Fig. 3: False negative predictions with low-grade morphologies from the SUH cohort.

At the base model threshold of 0.5, i.e., before sensitivity prioritization, the AI produced six false negative predictions exhibiting low-grade morphologies from the SUH cohort–the only cohort with digitized IHC stains available. These cases were reassessed by the study pathologist (A.B.), blinded to the AI result, and subsequently reviewed in a meeting with a second pathologist (L.E.). Case 1 (a, g) shows a sub-millimeter area of prostatic glands with enlarged nuclei and prominent nucleoli, suspicious for ISUP 1 cancer. However, features like a fuzzy luminal border, wavy contours, and cytoplasmic pigment make a definitive diagnosis difficult. IHC confirmed a small ISUP 1 cancer; otherwise, the case would have been diagnosed as atypia/SFC. Case 2 (b, h) shows small, angulated glands and stromal elastoid degeneration, suggesting postatrophic hyperplasia. Some nuclear irregularities and hyperchromasia are present, but not convincing of malignancy. IHC reveals partially retained basal cells. The pathologists would render a benign diagnosis. Case 3 (c, i) shows 3-4 glands (left) with nuclear enlargement and prominent nucleoli, highly suspicious for malignancy. IHC confirmed ISUP 1 cancer, though the quantity is borderline insufficient for a definitive diagnosis. Case 4 (d, j) shows approximately 10 glands with low-grade atypia, not convincing for malignancy. On IHC, basal cells are lost in some glands; however, other glands with similar H&E morphology retain them. The pathologists would diagnose this as atypia/SFC, even after IHC. Case 5 (e, k) shows glands with minimal nuclear atypia deemed insufficient for a definitive cancer diagnosis, regardless of the IHC result. The pathologists would diagnose this as atypia/SFC. Case 6 (f, l) shows two glands with obvious nuclear atypia and pathological secretion, but quantitatively, the suspicious glands are too few to render a definitive malignant diagnosis, even after IHC. The pathologists would diagnose atypia/SFC, noting a strong suspicion of ISUP 1 cancer. AI-generated attention maps (m–r) show that, although slides were ultimately predicted to be benign, the AI correctly identifies suspicious areas, which could help focus the pathologist’s attention in a clinical setting. AI artificial intelligence, H&E hematoxylin & eosin, HMWCK high-molecular-weight cytokeratin, IHC immunohistochemistry, ISUP International Society of Urological Pathology grade, SFC suspicious for cancer, SUH Stavanger University Hospital.