Deep-OCTA Ensemble Deep Learning Approaches for Diabetic Retinopathy Analysis on OCTA Images

2025-05-06 0 0 3.65MB 14 页 10玖币

侵权投诉

Deep-OCTA: Ensemble Deep Learning

Approaches for Diabetic Retinopathy Analysis

on OCTA Images

Junlin Hou1, Fan Xiao2, Jilan Xu1, Yuejie Zhang1*,

Haidong Zou3, and Rui Feng1,2*

1School of Computer Science, Shanghai Key Laboratory of Intelligent Information

Processing, Fudan University, China

{jlhou18,jilanxu18,yjzhang,fengrui}@fudan.edu.cn

2Academy for Engineering and Technology, Fudan University, China

21210860085@m.fudan.edu.cn

3Department of Ophthalmology, Shanghai General Hospital, School of Medicine,

Shanghai Jiao Tong University, China

zouhaidong@sjtu.edu.cn

Abstract. The ultra-wide optical coherence tomography angiography

(OCTA) has become an important imaging modality in diabetic retinopa-

thy (DR) diagnosis. However, there are few researches focusing on au-

tomatic DR analysis using ultra-wide OCTA. In this paper, we present

novel and practical deep-learning solutions based on ultra-wide OCTA

for the Diabetic Retinopathy Analysis Challenge (DRAC). In the seg-

mentation of DR lesions task, we utilize UNet and UNet++ to segment

three lesions with strong data augmentation and model ensemble. In

the image quality assessment task, we create an ensemble of Inception-

V3, SE-ResNeXt, and Vision Transformer models. Pre-training on the

large dataset as well as the hybrid MixUp and CutMix strategy are

both adopted to boost the generalization ability of our model. In the DR

grading task, we build a Vision Transformer (ViT) and ﬁnd that the ViT

model pre-trained on color fundus images serves as a useful substrate for

OCTA images. Our proposed methods ranked 4th, 3rd, and 5th on the

three leaderboards of DRAC, respectively. The source code will be made

available at https://github.com/FDU-VTS/DRAC.

Keywords: Diabetic retinopathy analysis ·Optical coherence tomogra-

phy angiography ·Deep learning.

1 Introduction

Diabetic Retinopathy (DR) is a chronic progressive disease that causes visual

impairment due to retinal microvascular damage. It has become a leading cause

of legal blindness in the working-age population worldwide [18]. DR is diagnosed

by the presence of retinal lesions, such as microaneurysms (MAs), intraretinal

microvascular abnormalities (IRMAs), nonperfusion areas (NPAs) and neovas-

cularization (NV) [8]. The traditional diagnosis of DR grading mainly relies

arXiv:2210.00515v1 [eess.IV] 2 Oct 2022

2 J. Hou et al.

OCTA IRMAs NPAs NV

Poor Good Excellent

(a) Segmentation of DR lesions

(b) Image quality assessment (c) DR grading

Normal NPDR PDR

Fig. 1. Samples from three tasks in the DRAC dataset, including (a) segmentation of

DR lesions, (b) image quality assessment, and (c) DR grading.

on fundus photography and ﬂuorescein fundus angiography (FFA). With rising

popularity, OCT angiography (OCTA) has become a reliable tool due to its ca-

pability of visualizing the retinal and choroidal vasculature at a microvascular

level [9]. Compared to fundus photography and FFA, the ultra-wide OCTA can

non-invasively detect the changes of DR neovascularization, thus it is an impor-

tant imaging modality to assist ophthalmologists in diagnosing Proliferative DR

(PDR).

Recently, deep learning approaches have achieved promising performance in

DR diagnosis. Based on fundus photography, numerous deep learning methods

are proposed for lesion segmentation, biomarkers segmentation, disease diagnosis

and image synthesis [5]. A deep learning system, named DeepDR, is developed

to perform real-time image quality assessment, lesion detection and DR grading

[1]. However, there are currently few works capable of automatic DR analysis

using ultra-wide OCTA images.

In this paper, we propose novel and practical methods for DR analysis based

on ultra-wide OCTA images. Our methods are demonstrated eﬀective in the

MICCAI 2022 Diabetic Retinopathy Analysis Challenge (DRAC) [8]. This chal-

lenge provides a standardized ultra-wide OCTA dataset, including lesion seg-

mentation, image quality assessment and DR grading. As illustrated in Fig. 1,

the segmentation subset includes three DR lesions, i.e. Intraretinal Microvascu-

lar Abnormalities (IRMAs), Nonperfusion Areas (NPAs), and Neovascularization

(NV). The image quality assessment subset divides images into three quality lev-

els, i.e. Poor quality level, Good quality level, and Excellent quality level. The

DR grading subset contains images of three diﬀerent DR grades, i.e. Normal,

Non-Proliferative DR (NPDR), and Proliferative DR (PDR). In our approach,

we ﬁrst develop two encoder-decoder networks, namely UNet [7] and UNet++

[19], to segment three lesions separately. Strong data augmentation and model

ensemble are demonstrated helpful to the generalization ability of our models.

Deep-OCTA 3

For image quality assessment, we create an ensemble of multiple state-of-the-

art neural networks. The networks are ﬁrst pre-trained on large-scale datasets,

and then ﬁne-tuned on DRAC dataset using the hybrid MixUp and CutMix

strategy. For DR grading, we adopt a Vision Transformer model, which utilizes

self-attention to integrate information across the entire OCTA image. Extensive

experiments on DRAC dataset show that our proposed solutions achieve supe-

rior DR analysis performance, ranking 4th, 3rd, 5th on the three leaderboards,

respectively.

2 Segmentation of Diabetic Retinopathy Lesions

2.1 Methodology

We adopt UNet [7] and UNet++ [19] networks with pre-trained encoders for DR

lesion segmentation. Customized strategies are designed to train the segmenta-

tion models of three diﬀerent lesions. For IRMAs segmentation, Step1 learning

rate schedule and color jittering augmentation are both adopted. However, they

are not utilized to boost the segmentation performance of NPAs and NV lesions.

Besides, we employ the model ensemble strategy when predicting the segmented

masks of IRMAs and NPAs. In the following section, we will introduce the UNet

and UNet++ networks, learning rate schedules, and loss functions.

Backbone network. (1) UNet [7] is an encoder-decoder network, where the

encoder includes down-sampling layers and the decoder consists of up-sampling

layers with skip connections. The feature structures of diﬀerent levels are com-

bined through skip connections. (2) UNet++ [19] proposes an eﬃcient ensemble

of U-Nets of varying depths, which partially share an encoder and co-learn si-

multaneously using deep supervision. It redesigns skip connections to exploit

multi-scale semantic features, leading to a ﬂexible feature aggregation scheme.

Learning rate schedule. In order to train our network eﬃciently, we ana-

lyze two diﬀerent schedules for learning rate decay. One method, named Step1,

divides the initial learning rate by 10 at 25% of the total number of training

epochs. The other method, named Step2, is to decay the learning rate to 0.6×

of the previous value every quarter epoch of the total epochs.

Loss function. We adopt two loss functions, i.e. Dice loss LDand Jaccard loss

LJ, to train the segmentation models of DR lesions. Speciﬁcally, the Dice loss

LDis expressed by:

LD=1

i=1

1−ScoreDice(i),(1)

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

Deep-OCTA:EnsembleDeepLearningApproachesforDiabeticRetinopathyAnalysisonOCTAImagesJunlinHou1,FanXiao2,JilanXu1,YuejieZhang1*,HaidongZou3,andRuiFeng1;2*1SchoolofComputerScience,ShanghaiKeyLaboratoryofIntelligentInformationProcessing,FudanUniversity,Chinafjlhou18,jilanxu18,yjzhang,fengruig@fudan.edu.c...

展开>> 收起<<

Deep-OCTA Ensemble Deep Learning Approaches for Diabetic Retinopathy Analysis on OCTA Images.pdf

共14页,预览3页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Deep-OCTA Ensemble Deep Learning Approaches for Diabetic Retinopathy Analysis on OCTA Images

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: