Fig. 1: Overview of the AI-based triaging method.
From: Artificial intelligence-based triaging of cutaneous melanocytic lesions
Each case, usually consisting of multiple slides, was digitized to obtain whole slide images. Tissue regions were segmented in each image to guide the tessellation, resulting in one set of tiles for the entire case. All tiles were converted to feature vectors using the encoder of the Hierarchical Image Pyramid Transformer. An ensemble of five Vision Transformers was used to predict the diagnostic complexity of the case based on the set of extracted feature vectors for assignment to a general or expert pathologist.
