A Self-Supervised Foundation Model Based on Three-Dimensional Chest CT Scans for Lung Cancer Diagnosis and Prognosis Prediction.

Purpose To develop a self-supervised chest CT foundation model and evaluate its performance in lung cancer clinical tasks. Materials and Methods In this retrospective multicenter study, the authors developed the Unified CT-Based Lung Cancer Imaging Foundation (UCLIF) model using self-supervised learning on 33 901 three-dimensional chest CT scans acquired between June 1958 and February 2019. The model was pretrained with a contrastive masked image modeling task and then fine-tuned for lung cancer histologic subtype classification, cancer staging, survival, and recurrence prediction using multicenter patient datasets. Histopathology, TNM stage, and follow-up outcomes served as reference standards. UCLIF was compared with mainstream deep learning and machine learning algorithms, and model performance was assessed by accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC); superiority was tested using the DeLong test. Results A total of 656 patients were included for downstream evaluation (mean age, 68.55 years ± 10.01 [SD]; 450 males). Compared with self-supervised pretraining on natural images or single tumor regions, UCLIF achieved superior performance (DeLong test,