Related Experiment Video
Updated: Sep 11, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.9K
Hard exudates segmentation for retinal fundus images based on longitudinal multi-scale fusion network
Shuang Liu1, Xiangyu Jiang1, Jie Zhang1
1School of Electronic and Information Engineering, Soochow University, Suzhou, 215006, China.
Medical & Biological Engineering & Computing
|August 12, 2025
Summary
A new deep learning model, the longitudinal multi-scale fusion network (LMSF-Net), enhances hard exudate segmentation in retinal images. This method improves the detection of small and ambiguous lesions for better early diagnosis of retinal diseases.
Area of Science:
- Ophthalmology
- Medical Imaging
- Computer Vision
Background:
- Accurate segmentation of hard exudates in fundus images is critical for early detection of retinal diseases.
- Current segmentation methods face challenges with small lesions and ambiguous boundaries.
Purpose of the Study:
- To propose a novel deep learning model, the longitudinal multi-scale fusion network (LMSF-Net), for accurate hard exudate segmentation.
- To improve the segmentation of hard exudates across various scales and shapes in fundus images.
Main Methods:
- The LMSF-Net incorporates an adjacent complementary correction module (ACCM) for feature fusion in the encoding path.
- A progressive iterative fusion module (PIFM) is utilized in the decoding path for adjacent feature fusion.
- A spatial awareness fusion module (SAFM) calibrates and aggregates decoding outputs.
Main Results:
- The LMSF-Net demonstrated superior performance in hard exudate segmentation.
- Achieved Area Under the Precision-Recall curve (AUPR) scores of 0.6954, 0.9017, and 0.6745 on the DDR, IDRID, and E-Ophtha EX datasets, respectively.
Conclusions:
- The proposed LMSF-Net effectively addresses challenges in hard exudate segmentation.
- The method shows significant potential for improving the early diagnosis of retinal diseases through enhanced image analysis.
