Pretraining dataset and model weights of CT Lesion Foundation Model

This repository provides pretraining data patches, pretrained model weights, and fine-tuned model weights associated with the study (https://github.com/xiawei999000/LFM-HCC): A CT Lesion Foundation Model for Non-invasive Assessment of Hepatocellular Carcinoma Aggressiveness and Prediction of Recurrence Risk The lesion foundation model (LFM) was developed using a DINOv2-style self-supervised learning framework to learn transferable lesion-level representations from contrast-enhanced CT images. The pretrained LFM encoder was further adapted for hepatocellular carcinoma (HCC) aggressiveness assessment, including microvascular invasion (MVI) and pathological grade assessment, and for recurrence-free survival (RFS) risk prediction. Repository contents deeplesion_patches.rarGeneric lesion CT patches extracted from the DeepLesion dataset and used for lesion-level self-supervised pretraining. HCC_patches.rarHCC-specific CT lesion patches extracted from public HCC CT datasets and used for HCC-specific lesion-level self-supervised pretraining. pretrained_models.rarPretrained LFM encoder weights obtained from different pretraining strategies, including generic lesion pretraining, mixed generic/HCC lesion pretraining, and sequential generic-to-HCC lesion pretraining. finetuned_models.rarFine-tuned model weights for downstream HCC tasks, including MVI assessment, pathological grade assessment, and LFM-derived feature extraction for recurrence-free survival prediction.