DiffSegLung uses diffusion radiomic distillation for unsupervised lung segmentation
No annotations needed: new AI segments lung pathologies from CT scans with expert-level accuracy.
DiffSegLung tackles a key bottleneck in medical imaging: the scarcity of annotated datasets for lung pathology segmentation. The framework introduces diffusion radiomic distillation, where handcrafted radiomic features (like texture and shape descriptors) act as a teacher during training, guiding the latent bottleneck of a 3D diffusion U-Net via a contrastive objective. This lets the model learn pathology-discriminative representations without any manual labels. At inference, the teacher is discarded; features from multiple timesteps are clustered with a Gaussian Mixture Model, using Hounsfield Units for label assignment, and refined via Sobel Diffusion Fusion for cleaner boundaries.
On 190 expert-annotated axial slices from four heterogeneous CT cohorts, DiffSegLung improved segmentation across all four pathology classes over existing unsupervised methods. It also enhanced generation fidelity compared to prior CT diffusion models. This work, from researchers at Université Paris-Saclay and Université Sorbonne Paris Nord, is published on arXiv (paper 2605.11758). By eliminating the need for annotations, DiffSegLung could accelerate lung disease diagnosis and large-scale screening, especially for rare or understudied pathologies.
- Framework uses diffusion radiomic distillation: handcrafted radiomic descriptors teach a 3D diffusion U-Net without annotations.
- Segmentation clusters multi-timestep bottleneck features with Gaussian Mixture Model, guided by Hounsfield Unit values for tissue distinction.
- Evaluated on 190 expert-labeled axial slices from 4 CT cohorts, outperforming unsupervised baselines across all pathology classes.
Why It Matters
Enables zero-shot lung pathology segmentation from CT scans, reducing reliance on expensive, scarce annotated medical data.