neural networks for outcome modelling of locally advanced head and neck squamous cell carcinoma, Sci. Rep., 10 (2020), 1–13. https://doi.org/10.1038/s41598-020-70542-9 doi: 10.1038/s41598-020-70542-9
![]() |
[25] | J. D. Cardosi, H. Shen, J. I. Groner, M. Armstrong, H. Xiang, Machine intelligence for outcome predictions of trauma patients during emergency department care, preprint, arXiv: 2009.03873. https://doi.org/10.48550/arXiv.2009.03873 |
[26] |
M. Pease, D. Arefan, J. Biligen, J. Sharpless, A. Puccio, K. Hochberger, et al., Deep neural network analysis of CT scans to predict outcomes in a prospective database of severe traumatic brain injury patients, Neurosurgery, 67 (2020), 417. https://doi.org/10.1093/neuros/nyaa447_417 doi: 10.1093/neuros/nyaa447_417
![]() |
[27] | D. Feng, X. Chen, Z. Zhou, H. Liu, Y. Wang, L. Bai, et al., A preliminary study of predicting effectiveness of anti-VEGF injection using OCT images based on deep learning, in 42nd Annual International Conference of the IEEE Engineering in Medicine and Biology Society, (2020), 5428–5431. https://doi.org/10.1109/EMBC44109.2020.9176743 |
[28] | M. Noroozi, A. Vinjimoor, P. Favaro, H. Pirsiavash, Boosting self-supervised learning via knowledge transfer, in 2018 IEEE Conference on Computer Vision and Pattern Recognition, (2018). https://doi.org/10.1109/CVPR.2018.00975 |
[29] | C. Doersch, A. Gupta, A. A. Efros, Unsupervised visual representation learning by context prediction, in 2015 IEEE International Conference on Computer Vision, (2015), 1422–1430. https://doi.org/10.1109/ICCV.2015.167 |
[30] | M. Noroozi, P. Favaro, Unsupervised learning of visual representations by solving jigsaw puzzles, preprint, arXiv: 1603.09246. https://doi.org/10.48550/arXiv.1603.09246 |
[31] |
L. Chen, P. Bentley, K. Mori, K. Misawa, M. Fujiwara, et al., Self-supervised learning for medical image analysis using image context restoration, Med. Image Anal., 58 (2019), 101539. https://doi.org/10.1016/j.media.2019.101539 doi: 10.1016/j.media.2019.101539
![]() |
[32] |
J. Zhu, Y. Li, Y. Hu, K. Ma, S. K. Zhou, Y. Zheng, Rubik's cube+: A self-supervised feature learning framework for 3D medical image analysis, Med. Image Anal., 64 (2020), 101746. https://doi.org/10.1016/j.media.2020.101746 doi: 10.1016/j.media.2020.101746
![]() |
[33] | Z. Zhou, V. Sodha, M. M. R. Siddiquee, R. Feng, N. Tajbakhsh, M. B. Gotway, et al., Models genesis: Generic autodidactic models for 3D medical image analysis, in International Conference on Medical Image Computing and Computer Assisted Intervention, (2019), 384–393. https://doi.org/10.1007/978-3-030-32251-9_42 |
[34] | G. Larsson, M. Maire, G. Shakhnarovich, Colorization as a proxy task for visual understanding, in 2017 IEEE Conference on Computer Vision and Pattern Recognition, (2017), 840–849. https://doi.org/10.1109/CVPR.2017.96 |
[35] | T. Chen, S. Kornblith, M. Norouzi, G. E. Hinton, A simple framework for contrastive learning of visual representations, in Proceedings of the 37th International Conference on Machine Learning, preprint, arXiv: 2002.05709. |
[36] | A. Van den Oord, Y. Li, O. Vinyals, Representation learning with contrastive predictive coding, preprint, arXiv: 1807.03748. |
[37] | G. Lorre, J. Rabarisoa, A. Orcesi, S. Ainouz, S. Canu, Temporal contrastive pretraining for video action recognition, in Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, (2020), 662–670. https://doi.org/10.1109/WACV45572.2020.9093278 |
[38] | L. Tao, X. Wang, T. Yamasaki, Self-supervised video representation learning using inter- intra contrastive framework, in MM'20: The 28th ACM International Conference on Multimedia, (2020), 2193–2201. https://doi.org/10.1145/3394171.3413694 |
[39] | Z. Wu, Y. Xiong, S. X. Yu, D. Lin, Unsupervised feature learning via non-parametric instance discrimination, in 2018 IEEE Conference on Computer Vision and Pattern Recognition, (2018), 3733–3742. https://doi.org/10.1109/CVPR.2018.00393 |
[40] | K. He, H. Fan, Y. Wu, S. Xie, R. B. Girshick, Momentum contrast for unsupervised visual representation learning, in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition, (2020), 9726–9735. https://doi.org/10.1109/CVPR42600.2020.00975 |
[41] | H. Spitzer, K. Kiwitz, K. Amunts, S. Harmeling, T. Dickscheid, Improving cytoarchitectonic segmentation of human brain areas with self-supervised siamese networks, in Medical Image Computing and Computer Assisted Intervention-MICCAI 201-21st International Conference, (2018), 663–671. https://doi.org/10.1007/978-3-030-00931-1 |
[42] |
X. Li, M. Jia, M. T. Islam, L. Yu, L. Xing, Self-supervised feature learning via exploiting multi-modal data for retinal disease diagnosis, IEEE Trans. Med. Imaging, 39 (2020), 4023–4033. https://doi.org/10.1109/TMI.2020.3008871 doi: 10.1109/TMI.2020.3008871
![]() |
[43] |
D. S. Kermany, M. Goldbaum, W. Cai, C. C. Valentim, H. Liang, S. L. Baxter, et al., Identifying medical diagnoses and treatable diseases by image-based deep learning, Cell, 172 (2018), 1122–1131. https://doi.org/10.1016/j.cell.2018.02.010 doi: 10.1016/j.cell.2018.02.010
![]() |
[44] | M. E. Mortenson, Mathematics For Computer Graphics Applications, Industrial Press Inc., 1999. |
[45] | O. Ronneberger, P. Fischer, T. Brox, U-net: Convolutional networks for biomedical image segmentation, in Medical Image Computing and Computer-Assisted Intervention-MIC-CAI 2015-18th International Conference Munich, (2015), 234–241. https://doi.org/10.1007/978-3-319-24574-4 |
[46] |
C. Chang, C. Lin, LIBSVM: A library for support vector machines, ACM Trans. Intell. Syst. Technol., 2 (2011), 1–27. https://doi.org/10.1145/1961189.1961199 doi: 10.1145/1961189.1961199
![]() |
[47] |
V. N. Vapnik, An overview of statistical learning theory, IEEE Trans. Neural Networks, 10 (1999), 988–999. https://doi.org/10.1109/72.788640 doi: 10.1109/72.788640
![]() |
[48] |
D. Lu, K. Popuri, G. W. Ding, R. Balachandar, M. F. Beg, Multiscale deep neural network ased analysis of FDG-PET images for the early diagnosis of Alzheimer's disease, Med. Image Anal., 46 (2018), 26–34. https://doi.org/10.1016/j.media.2018.02.002 doi: 10.1016/j.media.2018.02.002
![]() |
[49] |
Y. Y. Liu, M. Chen, H. Ishikawa, G. Wollstein, J. S. Schuman, J. M. Rehg, Automated macular pathology diagnosis in retinal OCT images using multi-scale spatial pyramid and local binary patterns in texture and shape encoding, Med. Image Anal., 15 (2011), 748–59. https://doi.org/10.1016/j.media.2011.06.005 doi: 10.1016/j.media.2011.06.005
![]() |
[50] |
P. P. Srinivasan, L. A. Kim, P. S. Mettu, S. W. Cousins, G. M. Comer, J. A. Izatt, et al., Fully automated detection of diabetic macular edema and dry age-related macular degeneration from optical coherence tomography images, Biomed. Optics Express, 5 (2014), 3568–577. https://doi.org/10.1364/BOE.5.003568 doi: 10.1364/BOE.5.003568
![]() |
[51] |
R. Rasti, H. Rabbani, A. Mehridehnavi, F. Hajizadeh, Macular OCT classification using a multi-scale convolutional neural network ensemble, IEEE Trans. Med. Imaging, 37 (2017), 1024–034. https://doi.org/10.1109/TMI.2017.2780115 doi: 10.1109/TMI.2017.2780115
![]() |
[52] |
X. Wang, H. Chen, A. R. Ran, L. Luo, P. P. Chan, C. C. Tham, et al., Towards multi-center glaucoma OCT image screening with semi-supervised joint structure and function multi-task learning, Med. Image Anal., 63 (2020), 101695. https://doi.org/10.1016/j.media.2020.101695 doi: 10.1016/j.media.2020.101695
![]() |
[53] |
H. Bogunović, S. M. Waldstein, T. Schlegl, G. Langs, A. Sadeghipour, X. Liu, et al., Prediction of Anti-VEGF treatment requirements in neovascular amd using a machine learning approach, Investi. Ophth. Vis. Sci., 58 (2017), 3240–3248. https://doi.org/10.1167/iovs.16-21053 doi: 10.1167/iovs.16-21053
![]() |
[54] |
Z. Zhou, M. Folkert, P. Iyengar, K. Westover, Y. Zhang, H. Choy, et al., Multi-objective radiomics model for predicting distant failure in lung SBRT, Phys. Med. Biol., 62 (2017), 4460. https://doi.org/10.1088/1361-6560/aa6ae5 doi: 10.1088/1361-6560/aa6ae5
![]() |
[55] | K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in 2016 IEEE Conference on Computer Vision and Pattern Recognition, (2016), 770–78. https://doi.org/10.1109/CVPR.2016.90 |
[56] | R. R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, D. Batra, Grad-CAM: Visual explanations from deep networks via gradient-based localization, in IEEE International Conference on Computer Vision, (2017), 618–626. https://doi.org/10.1109/ICCV.2017.74 |