Abstract
This paper introduces DP-Net, the first framework to leverage the rich semantic features of Diffusion Probabilistic Models (DPMs) for Unsupervised Domain Adaptation (UDA) in medical imaging. By combining feature-level alignment within a diffusion process and a prototype-guided consistency module, DP-Net achieves state-of-the-art performance in segmenting fundus images across different clinical datasets without requiring target-domain labels. Our results demonstrate that diffusion-based representations are more robust and semantically richer than standard convolutional features for bridging domain gaps.
Introduction
In medical AI, models trained on data from a specific source domain (e.g., one hospital) often suffer from significant performance degradation when deployed in a target domain (e.g., another hospital). This "domain shift" is driven by variations in imaging equipment, lighting conditions, and patient demographics. Unsupervised Domain Adaptation (UDA) seeks to mitigate this by aligning source and target distributions. However, conventional CNNs often fail to capture invariant semantic features, and the lack of target-domain labels necessitates reliance on noisy pseudo-labels, which can lead to error propagation.
Methodology
DP-Net addresses these challenges through a two-stage framework that utilizes generative diffusion models as powerful feature extractors.
Stage 1: Distribution Aligned Diffusion (DADiff)
Rather than using DPMs solely for image generation, DP-Net extracts intermediate latent activations that contain deep semantic information. A domain discriminator with a Gradient Reversal Layer (GRL) is integrated into the diffusion process. This forces the model to produce domain-invariant features, ensuring that representations from the source and target datasets are indistinguishable to the discriminator.
Stage 2: Prototype-guided Consistency Learning (PCL)
To refine segmentation without target labels, the model employs two key strategies:
- Prototypes: The network calculates "feature centroids" for each anatomical class (e.g., optic disc and cup). A consistency loss minimizes the distance between these prototypes across domains, acting as a regularizer.
- Uncertainty Denoising: To handle noisy pseudo-labels, Monte Carlo Dropout is used to measure prediction variance. By filtering out unreliable, high-variance predictions, the model ensures it only trains on stable and accurate pseudo-labels.
Experimental Results
DP-Net was evaluated on three public fundus image datasets for glaucoma screening: REFUGE (source), RIM-ONE-r3 (target), and Drishti-GS (target). The primary task was the automated segmentation of the optic disc and optic cup.
Quantitative Performance
DP-Net outperformed seven state-of-the-art UDA methods. Performance was measured using the Dice coefficient, where 1.0 represents a perfect match. On the RIM-ONE-r3 dataset, DP-Net achieved a Dice score of 0.913 for the optic disc and 0.852 for the optic cup, surpassing the previous best method, CLR. On Drishti-GS, the model tied for the top spot in disc segmentation (0.966) and remained highly competitive in cup segmentation.
Ablation Studies
Ablation experiments confirmed that both the DADiff and PCL modules are essential. Removing either component led to a significant drop in accuracy, proving that feature alignment and prototype consistency are complementary mechanisms for robust adaptation.
Conclusion and Future Work
DP-Net demonstrates that the internal representations of diffusion models are ideally suited for the "heavy lifting" of domain adaptation. Despite being trained on smaller medical datasets compared to competitors using ImageNet-pretrained backbones, DP-Net achieved superior results. Future work will explore applying this diffusion-based UDA approach to 3D medical data, such as MRI or CT scans, where domain shift remains a significant barrier to clinical deployment.
Summary updated