“…33 Given a large enough repository of diagnostic WSIs (n > 100), 34 deep learning can be used to formulate and solve new clinical tasks beyond human pathologist capabilities such as metastatic origin of cancer prediction, 35 cancer prognostication, [36][37][38][39][40] and microsatellite instability (MSI) prediction. [41][42][43] Looking beyond supervised learning applications via CNNs and MIL, the development of other techniques such as generative AI modeling, [44][45][46] geometric deep learning, 47,48 unsupervised learning, 49,50 and multimodal deep learning 51 may soon enable new clinical capabilities that could enter pathology and laboratory medicine workflows, such as virtual staining, [52][53][54] elucidating cell and tissue interactions, 55 untargeted biomarker discovery, 56,57 data fusion with genomics and other rich biomedical data streams. [58][59][60][61] Though direct application of many deep learning techniques may appear to work "out-of-the-box" and have emergent capabilities in computational pathology (CPATH), there exists a variety of technical challenges that would limit their adoption in clinical translation, deployment, and commercialization.…”