Diabetic Retinopathy Detection Based on Weakly Supervised Object Localization and Knowledge Driven Attribute Mining
Disease grading and lesion identification are two important tasks for diabetic retinopathy detection. Disease grading uses image-level annotation but lesion identification often needs the fine-grained annotations, which requires a lot of time and effort of professional doctors. Therefore, it is a gr...
Saved in:
Published in | Ophthalmic Medical Image Analysis pp. 32 - 41 |
---|---|
Main Authors | , , , |
Format | Book Chapter |
Language | English |
Published |
Cham
Springer International Publishing
21.09.2021
|
Series | Lecture Notes in Computer Science |
Subjects | |
Online Access | Get full text |
Cover
Loading…
Summary: | Disease grading and lesion identification are two important tasks for diabetic retinopathy detection. Disease grading uses image-level annotation but lesion identification often needs the fine-grained annotations, which requires a lot of time and effort of professional doctors. Therefore, it is a great challenge to complete disease grading and lesion identification simultaneously with the limited labeled data. We propose a method based on weakly supervised object localization and knowledge driven attribute mining to conduct disease grading and lesion identification using only image-level annotation. We first propose an Attention-Drop-Highlight Layer (ADHL), which enables the CNN to accurately and comprehensively focus on the various lesion features. Then, we design a search space and employ neural architecture search (NAS) to select the best settings of the ADHL, to maximize the performance of the model. Finally, we regard the lesion attributes corresponding to different disease grades as weakly supervised classification labels representing prior knowledge, and propose an Attribute Mining (AM) method to further improve the effect of disease grading and complete lesion identification. Extensive experiments and a user study have proved that our method can capture more lesion features, improve the performance of disease grading, and obtain state-of-the-art results compared to the methods only using image-level annotation. |
---|---|
Bibliography: | Electronic supplementary materialThe online version of this chapter (https://doi.org/10.1007/978-3-030-87000-3_4) contains supplementary material, which is available to authorized users. |
ISBN: | 9783030869991 3030869997 |
ISSN: | 0302-9743 1611-3349 |
DOI: | 10.1007/978-3-030-87000-3_4 |