Research map / Label-efficient and robust learning / Semi-supervised learning
beta-FFT: Nonlinear Interpolation and Differentiated Training Strategies for Semi-Supervised 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 β-FFT, a semi-supervised medical image segmentation method that mitigates co-training homogenization via FFT low-frequency swapping between augmented images and differentiated training (extra labeled/Beta-interpolation training for one student).
Problem
Co-training for semi-supervised medical image segmentation suffers from homogenization, where multiple models converge to similar decision boundaries, reducing diversity and generalization.
Approach
Swap low-frequency Fourier components between weakly and strongly augmented images to create smooth intermediate training samples, and apply extra training (supervised plus linear/noise interpolation consistency) to only one student model so the two students stay differentiated.
Key results
- On ACDC with 10% labeled data, β-FFT reaches 90.50±0.04% Dice and 83.12±0.12% Jaccard, above ABD (89.81 Dice) and AD-MT (89.46 Dice).
- On PROMISE12 with 20% labeled data, β-FFT achieves 83.75±0.65% Dice and 1.20±0.07 ASD, versus ABD 82.06 Dice.
- On MS-CMRSeg 2019 with 20% labeled data, β-FFT achieves 87.79% Dice, versus ABD 87.25 and DiffRect 86.78.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use β-FFT instead of ABD | semi-supervised segmentation · ACDC test, 5% labeled | DSC: 88.96 -> 89.46±0.12 | Helped |
| Use β-FFT instead of ABD | semi-supervised segmentation · ACDC test, 10% labeled | DSC: 89.81 -> 90.50±0.04 | Helped |
| Use β-FFT instead of ABD | semi-supervised segmentation · PROMISE12 test, 20% labeled | DSC: 82.06 -> 83.75±0.65 | Helped |
| Use β-FFT instead of ABD | semi-supervised segmentation · MS-CMRSEG 2019, 20% labeled | Dice: 87.25 -> 87.79±0.04 | Helped |
| Add low-frequency component exchange with H=30x30 (ablation) | β-FFT on ACDC · ACDC test, 10% labeled | Dice: 89.60 -> 89.96 | Helped |
| Add LICR(w) to Student Model 1 (ablation) | β-FFT on ACDC · ACDC test, 10% labeled | Dice: 89.60 -> 90.24 | Helped |
| Add NICR(w) to Student Model 1 (ablation) | β-FFT on ACDC · ACDC test, 10% labeled | Dice: 89.60 -> 90.23 | Helped |
| Add LICR(w) and LICR(s) to Student Model 1 (ablation) | β-FFT on ACDC · ACDC test, 10% labeled | Dice: 89.60 -> 90.54 | Helped |
| Apply LICR(w) and LICR(s) to both student models instead of only Student Model 1 (ablation) | β-FFT on ACDC · ACDC test, 10% labeled | Dice: 90.54 -> 90.10 | Hurt |
| Use Beta(0.1,0.1) for sample mixing (ablation) | β-FFT, differentiated training on Student Model 1 · ACDC | Dice: 90.77 / 83.81; 90.54 / 83.23 | Helped |
Surprising findings
- Applying LICR to both student models simultaneously performed worse than applying it to only one student, indicating a single-model differentiation is more effective.
- Smaller Beta parameters (Beta(0.1,0.1)) gave better results, with performance peaking at the smallest tested values.
Limitations
- Ablations were run only on ACDC with 10% labels, so generalization of design choices to other settings is untested.
- Low-pass filter size selection favored different sizes on validation vs test sets, suggesting sensitivity to hyperparameters.
- Reported gains over strong baselines are often small (e.g., ~0.5-1 Dice points on ACDC) relative to reported standard deviations.
Experimental setup
Datasets: ACDC PROMISE12 MS-CMRSeg 2019
Benchmarks: ACDC PROMISE12 MS-CMRSeg 2019
Metrics: Dice Jaccard 95HD ASD
Baselines: UA-MT DTC URPC MC-Net SS-Net SCP-Net Cross Teaching BCP DiffRect ABD AD-MT PLGCL
Builds on: BCP ABD Cross-Teaching Mixup FDA
Closest papers by meaning
- Steady Progress Beats Stagnation: Mutual Aid of Foundation and Conventional Models in Mixed Domain Semi-Supervised Medical Image Segmentation (CVPR 2025 · similarity 0.86)
- Semi-supervised TEE Segmentation via Interacting with SAM Equipped with Noise-Resilient Prompting (AAAI 2024 · similarity 0.86)
- SGTC: Semantic-Guided Triplet Co-training for Sparsely Annotated Semi-Supervised Medical Image Segmentation (AAAI 2025 · similarity 0.85)
- Constructing and Exploring Intermediate Domains in Mixed Domain Semi-supervised Medical Image Segmentation (CVPR 2024 · similarity 0.85)
- Dual-calibrated Co-training Framework for Personalized Federated Semi-Supervised Medical Image Segmentation (AAAI 2025 · similarity 0.85)
- Combinatorial CNN-Transformer Learning with Manifold Constraints for Semi-supervised Medical Image Segmentation (AAAI 2024 · 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.