Deep learning-based classification of gh-secreting pituitary neuroendocrine tumors granulation patterns using...

Accurate classification of granulation patterns in growth hormone-secreting pituitary neuroendocrine tumors (GH-PitNETs) is clinically essential but traditionally dependent on labor-intensive immunohistochemistry (IHC). To address this, we developed the Cross-Attention Pituitary Neuroendocrine Tumor Network (CA-PITNET), a deep learning model for automated subtyping directly from routine H&E-stained whole-slide images (WSIs). Utilizing 95 GH-PitNETs WSIs (54,610 patches) and 36 nonfunctioning PitNET WSIs (39,590 patches), the model was evaluated on two tasks: distinguishing GH-PitNETs from nonfunctioning PitNETs (Task I) and differentiating densely from sparsely granulated GH-PitNETs (Task II). Following staining normalization and rigorous validation, CA-PITNET demonstrated exceptional diagnostic accuracy. For Task I, CA-PITNET achieved patient-level AUCs of 0.948 (95% CI: 0.854-1.000; internal validation, n = 48) and 0.948 (95% CI: 0.837-1.000; external test, n = 33), with accuracy of 0.938 and 0.909, respectively. For Task II, patient-level AUCs were 0.935 (95% CI: 0.890-1.000; internal validation, n = 38) and 0.875 (95% CI: 0.670-1.000; external test, n = 25), with accuracy of 0.895 and 0.840, respectively. Generated heatmaps precisely localized diagnostic regions, highlighting diagnostically relevant regions consistent with pathologist-reviewed areas. This model is designed to augment the pathological workflow by improving the subtyping efficiency of GH-PitNETs, serving as a valuable adjunct to the standard IHC-based classification.