Unveiling the Capabilities of Latent Diffusion Models for Classification of Lung Diseases in Chest X-Rays

Diffusion models have demonstrated remarkable ability in synthesizing chest X-ray (CXR) images, particularly by generating high-quality samples to address the scarcity and imbalance of annotated CXR datasets. While these models excel in generating realistic samples-suggesting that they contain rich discriminative information effectively harnessing this capability for disease classification and decomposition remains a challenge. This study investigates an approach that leverages latent conditional diffusion models, which are conditioned on corresponding radiology reports, for lung disease classification in CXRs. Specifically, we employ a pre-trained latent conditional diffusion model for CXRs to predict noise estimates for a noisy input lung CXR under various disease conditions. By comparing the noise estimation errors associated with different class prompts, we determine the most probable disease classification based on the minimal error. Through the experiments, we demonstrate that the CXR diffusion-based classifier not only achieves zero-shot classification performance comparable to existing models but also reveals lesion regions aligning with ground truth lesion areas in CXRs.