Bag of Tricks for Developing Diabetic Retinopathy Analysis Framework to Overcome Data Scarcity

2025-05-02 0 0 3.08MB 16 页 10玖币

侵权投诉

Bag of Tricks for Developing Diabetic

Retinopathy Analysis Framework to Overcome

Data Scarcity

Gitaek Kwon, Eunjin Kim, Sunho Kim, Seongwon Bak, Minsung Kim, and

Jaeyoung Kim?

VUNO Inc., Seoul, South Korea

{gitaek.kwon,eunjin.kim,ksunho0660,seongwon.bak,minsung.kim,

jaeyoung.kim}@vuno.co

Abstract. Recently, diabetic retinopathy (DR) screening utilizing ultra-

wide optical coherence tomography angiography (UW-OCTA) has been

used in clinical practices to detect signs of early DR. However, develop-

ing a deep learning-based DR analysis system using UW-OCTA images is

not trivial due to the diﬃculty of data collection and the absence of pub-

lic datasets. By realistic constraints, a model trained on small datasets

may obtain sub-par performance. Therefore, to help ophthalmologists

be less confused about models’ incorrect decisions, the models should

be robust even in data scarcity settings. To address the above practical

challenging, we present a comprehensive empirical study for DR analysis

tasks, including lesion segmentation, image quality assessment, and DR

grading. For each task, we introduce a robust training scheme by leverag-

ing ensemble learning, data augmentation, and semi-supervised learning.

Furthermore, we propose reliable pseudo labeling that excludes uncertain

pseudo-labels based on the model’s conﬁdence scores to reduce the nega-

tive eﬀect of noisy pseudo-labels. By exploiting the proposed approaches,

we achieved 1st place in the Diabetic Retinopathy Analysis Challenge.

Keywords: Diabetic Retinopathy Analysis ·Semi-supervised learning.

1 Introduction

Diabetic retinopathy (DR) is an eye disease that can result in vision loss and

blindness in people with diabetes, but early DR might cause no symptoms or

only mild vision problems [7]. Therefore, early detection and management of

DR play a crucial role in improving the clinical outcome of eye condition. Color

fundus photography, ﬂuorescein angiography (FA), and optical coherence tomog-

raphy angiography (OCTA) have been used in diabetic eye screening to acquire

valuable information for DR diagnosis and treatment planning. Recently, in the

screening, ultra-wide OCTA (UW-OCTA) images have been widely used lever-

aging their advantages such as more detailed visualization of vessel structures,

?Correspondence to Jaeyoung Kim

arXiv:2210.09558v1 [eess.IV] 18 Oct 2022

2 Gitaek Kwon et al.

and ability to capture a much wider view of the retinal compared to previous

standard approaches [34].

With the advancements of deep learning (DL), applying DL-based methods

for medical image analysis has become an active research area in the ophthalmol-

ogy ﬁelds [13, 23, 27, 28]. Notably, the availability to large amounts of annotated

fundus photography has been one of the key elements driving the quick growth

and success of developing automated DR analysis tools. Sun et al. [29] develop

the automatic DR diagnostic models using color fundus images, and Zhou et al.

[36] propose a collaborative learning approach to improve the accuracy of DR

grading and lesion segmentation by semi-supervised learning on the color fundus

photography. Although previous studies investigate the eﬀectiveness of applying

DL to DR grading and lesion detection tasks based on color fundus images, DR

analysis tool leveraging UW-OCTA are still under-consideration. One of the rea-

sons lies in the fact that annotating high-quality UW-OCTA images is inherently

diﬃcult because the annotation of medical images requires manual labeling by

experts. Consequently, when we consider about the practical restrictions, it is

one of the most crucial things to develop a robust model even in the lack of data.

To address the above real-world setting, we introduce the bag of tricks for

DR analysis tasks using the Diabetic Retinopathy Analysis Challenge (DRAC22)

dataset, which consists of three tasks (i.e., lesion segmentation, image quality

assessment, and DR grading) [25]. To alleviate the negative eﬀect introduced by

the lack of labeled data, we investigate the eﬀectiveness of data augmentations,

ensembles of deep neural networks, and semi-supervised learning. Furthermore,

we propose reliable pseudo labeling (RPL) that selects reliable pseudo-labels

based on a trained classiﬁer’s conﬁdence scores, and then the classiﬁer is re-

trained with labeled and trustworthy pseudo-labeled data.

In our study, we ﬁnd that Deep Ensembles [11], test-time data augmentation

(TTA), and RPL have powerful eﬀects for DR analysis tasks. Our solutions are

combinations of the above techniques and achieved 1st place in all tasks for

DRAC22.

2 Related Work

In this section, we overview previous studies on the DR analysis (Sec. 2.1), and

semi-supervised learning algorithms (Sec. 2.2).

2.1 Diabetic Retinopathy Analysis

Automatic DR assessment methods based on neural networks have been de-

veloped to assist ophthalmologists [20, 21, 35]. Gulshan et al. [8] develop the

convolution neural network (CNN) for detecting DR, and the proposed method

shows the competitive result with ophthalmologists in detection performance.

They demonstrates the feasibility of the DL-based computer-aided diagnosis

system for fundus photography. Dai et al. [4] suggest a uniﬁed framework called

DeepDR in order to improve the interpretability of CNNs. DeepDR provides

Bag of Tricks for Diabetic Retinopathy Analysis 3

comprehensive predictions, including DR grade, location of DR-related lesions,

and an image quality assessment of color fundus photography.

On the other hand, a series of approaches based on the FA [5, 17], OCT

[6, 10], and OCTA [22, 33] have been studied to detect DR. Pan et al. [17]

propose the CNN-based model, which classiﬁes DR ﬁndings (i.e., non-perfusion

regions, microaneurysms, and laser scars) with FA. Heisler et al. [10] suggest

an ensemble network for DR classiﬁcation. Each ensemble member is trained

with OCT and OCTA, respectively. For a testing time, they use aggregated

predictions of the ensemble model to provide robust and calibrated predictions.

Although the previous methods have shown remarkable results in promoting

the accuracy of DR grading, a comprehensive empirical study of applying UW-

OCTA to DL has yet to be conducted.

2.2 Semi-supervised Learning

In the medical imaging domain, collecting labeled data is challenging due to ex-

pensive costs and time-consuming. Instead, it is much easier to obtain unlabeled

data. Thanks to the recent success of semi-supervised learning (SSL), various

SSL algorithms [2, 18] show impressive performance on various tasks such as

semantic segmentation, object detection, and image recognition.

Pseudo-labeling (PL) [12] is a simple and eﬀective method in SSL approaches,

in which pseudo labels are generated based on the pretrained-network’s predic-

tions, and then the network is re-trained both labeled and pseudo-labeled data

simultaneously. Following the pioneering approach of pseudo-labeling, Sohn et al.

[26] propose FixMatch, which produces pseudo labels using both consistency

regularization and pseudo-labeling. FixMatch only retains a pseudo label if the

network produces a high probability for a weak-augmented image in order to

reduce an error of the prediction caused by the distortions of a given image. Xie

et al. [32] suggests an iterative training scheme for SSL, called noisy student

training. In their process, they ﬁrst train a model on labeled data and use it as

a teacher network to generate pseudo labels for unlabeled data. They then train

an equal-or-larger model as a student network on the combination of labeled and

pseudo-labeled samples and iterate the above process by assigning the student

as the teacher.

3 Method

In this section, we ﬁrst introduce a simple and eﬀective technique, reliable pseudo

labeling (RPL), for improving classiﬁcation performance in the data scarcity

setting. Then, we describe our solutions for each task in detail.

Scope of the paper. We consider a standard multi-class classiﬁcation problem

which comprise a common setting in computer vision tasks. In this paper, our

goal is for a model to classify as ˆyi={1, ..., C} ∈ N(i.e., discrete random

variable) for a given xi.

4 Gitaek Kwon et al.

Fig. 1. Overall procedure of RPL.

3.1 Reliable Pseudo Labeling

Pitfall of pseudo-labels. While PL shows powerful performance in the data

scarcity setting, networks can produce an incorrect prediction on unseen data.

If the model is trained using incorrect pseudo-labels, errors accumulate and

conﬁrmation bias can appear since modern over-parameterized neural networks

easily overﬁt to noisy samples [1]. Hence, it would be reckless to consider all

pseudo-labels generated by a network trained with a small amount of data as

correct predictions. To address the vulnerability of PL, we propose RPL, and

overall procedure is described in Fig. 1.

Notation. Let [m] := {1, ..., m},σ(·) is the softmax function, and b·e is the

nearest integer function. We denote a given dataset by D=Dl∪ Du, where

Dl={(xl

i, yl

i) : i∈[n]}is the labeled set and Du={xu

i:i∈[m]}is the

unlabeled set. For multi-class classiﬁcation tasks, a softmax classiﬁer fcls maps

an input xi∈RW×Hinto a predictive distribution ˆp(y|σ(zi)), where ziis a vector

of logits fcls(xi) and y∈[C] is a discrete class label. When the classiﬁcation task

is formulated by regression problem, a class prediction of a regressor freg can be

calculated by bfreg(xi)ewhen a label is deﬁned as y∈N.

Procedure of RPL. We ﬁrst train fusing an arbitrary loss function using

train. After training, we collect pseudo-labeled set per predicted class ˆyde-

ﬁned as ˆ

Du={ˆ

k}C

k=1, where ˆ

k={(xu

i,ˆyu

i)|ˆyu

i=k, i ∈[m]}. Then we sort

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：

BagofTricksforDevelopingDiabeticRetinopathyAnalysisFrameworktoOvercomeDataScarcityGitaekKwon,EunjinKim,SunhoKim,SeongwonBak,MinsungKim,andJaeyoungKim?VUNOInc.,Seoul,SouthKoreafgitaek.kwon,eunjin.kim,ksunho0660,seongwon.bak,minsung.kim,jaeyoung.kimg@vuno.coAbstract.Recently,diabeticretinopathy(DR)scr...

展开>> 收起<<

Bag of Tricks for Developing Diabetic Retinopathy Analysis Framework to Overcome Data Scarcity.pdf

共16页,预览4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！

Bag of Tricks for Developing Diabetic Retinopathy Analysis Framework to Overcome Data Scarcity

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: