Fig. 7: Overview of the self-supervised learning and two-step training framework for biomarker prediction.
From: Predicting ROS1 and ALK fusions in NSCLC from H&E slides with a two-step vision transformer approach
A Splitting the ROS1 dataset into a cross-validation (CV) dataset and a holdout dataset, followed by training the MoCo-v3 model using the CV dataset. The trained MoCo-v3 model maps each tile to a feature vector of dimensions 1 × 384. These vectors are subsequently utilized to train a vision transformer-based model through a two-step process. Q query encoder, K key encoder, S similarity metric, CL contrastive loss, ViT vision transformer. B Dataset relabeling and two-step model training for ROS1 prediction, with a similar methodology applied for ALK prediction. At each fold, the negative samples in the training data are randomly downsampled to a size five times that of the positive cases. The validation and test sets remain unchanged.
