Material composition characterization from computed tomography via self-supervised learning promotes pulmonary disease diagnosis

Computed tomography (CT) images primarily provide tissue morphological information, while material composition analysis may enable a more fundamental body assessment. However, existing methods for this suffer from low accuracy and severe degradation. Furthermore, the complex composition of bodies and the absence of labels constrain the potential use of deep learning. Here, we present a self-supervised learning approach, generating multiple basis material images with no labels (NoL-MBMI), for analyzing material composition without labels. Results from phantom and patient experiments demonstrate that NoL-MBMI can provide results with superior visual quality and accuracy. Notably, to extend the clinical usage of NoL-MBMI, we construct an automated system to extract material composition information directly from standard single-energy CT (SECT) data for diagnosis. We evaluate the system on two pulmonary diagnosis tasks and observe that deep-learning models using material composition features significantly outperform those using morphological features, suggesting the clinical effectiveness of diagnosing utilizing material composition and its potential for advancing medical imaging technology.