Unsupervised Anatomical Feature Learning via Diffusion Models: Enhanced Medical Image Segmentation with Denoising Diffusion Probabilistic Models

Medical image segmentation typically relies on large, expert-annotated datasets, which are expensive and time-consuming to produce. Traditional U-Net architectures often struggle with boundary delineation and generalization because they focus on local texture patterns rather than global anatomical structure. This paper investigates whether unsupervised pretraining using Denoising Diffusion Probabilistic Models (DDPMs) can force a neural network to learn meaningful anatomical priors from unlabeled data, thereby improving segmentation performance and data efficiency.

The authors trained a DDPM on 21 unlabeled abdominal CT scans to learn structural representations through an iterative denoising process. By training the model to reverse noise corruption, the encoder is forced to learn global anatomical relationships, organ shapes, and boundary characteristics. After this unsupervised pretraining, the encoder weights were transferred to a U-Net architecture for downstream segmentation tasks on the BTCV multi-organ dataset. The researchers evaluated three strategies: random initialization, fine-tuning the pretrained encoder, and keeping the pretrained encoder frozen.

Diffusion pretraining led to statistically significant improvements in segmentation accuracy and boundary precision. For liver segmentation, the Dice coefficient increased from 0.75 to 0.93, with a 66% reduction in Average Surface Distance. The models also demonstrated remarkable robustness in low-data scenarios, maintaining high performance even when using only 10% of the labeled data. Crucially, the frozen encoder experiments showed that the model could retain over 80% of fine-tuned performance without seeing any segmentation labels, confirming that the diffusion process successfully embeds anatomical knowledge into the network. Furthermore, the approach significantly reduced prediction variance, leading to more reliable and consistent clinical performance.