Fig. 1: AI algorithm development and validation.

A A two-step approach was used to generate an ALK/ROS1 classifier; first, initial weights were generated from WSIs from Imagene-AI’s internal database, using an unsupervised learning algorithm. This was followed by a fine-tuning step performed using a semi-supervised algorithm, specifically for the tested biomarkers. B A retrospective set of de-identified scanned H&E images of 72 NSCLC samples were collected and analyzed by the ALK/ROS1 classifier. Results of the AI solution and the routine practice (IHC, NGS, and/or FISH) were compared for performance assessment. FFPE Formalin-Fixed Paraffin-Embedded, WSI whole image slide.