Research map / Label-efficient and robust learning / Semi-supervised learning
PH-Net: Semi-Supervised Breast Lesion Segmentation via Patch-wise Hardness
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes PH-Net, a semi-supervised breast ultrasound lesion segmentation framework that shields high-entropy (hard) patches during cutmix and adds hard-patch contrastive learning on a teacher-student model.
Problem
Semi-supervised breast ultrasound (BUS) lesion segmentation, where lesions have large scale/shape variation and ambiguous boundaries under speckle noise, and existing methods under-train hard regions.
Approach
Computes patch-wise hardness as average prediction entropy in the teacher, shields the hardest patches from patch cutmix so they keep their content, and applies contrastive learning restricted to top-k hard patches with reliable (dual-consistent) pixel samples and a memory bank.
Key results
- On UDIAT with ResNet-50 at 1/8 labeled partition, PH-Net reaches 87.76% Dice, outperforming AugSeg by 0.98% Dice.
- On BUSI with U-Net, PH-Net beats RA-UGMT by +0.48%, +0.47%, +0.38% Dice under 1/2, 1/4, 1/8 partitions.
- Ablation on UDIAT 1/4 partition: SupOnly 83.45% Dice rises to 86.23% with APA and HPC with memory bank.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use PH-Net instead of SupOnly | UDIAT · UDIAT 1/2, 1/4, and 1/8 | Dice (%): +3.14%, +2.78%, and +3.89% | Helped |
| use PH-Net instead of AugSeg | UDIAT · UDIAT 1/8 | Dice (%): +0.98% | Helped |
| use PH-Net with U-Net instead of RA-UGMT | BUSI · BUSI 1/2, 1/4, and 1/8 | Dice (%): +0.48%, +0.47%, and +0.38% | Helped |
| add adaptive patch augmentation to supervised-only training (ablation) | PH-Net on UDIAT 1/4 · UDIAT 1/4 | Dice (%): +1.89% | Helped |
| add hard-patch contrastive learning without a memory bank (ablation) | PH-Net on UDIAT 1/4 · UDIAT 1/4 | Dice (%): +2.48% | Helped |
| add a memory bank to APA + HPC (ablation) | PH-Net on UDIAT 1/4 · UDIAT 1/4 | Dice (%) | Helped |
| use hard-patch shielding instead of standard CutMix (ablation) | PH-Net on UDIAT 1/4 · UDIAT 1/4 | Dice (%): +2.38% | Helped |
| use standard CutMix instead of weak augmentation only (ablation) | PH-Net on UDIAT 1/4 · UDIAT 1/4 | Dice (%): +0.60% | Helped |
| shield hard patches instead of randomly selected patches (ablation) | PH-Net on UDIAT 1/4 · UDIAT 1/4 | Dice (%): 84.31; 86.23 | Helped |
| set patch shielding parameter β to 10%, 20%, 30%, or 40% rather than 0 (ablation) | PH-Net on UDIAT 1/4 · UDIAT 1/4 | Dice (%): 84.31; 84.69; 85.35; 86.23; 85.96 | Mixed |
| set confidence threshold γ to 0.70, 0.80, 0.85, or 0.95 rather than 0.90 (ablation) | PH-Net on UDIAT 1/4 · UDIAT 1/4 | Dice (%): 85.03; 85.61; 86.07; 86.23; 86.11 | Mixed |
Surprising findings
- Shielding easy patches instead of hard ones gives only +0.32% Dice, while random patch shielding gives +0.46%, versus +2.38% for shielding hard patches with patch cutmix.
Limitations
- Struggles with lesions having extremely ambiguous contours or minimal distinction from background (failure cases).
- Evaluated only on two small public BUS datasets (UDIAT, BUSI).
- Contrastive learning is inapplicable to few-class tasks per authors' claim about U2PL, not a limitation of their own method.
Experimental setup
Datasets: UDIAT BUSI
Benchmarks: UDIAT BUSI
Metrics: Dice IoU standard deviation Wilcoxon signed-rank p-value
Baselines: SupOnly CPS PS-MT U2PL iMAS AugSeg BCP PDF-UNet RA-UGMT
Builds on: mean teacher CutMix U2PL UniMatch
Closest papers by meaning
- CSC-PA: Cross-image Semantic Correlation via Prototype Attentions for Single-network Semi-supervised Breast Tumor Segmentation (CVPR 2025 · similarity 0.87)
- Semi-supervised TEE Segmentation via Interacting with SAM Equipped with Noise-Resilient Prompting (AAAI 2024 · similarity 0.83)
- Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation (CVPR 2025 · similarity 0.82)
- PersonalizedUS: Interpretable Breast Cancer Risk Assessment with Local Coverage Uncertainty Quantification (AAAI 2025 · similarity 0.82)
- Neighbor Does Matter: Density-Aware Contrastive Learning for Medical Semi-supervised Segmentation (AAAI 2025 · similarity 0.82)
- DyCON: Dynamic Uncertainty-aware Consistency and Contrastive Learning for Semi-supervised Medical Image Segmentation (CVPR 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.