Related Experiment Video
Updated: Aug 15, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.9K
TSSK-Net: Weakly supervised biomarker localization and segmentation with image-level annotation in retinal OCT images
Xiaoming Liu1, Qi Liu1, Ying Zhang2
1School of Computer Science and Technology, Wuhan University of Science and Technology, Wuhan, 430070, China; Hubei Province Key Laboratory of Intelligent Information Processing and Real-Time Industrial System, Wuhan, 430070, China.
Computers in Biology and Medicine
|December 30, 2022
Summary
This study introduces TSSK-Net, a novel method for segmenting retinal biomarkers in OCT images using only image-level annotations. It significantly reduces the annotation burden for ophthalmologists while maintaining high accuracy.
Area of Science:
- Ophthalmology
- Medical Imaging
- Artificial Intelligence
Background:
- Accurate segmentation of biomarkers in Optical Coherence Tomography (OCT) images is crucial for diagnosing retinal diseases.
- Fully supervised deep learning models require extensive pixel-level annotations, which are time-consuming and costly.
- Weakly supervised methods using image-level annotations offer a reduced annotation burden but face challenges like model collapse and anatomical mismatch.
Purpose of the Study:
- To develop a novel weakly supervised method for localizing and segmenting biomarkers in OCT images using only image-level annotations.
- To reduce the manual annotation workload for ophthalmologists in retinal disease diagnosis.
- To overcome limitations of existing weakly supervised methods, such as training instability and anatomical mismatch.
Main Methods:
- Proposed a Teacher-Student network named TSSK-Net, integrating self-supervised contrastive learning and knowledge distillation-based anomaly localization.
- Introduced a pre-training strategy using supervised contrastive learning to learn normal OCT image anatomy.
- Designed a fine-tuning module with a hybrid network structure incorporating supervised contrastive loss and cross-entropy loss, combined efficiently to preserve anatomical structure and enhance feature representation.
- Developed a knowledge distillation-based anomaly segmentation method to address insufficient supervision.
Main Results:
- Experimental results on local and public datasets demonstrated the effectiveness of the proposed TSSK-Net method.
- The method successfully reduced the annotation burden for ophthalmologists.
- Achieved accurate localization and segmentation of pathological regions in OCT images with only image-level annotations.
Conclusions:
- TSSK-Net offers an effective solution for weakly supervised biomarker localization and segmentation in OCT images.
- The approach significantly alleviates the annotation burden in clinical practice.
- This method holds promise for improving the efficiency and accuracy of retinal disease diagnosis using OCT imaging.
