A Lung CT Foundation Model Facilitating Disease Diagnosis and Medical Imaging

The concomitant development and evolution of lung computed tomography (CT) and artificial intelligence (AI) has allowed non-invasive lung imaging to be a key part of the clinical care of patients with major diseases, such as lung cancer. However, the paucity of labeled lung CT data has limited the training highly efficacious AI models and thereby has retarded broad-scale adoption and deployment of AI-based lung CT imaging in the real-world clinical setting. In this paper, We introduce LCTfound, a foundational model that encodes images along with correlated clinical information, into a neural network. LCTfound used self-supervised learning pre-trained by diffusion models using a large dataset containing 105,184 lung CT scans (totaling more than 28 million images) from multiple centers. LCTfound was evaluated on 8 categories of lung CT tasks, ranging from scanning-level clinical diagnosis to pixel-level image restoration, including segmentation of mediastinal neoplasm, diagnosis of pulmonary alveolar proteinosis, prognosis of non-small cell lung cancer, prediction of major pathological response to neoadjuvant chemoimmunotherapy, whole lung 3D modeling for surgical navigation, virtual lung computed tomography angiography(CTA), reconstruction of lung CT from sparse views, and enhancement of low-dose CT images. Equipped with the robust few-shot learning capability, LCTfound outperformed the previously state-of-the-art pre-trained models in all the above tasks. LCTfound is a major advancements in self-supervised representation learning on lung CT, laying the groundwork for a foundational model that operates with high efficacy across the spectrum of low-level and high-level clinical tasks and serving a dual purpose in aiding in clinical diagnosis of lung diseases and improving the quality of lung CT imaging.