Related Experiment Video
Updated: Jun 23, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
DuoMod-Net: Logarithmic balancing and geometric refinement for imbalanced semi-supervised medical image segmentation
Wang Bo1,2, Along He3, Ting Xue4
1School of Biomedical Engineering, Division of Life Sciences and Medicine, University of Science and Technology of China, Hefei, Anhui, P.R. China.
Patterns (New York, N.Y.)
|June 22, 2026
Summary
DuoMod-Net tackles class imbalance in medical image segmentation by using relative logarithmic modulation and disagreement-driven adaptive feature refinement. This improves learning for rare classes and enhances detection reliability.
Area of Science:
- Medical Imaging
- Computer Vision
- Machine Learning
Background:
- Class imbalance in semi-supervised medical image segmentation hinders learning for underrepresented classes and biases model training towards the background.
- This imbalance leads to compromised feature learning and unreliable segmentation results, particularly for critical anatomical structures.
Purpose of the Study:
- To introduce a novel framework, DuoMod-Net, to effectively address class imbalance challenges in semi-supervised medical image segmentation.
- To improve feature learning for tail classes and enhance the overall reliability and generalization of segmentation models.
Main Methods:
- Developed DuoMod-Net, a synergistic framework with two key components: Relative Logarithmic Modulation (RLM) and Disagreement-Driven Adaptive Feature Refinement (DAFR).
- RLM decouples background magnitude from foreground balancing, using logarithmic scaling anchored by percentiles to preserve foreground organ dynamics.
- DAFR employs inter-model disagreement for geometric regularization, expanding the feature space during training to refine decision boundaries, which is removed during inference.
Main Results:
- DuoMod-Net demonstrated substantial improvements in segmenting tail classes across various data regimes (5%, 10%, 20%).
- The framework significantly increased detection reliability by minimizing catastrophic failures and maintaining a safety margin.
- Achieved robust zero-shot generalization capabilities on unseen medical image datasets.
Conclusions:
- DuoMod-Net effectively mitigates the dual challenges of class imbalance in semi-supervised medical image segmentation.
- The proposed method enhances performance on underrepresented classes and improves the robustness and generalizability of segmentation models.
- DuoMod-Net offers a promising solution for reliable medical image analysis in data-scarce scenarios.