Research map / Detection and segmentation / Weakly supervised detection and segmentation
CSC-PA: Cross-image Semantic Correlation via Prototype Attentions for Single-network Semi-supervised Breast Tumor Segmentation
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 CSC-PA, a single-network semi-supervised BUS tumor segmentation framework with foreground and edge prototype attentions plus a pixel affinity loss, reporting SOTA on BUSI and UDIAT.
Problem
Semi-supervised breast ultrasound tumor segmentation suffers from scarce labels, variable lesion scale/shape, and ambiguous boundaries; mean-teacher methods learn only within-image correlations and depend heavily on the teacher.
Approach
A single-network semi-supervised framework that transfers foreground information across labeled and unlabeled images via channel-prototype attention (FPA), enhances lesion edges with an adaptive edge container and edge prototype (EPA), and enforces consistency of pixel affinity matrices between predictions and supervision (PAL).
Key results
- On UDIAT, CSC-PA improves IoU over the supervised-only baseline by +4.96%, +7.26%, +8.30% under 1/2, 1/4, 1/8 protocols.
- On BUSI with 1/4 labels, CSC-PA reaches 76.93% Dice (68.20% IoU), the best among compared methods.
- Ablation on BUSI 1/4: FPA, EPA and PAL together raise Dice from 71.49% (supervised only) to 76.93%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add FPA (ablation) | CSC-PA · BUSI 1/4 | Dice: 71.49±2.27 -> 76.20±2.73 | Helped |
| Add FPA (ablation) | CSC-PA · UDIAT 1/4 | Dice: 74.78±2.44 -> 78.79±3.35 | Helped |
| Add EPA to FPA (ablation) | CSC-PA · BUSI 1/4 | Dice: 76.20±2.73 -> 76.48±2.75 | Helped |
| Add EPA to FPA (ablation) | CSC-PA · UDIAT 1/4 | Dice: 78.79±3.35 -> 80.29±1.68 | Helped |
| Add PAL to the supervised baseline (ablation) | CSC-PA · BUSI 1/4 | Dice: 71.49±2.27 -> 74.19±3.09 | Helped |
| Add PAL to the supervised baseline (ablation) | CSC-PA · UDIAT 1/4 | Dice: 74.78±2.44 -> 75.91±3.96 | Helped |
| Use random edge-feature selection and linear-layer container updating instead of top-k selection and memory-bank updating (ablation) | CSC-PA · UDIAT 1/4 | Dice: 80.28±1.55 -> 81.26±1.75 | Helped |
| Use random rather than top-k edge-feature selection with memory-bank updating (ablation) | CSC-PA · UDIAT 1/4 | Dice: 80.28±1.55 -> 79.95±1.80 | Hurt |
| Use a linear layer rather than a convolutional layer for edge-container updating with top-k selection (ablation) | CSC-PA · UDIAT 1/4 | Dice: 80.45±2.22 -> 81.26±1.75 | Helped |
Limitations
- Fails under extremely low contrast in BUS images.
- Severe artifacts degrade segmentation performance.
- Evaluated only on two small BUS datasets.
Experimental setup
Datasets: BUSI UDIAT
Benchmarks: BUSI UDIAT
Metrics: Dice IoU
Baselines: U2PL AugSeg ABD CISC-R PS-MT UniMatch CorrMatch MFMSNet DK-HRS PH-Net
Builds on: Deeplabv3+ UniMatch CorrMatch mean-teacher
Closest papers by meaning
- PH-Net: Semi-Supervised Breast Lesion Segmentation via Patch-wise Hardness (CVPR 2024 · similarity 0.87)
- SGTC: Semantic-Guided Triplet Co-training for Sparsely Annotated Semi-Supervised Medical Image Segmentation (AAAI 2025 · similarity 0.85)
- Weakly Supervised Gland Segmentation with Class Semantic Consistency and Purified Labels Filtration (AAAI 2025 · similarity 0.84)
- Enhancing SAM with Efficient Prompting and Preference Optimization for Semi-supervised Medical Image Segmentation (CVPR 2025 · similarity 0.84)
- Boost the Inference with Co-training: A Depth-guided Mutual Learning Framework for Semi-supervised Medical Polyp Segmentation (CVPR 2025 · similarity 0.84)
- FAMNet: Frequency-aware Matching Network for Cross-domain Few-shot Medical Image Segmentation (AAAI 2025 · similarity 0.84)
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.