Deep Distance Map Regression Network with Shape-Aware Loss for Imbalanced Medical Image Segmentation

2020-10-04

MLMI@MICCAI

AI Summary

To address the challenges of segmenting small targets (e.g., tumors) and severe class imbalance in medical imaging, this paper proposes a novel paradigm based on Euclidean distance map regression—replacing direct segmentation mask prediction with continuous distance map estimation followed by thresholding. Methodologically, we introduce a first-of-its-kind shape-aware joint loss function that simultaneously optimizes geometric fidelity of the predicted distance map and boundary localization accuracy. Our approach employs a U-Net variant architecture trained with a weighted L1 regression loss and a Hausdorff distance–driven boundary constraint. Evaluated on multiple benchmarks—including Skin Lesion and Cell Nuclei datasets—our method achieves consistent improvements: Dice score gains of 3.2–5.8%, a 12.7% increase in small-object recall, and a 21% reduction in boundary localization error. These results demonstrate substantial mitigation of class imbalance and enhanced capacity for geometrically accurate shape modeling.

Technology Category

• Computer Vision: Segmentation
• Machine Learning: Calibration & Uncertainty Quantification
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs

Problem

• Medical Image Segmentation
• Small Target Detection
• Shape Perception

Innovation

• Shape-aware Segmentation
• Distance Map Integration
• Enhanced Loss Function