a. Self-supervised pre-training of LFM
b. Downstream adaptation with attention-based multiple instance learning
c. Biomarker discovery via principal component analysis
Lesion patches from DeepLesion dataset used for LFM pretraining and Model weights of LFM
