This study explores the use of foundation models in deep learning for cancer imaging biomarker discovery. Foundation models, characterized by large-scale training on vast datasets, are trained using self-supervised learning to reduce the need for labeled data, which is often scarce in medicine. The researchers developed a convolutional encoder foundation model using 11,467 radiographic lesions from 2,312 patients. This model was evaluated in three clinically relevant applications: lesion anatomical site classification, lung nodule malignancy prediction, and non-small cell lung cancer (NSCLC) tumor prognostication. The foundation model outperformed conventional supervised and other state-of-the-art models, especially in limited dataset scenarios. It demonstrated improved performance, stability, and biological relevance, with strong associations to underlying biology. The study highlights the potential of foundation models in discovering new imaging biomarkers and accelerating their translation into clinical settings.This study explores the use of foundation models in deep learning for cancer imaging biomarker discovery. Foundation models, characterized by large-scale training on vast datasets, are trained using self-supervised learning to reduce the need for labeled data, which is often scarce in medicine. The researchers developed a convolutional encoder foundation model using 11,467 radiographic lesions from 2,312 patients. This model was evaluated in three clinically relevant applications: lesion anatomical site classification, lung nodule malignancy prediction, and non-small cell lung cancer (NSCLC) tumor prognostication. The foundation model outperformed conventional supervised and other state-of-the-art models, especially in limited dataset scenarios. It demonstrated improved performance, stability, and biological relevance, with strong associations to underlying biology. The study highlights the potential of foundation models in discovering new imaging biomarkers and accelerating their translation into clinical settings.