HEAL: Learning-Free Source Free Unsupervised Domain Adaptation for Cross-Modality Medical Image Segmentation

Growing demands for clinical data privacy and storage constraints have spurred advances in Source Free Unsupervised Domain Adaptation (SFUDA). SFUDA addresses the domain shift by adapting models from the source domain to the unseen target domain without accessing source data, even when target-domain labels are unavailable. However, SFUDA faces significant challenges: the absence of source domain data and label supervision in the target domain due to source free and unsupervised settings. To address these issues, we propose HEAL, a novel SFUDA framework that integrates Hierarchical denoising, Edge-guided selection, size-Aware fusion, and Learning-free characteristic. Large-scale cross-modality experiments demonstrate that our method outperforms existing SFUDA approaches, achieving state-of-the-art (SOTA) performance. The source code is publicly available at: https://github.com/derekshiii/HEAL.

💡 Research Summary

**
The paper introduces HEAL, a novel source‑free unsupervised domain adaptation (SFUDA) framework designed for cross‑modality medical image segmentation. Unlike most existing SFUDA methods that rely on self‑training, pseudo‑label refinement, or auxiliary networks, HEAL adopts a “learning‑free” paradigm: the pre‑trained segmentation model and a diffusion model are frozen, and all adaptation is performed at inference time. HEAL consists of three complementary modules.

1. Hierarchical Denoising (HD) refines the initial pseudo‑labels generated by the frozen segmentation model on the target data. First, voxel‑wise entropy is computed; voxels with entropy above a threshold τ₁ are discarded (entropy denoising). Second, a Normal‑Inverse Gaussian (NIG) prior is fitted to the entropy‑filtered labels, using regional entropy to adapt the NIG parameters (α, β, ω). Voxels whose NIG variance exceeds a second threshold τ₂ are also removed, yielding the final refined pseudo‑labels Y*ₜ. This two‑stage process mitigates both coarse and fine‑grained labeling errors.

2. Edge‑Guided Selection (EGS) leverages the stochastic nature of diffusion models. Six candidate images are synthesized conditioned on Y*ₜ. For each candidate, Canny edge maps are extracted and compared with the edge map of the conditioning label. The structural consistency score Sᵢ (intersection over the conditioning edge count) quantifies how well the candidate preserves the original structure. The candidate with the highest Sᵢ, denoted I_B, is selected as the most reliable source‑like sample.

3. Size‑Aware Fusion (SAF) dynamically combines the predictions from the HD‑refined pseudo‑labels and the EGS‑selected sample. Small anatomical structures benefit more from the refined pseudo‑labels, while large structures are better captured by the diffusion‑generated image. SAF computes class‑wise inverse‑volume weights λ_k and fuses the two predictions accordingly, producing the final segmentation b_Yₜ.

The authors evaluate HEAL on three public datasets: BraTS‑2021 (3‑D brain MRI with four modalities), Kvasir‑SEG (endoscopic images), and CVC‑ClinicDB (colonoscopy frames). Four domain‑shift directions are tested: T1→T1ce and T2→FLAIR for brain tumor segmentation, and Kvasir‑SEG↔CVC‑ClinicDB for polyp segmentation. Performance is measured with Dice coefficient and Average Surface Distance (ASD).

Results show that HEAL consistently outperforms four recent SFUDA baselines (ProtoContra, DPL, IAPC, UPL). For brain tumor segmentation, HEAL achieves mean Dice scores of 77.3 % (T1→T1ce) and 73.5 % (T2→FLAIR) with ASD around 2 mm, narrowing the gap to fully supervised training. For polyp segmentation, HEAL reaches Dice 81.8 % (Kvasir‑SEG→CVC‑ClinicDB) and 66.5 % (reverse direction), again substantially better than the “No Adaptation” baseline.

Ablation studies demonstrate the contribution of each component: entropy‑only denoising improves Dice modestly; adding NIG denoising yields further gains; incorporating EGS provides the largest jump; and the full HEAL (HD + EGS + SAF) attains the best results. Qualitative visualizations confirm that hierarchical denoising removes spurious predictions, while edge‑guided selection preserves global anatomical consistency.

The key advantage of HEAL is its learning‑free nature. Because the segmentation model’s parameters are never updated, the method incurs negligible computational overhead during deployment and eliminates any risk of privacy leakage from target‑domain fine‑tuning. The diffusion model, trained once on the source data, serves only as a generative prior for sample synthesis. However, training the diffusion model itself can be resource‑intensive, and the thresholds τ₁ and τ₂ require careful tuning for each new modality. Future work may explore adaptive threshold selection or lightweight diffusion alternatives.

In summary, HEAL offers a practical, privacy‑preserving solution for cross‑modality medical image segmentation, achieving state‑of‑the‑art SFUDA performance without any source data access or target‑domain training. Its modular design, strong empirical results, and inference‑only workflow make it a compelling candidate for real‑world clinical deployment where data sharing constraints are stringent.