Directional Dual-Head Supervision for Real-Time Source-Free Polyp Segmentation
Organizations: University of Science, Ho Chi Minh City, Vietnam · Vietnam National University, Ho Chi Minh City, Vietnam
Abstract
Source-free domain adaptation (SFDA) is critical for medical image segmentation as it adapts a model to an unlabeled target domain without accessing source data, thereby preserving data privacy. However, its effectiveness is often limited by confirmation bias: self-training models tend to miss small polyps, hallucinate foreground, and produce imprecise contours. We propose MedRT-SFSeg, a real-time SFDA framework built on a dual-head one-to-one/one-to-many YOLO segmentation architecture. Mask-DHF fuses precision- and recall-oriented predictions into stable pseudo-instances, while DURR uses the sign of pixel-level dual-head disagreement as a directional cue to expand or shrink masks, and additionally recovers missed instances and suppresses hallucinations. Seg-MARD further regularizes feature representations through mask-guided, region-stratified sampling. Experiments on bidirectional adaptation between Kvasir-SEG and CVC-ClinicDB show that MedRT-SFSeg achieves 80.43%/80.77% Dice and 5.53/8.44 mm ASD, improving the non-adapted baseline by 7.7 and 9.8 Dice points while running at more than 62 FPS end-to-end. These results suggest that segmentation-aware dual-head supervision can improve pseudo-label quality without sacrificing real-time inference. Project page: https://danielnguyen-05.github.io/MedRT-SFSeg
Figures & tables
| Method | Kvasir-SEG CVC-ClinicDB | CVC-ClinicDB Kvasir-SEG | ||
| Dice (%) | ASD (mm) | Dice (%) | ASD (mm) | |
| Baseline | 72.8 33.0 | 5.9 12.5 | 70.9 25.2 | 9.8 18.5 |
| Supervised | 96.0 4.3 | 0.6 0.6 | 90.3 16.6 | 3.7 10.2 |
| ProtoContra [ 16 ] | 78.0 12.1 | 0.9 0.6 | 68.9 26.4 | 9.5 15.0 |
| DPL [ 3 ] | 67.2 24.0 | 8.0 12.4 | 66.1 33.1 | 7.8 17.7 |
| IAPC [ 2 ] | 73.3 29.6 | 4.4 7.9 | 61.2 26.3 | 13.7 13.2 |
| Kvasir-SEG CVC-ClinicDB | CVC-ClinicDB Kvasir-SEG | |||||||
| Method | Dice | IoU | ASD | HD95 | Dice | IoU | ASD | HD95 |
| Baseline | 72.75 | 65.54 | 5.92 | 38.32 | 70.94 | 63.07 | 9.81 | 68.29 |
| Mask-DHF | 75.83 | 68.17 | 5.78 | 37.85 | 75.04 | 66.92 | 9.25 | 68.34 |
| Mask-DHF+Seg-MARD | 77.89 | 70.02 | 5.62 | 37.79 | 77.59 | 69.21 | 8.84 | 68.07 |
| Mask-DHF+DURR | 79.02 | 71.92 | 5.58 | 37.77 | 79.33 | 71.04 | 8.59 | 67.83 |
| Full MedRT-SFSeg | 80.43 | 72.57 | 5.53 | 36.92 | 80.77 | 72.52 | 8.44 | 62.74 |
| Metric | Kvasir ClinicDB | ClinicDB Kvasir |
| Full-graph params (M) | 11.43 | 11.43 |
| Fused deployment params (M) | 10.37 | 10.37 |
| Deployment GFLOPs | 34.05 | 34.05 |
| Checkpoint size (MB) | 22.32 | 22.34 |
| Network mean latency (ms) | 11.81 | 11.49 |
| Network FPS | 84.69 | 87.01 |
Polyp Segmentation
Polyp segmentation uses computer vision to delineate colorectal polyps in colonoscopy images and video for detection and clinical assessment.
See all 25 papers