Related Experiment Video
Updated: Jan 8, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.3K
Uncertainty- and hardness-weighted loss functions for medical image segmentation
Yanyan Zheng1, Yabo Wu2, Jie Chen1
1Wenzhou Third Clinical Institute Affiliated to Wenzhou Medical University, Third Affiliated Hospital of Shanghai University, Wenzhou People's Hospital, Wenzhou 32500, China.
Engineering Applications of Artificial Intelligence
|December 12, 2025
Summary
This study introduces novel weighted loss functions for medical image segmentation, improving accuracy by focusing on difficult pixels and reducing prediction uncertainty. These new methods enhance segmentation performance across diverse datasets.
Area of Science:
- Medical Image Analysis
- Deep Learning
- Computer Vision
Background:
- Accurate medical image segmentation is crucial for diagnosis and treatment planning.
- Current deep learning models often struggle with segmentation errors, especially at object boundaries, due to unaddressed pixel-level uncertainties and hardness.
- Existing loss functions do not adequately account for variations in prediction difficulty across pixels.
Purpose of the Study:
- To develop novel uncertainty- and hardness-weighted loss functions for improved medical image segmentation.
- To address the limitations of existing loss functions in handling pixel-level prediction uncertainties and hardness.
- To enhance the accuracy of deep learning-based segmentation models, particularly in challenging boundary regions.
Main Methods:
- Introduced two pixel-wise weighting schemes: probability-guided uncertainty (PGU) and region-enhanced hardness (REH) weights.
- Derived weights from the differences between network predictions and ground truths to emphasize difficult pixels.
- Integrated the novel loss functions with Swin-Unet and V-Net architectures for 2D and 3D segmentation tasks.
- Validated the approach on four diverse datasets: REFUGE, RETA, OCT, and ASC.
Main Results:
- The proposed uncertainty- / hardness-weighted loss functions significantly outperformed classical loss functions like cross-entropy (CE) and Dice losses.
- Demonstrated superior performance across multiple datasets and segmentation tasks, indicating effectiveness and generalization.
- Achieved more accurate segmentation, especially in object boundary regions, by effectively handling pixel-level uncertainties.
Conclusions:
- The developed uncertainty- / hardness-weighted loss functions represent a significant advancement in medical image segmentation.
- These novel weighting schemes effectively address pixel-level prediction uncertainties and hardness, leading to improved segmentation accuracy.
- The proposed methods show strong potential for enhancing various deep learning-based medical image analysis applications.
