Related Experiment Video
Updated: Jul 12, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.8K
AHU-MultiNet: Adaptive loss balancing based on homoscedastic uncertainty in multi-task medical image segmentation
Shasha Liu1, Hailing Wang1, Yan Li1
1The MOE Research Center for Software/Hardware Co-Design Engineering, East China Normal University, Shanghai, China.
Computers in Biology and Medicine
|October 20, 2023
Summary
This study introduces AHU-MultiNet, a novel semi-supervised method for medical image segmentation that improves accuracy by focusing on ambiguous regions. The adaptive loss balancing strategy enhances segmentation performance on benchmark datasets.
Area of Science:
- Medical image analysis
- Computer-aided diagnosis
- Machine learning for healthcare
Background:
- Semi-supervised learning is crucial for medical image segmentation due to annotation challenges.
- Existing methods often neglect ambiguous regions, limiting segmentation accuracy.
- Multi-task learning can enhance feature extraction and improve segmentation performance.
Purpose of the Study:
- To propose a novel semi-supervised method, AHU-MultiNet, for improved medical image segmentation.
- To address limitations in handling ambiguous regions in current semi-supervised approaches.
- To introduce an adaptive loss balancing strategy for effective multi-task learning.
Main Methods:
- Developed AHU-MultiNet, a multi-task network with segmentation, signed distance, and contour detection tasks.
- Implemented an inter-task consistency mechanism for improved information extraction.
- Introduced an adaptive loss balancing strategy based on homoscedastic uncertainty to optimize task weighting.
Main Results:
- Auxiliary tasks enforce shape-priors, leading to more accurate segmentation masks.
- The adaptive loss balancing strategy effectively manages multi-task weighting.
- Achieved state-of-the-art performance on the 2018 Atrial Segmentation Challenge and 2017 Liver Tumor Segmentation Challenge benchmarks.
Conclusions:
- AHU-MultiNet significantly improves semi-supervised medical image segmentation.
- The proposed adaptive loss balancing strategy enhances multi-task learning effectiveness.
- The method demonstrates superior performance in segmenting challenging medical images.
