Figure 1.

attention-based multiple instance learning (aMIL) models use feature vectors as inputs, grouped in bags, to make predictions aggregated from all vectors within a bag. (A) Number of slides in the training and test cohorts by pathologic category. (B) Models were pre-trained with histology-specific digital images using unsupervised domain-specific learning to extract features with CTransPath. (C) Whole slide images (WSI) were divided into tiles, passed through the fine-tuned network to generate neuroblastoma-specific feature vectors, which are divided into bags per WSI. The aMIL network assigns attention scores to vectors, and a slide-level prediction is determined based on the aggregated predictions weighted by attention scores.