DiffSegLung: Diffusion Radiomic Distillation for Unsupervised Lung Pathology Segmentation

2026-05-12

AI Summary

This work addresses the challenges of unsupervised pathological segmentation in lung CT scans, which are hindered by the scarcity of multi-disease annotations and the difficulty of diffusion models in effectively leveraging physically meaningful Hounsfield Unit (HU) signals. To overcome these limitations, the authors propose a Diffusion Radiomic Distillation framework that, for the first time, employs handcrafted radiomic features as physics-informed teacher signals to shape representations in the bottleneck layer of a 3D diffusion U-Net via contrastive learning, enabling annotation-free learning of pathology-discriminative features. During inference, segmentation boundaries are refined through a combination of HU-guided Gaussian mixture model clustering and Sobel-enhanced diffusion fusion. Evaluated on 190 expert-annotated slices across four heterogeneous CT cohorts, the method consistently outperforms existing baselines in unsupervised segmentation of four pulmonary pathologies while simultaneously improving CT synthesis fidelity.

Abstract

Unsupervised segmentation of pulmonary pathologies in CT remains an open challenge due to the absence of annotated multi pathology cohorts and the failure of existing diffusion-based methods to exploit the quantitative Hounsfield Unit (HU) signal that physically distinguishes tissue classes. To address this, we propose DiffSegLung,a framework that introduces Diffusion Radiomic Distillation, in which handcrafted radiomic descriptors serve as a physics grounded teacher to shape the bottleneck of a 3D diffusion U-Net via a contrastive objective, transferring pathology discriminative structure into the learned representation without any annotations. At inference, the teacher is discarded and multitimestep bottleneck features are clustered by a Gaussian Mixture Model with HU-guided label assignment, followed by Sobel Diffusion Fusion for boundary refinement. Evaluated on 190 expert annotated axial slices drawn from four heterogeneous CT cohorts, Diff-SegLung improves segmentation across all four pathology classes over unsupervised baselines and improves generation fidelity over prior CT diffusion models.

Problem

• unsupervised segmentation
• lung pathology
• CT imaging
• Hounsfield Unit
• diffusion models

Innovation

• Diffusion Radiomic Distillation
• Unsupervised Segmentation
• Radiomic Descriptors
• Hounsfield Unit
• 3D Diffusion U-Net