Research map / Label-efficient and robust learning / Semi-supervised learning
A Semantic Knowledge Complementarity based Decoupling Framework for Semi-supervised Class-imbalanced Medical Image 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 SKCDF, a semi-supervised class-imbalanced multi-organ CT segmentation framework that decouples encoder/decoder data flows, adds a labeled-unlabeled cross-attention complementarity module, and auxiliary balanced segmentation heads.
Problem
Semi-supervised multi-organ segmentation in CT suffers because pseudo-labeled unlabeled data degrades decoders that depend on accurate labels, and class imbalance causes small organs to be under-segmented or ignored.
Approach
Decouple data flows so the shared encoder learns from all data while separate labeled and unlabeled decoders are trained on their respective data; use channel-wise cross-attention between labeled and unlabeled features so labeled data guides pseudo-labels and unlabeled data enriches labeled features; train auxiliary heads with voxel-frequency-based Bernoulli masks to emphasize small organs, with E
Key results
- On 20% labeled Synapse, SKCDF reaches 64.27% average Dice and 1.45 ASD, the best among compared methods (GenericSSL 60.88%).
- On 5% labeled AMOS, SKCDF reaches 53.81% average Dice, a 3.78% improvement over the prior state of the art.
- Ablation on Synapse: the full method improves average Dice by 19.61% over the pseudo-labeling baseline, with data flow decoupling contributing the largest gain (+12.74%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply SKCDF instead of GenericSSL | VNet semi-supervised segmentation · Synapse (20% labeled) | average Dice (%): 60.88 ± 0.7 -> 64.27 ± 1.36 | Helped |
| Apply SKCDF instead of GenericSSL | VNet semi-supervised segmentation · Synapse (20% labeled) | average ASD (voxel): 2.52 ± 0.4 -> 1.45 ± 0.09 | Helped |
| Apply SKCDF instead of GenericSSL | VNet semi-supervised segmentation · AMOS (5% labeled) | average Dice (%): 50.03 -> 53.81 | Helped |
| Apply SKCDF instead of GenericSSL | VNet semi-supervised segmentation · AMOS (5% labeled) | average ASD (voxel): 5.21 -> 5.97 | Hurt |
| Add data flow decoupling to the basic pseudo-labeling baseline (ablation) | SKCDF · Synapse (20% labeled) | average Dice (%): 44.66 ± 2.5 -> 57.4 ± 0.9 | Helped |
| Add semantic knowledge complementarity module to data flow decoupling (ablation) | SKCDF · Synapse (20% labeled) | average Dice (%): 57.4 ± 0.9 -> 59.17 ± 1.95 | Helped |
| Add auxiliary balanced segmentation head training strategy to data flow decoupling (ablation) | SKCDF · Synapse (20% labeled) | average Dice (%): 57.4 ± 0.9 -> 63.57 ± 1.21 | Helped |
| Add semantic knowledge complementarity module and auxiliary balanced segmentation head training strategy (ablation) | SKCDF data flow decoupling framework · Synapse (20% labeled) | average Dice (%): 44.66 ± 2.5 -> 64.27 ± 1.36 | Helped |
| Apply decoder EMA knowledge transfer (ablation) | SKCDF decoders; labeled decoder used for inference · Synapse (20% labeled) | average Dice: 53.71 ± 1.64 -> 57.21 ± 0.85 | Helped |
| Apply decoder EMA knowledge transfer (ablation) | SKCDF decoders; unlabeled decoder used for inference · Synapse (20% labeled) | average Dice: 50.53 ± 1.00 -> 57.40 ± 0.90 | Helped |
| Use unlabeled rather than labeled decoder for inference with EMA (ablation) | SKCDF · Synapse (20% labeled) | average Dice: 57.21 ± 0.85 -> 57.40 ± 0.90 | Helped |
| Use head-to-decoder-to-head knowledge transfer sequence instead of decoder-only transfer (ablation) | SKCDF segmentation heads and decoder · Synapse (20% labeled) | average Dice: 54.79 ± 1.88 -> 63.57 ± 1.21 | Helped |
Surprising findings
- Using the labeled decoder at inference without knowledge transfer scores higher than the unlabeled decoder (by 3.18% Dice), though knowledge transfer reverses this so the unlabeled decoder is best.
Limitations
- Evaluated only on two abdominal CT multi-organ datasets (Synapse, AMOS).
- Hyperparameters (lambda_u=10, omega_ema=0.99) set empirically; sensitivity not extensively studied.
- Synapse results averaged over only three seeds with limited test set (6 scans).
Experimental setup
Datasets: Synapse AMOS
Benchmarks: Synapse AMOS
Metrics: Dice Score Average Surface Distance
Baselines: UA-MT URPC CPS SS-Net DST DePL Adsh CReST SimiS CLD DHC GenericSSL
Builds on: DHC ABC GenericSSL AllSpark GuidedNet VNet
Closest papers by meaning
- S2S2: Semantic Stacking for Robust Semantic Segmentation in Medical Imaging (AAAI 2025 · similarity 0.87)
- Versatile Medical Image Segmentation Learned from Multi-Source Datasets via Model Self-Disambiguation (CVPR 2024 · similarity 0.87)
- Unleashing the Potential of SAM for Medical Adaptation via Hierarchical Decoding (CVPR 2024 · similarity 0.87)
- Towards Realistic Semi-supervised Medical Image Classification (AAAI 2025 · similarity 0.87)
- SGTC: Semantic-Guided Triplet Co-training for Sparsely Annotated Semi-Supervised Medical Image Segmentation (AAAI 2025 · similarity 0.87)
- DyCON: Dynamic Uncertainty-aware Consistency and Contrastive Learning for Semi-supervised Medical Image Segmentation (CVPR 2025 · similarity 0.87)
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.