AI-based selection of tumor regions for genomic profiling in neuropathology

Automating pathology workflows with deep learning is increasingly feasible and clinically relevant. We present an AI-based method that identifies diagnostically relevant areas directly from H&E-stained slides, trained on 250 glioma cases using sparse, incomplete annotations. First, we show that attention-based multiple instance learning achieves accurate predictions despite noisy labels, easing the annotation burden. Second, the model highlights tumor regions with high cellularity or grade, offering reproducible guidance for tissue selection. In a prospective evaluation, AI-selected regions achieved a mean Dice score of 0.743 [±0.077], supporting integration into neuropathology workflows as reliable guidance for molecular diagnostics.