A clinically conditional deep learning model for predicting histological subtypes of pituitary neuroendocrine tumors from hematoxylin-eosin whole-slide images

Histological subtype classification of pituitary neuroendocrine tumors is important for prognostic assessment and therapeutic decision-making. However, current subtype classification relies heavily on extensive immunohistochemistry, which can be time-consuming, costly, and tissue-consuming. We aimed to develop a clinically conditional deep learning model that integrates routinely available clinical features with hematoxylin-eosin-stained whole-slide images to predict five subtypes of pituitary neuroendocrine tumors. To address the histomorphologic characteristics of pituitary neuroendocrine tumors, we designed an automatic pathology-informed patch selection strategy focusing on diagnostically relevant tumor regions, and applied a masking strategy to simulate missing clinical features. Model performance was evaluated using ensemble inference from five cross-validated models on an internal test set and a temporal and external validation set. Models were assessed by accuracy, sensitivity, specificity, predictive values, and area under the receiver operating characteristic curve, with 95% confidence intervals. Paired DeLong and McNemar tests were used to compare models, with P < 0.05 considered significant. Attention heatmaps were visualized for interpretability. On the internal hold-out test set, the clinically conditional model achieved an overall accuracy of 91.8% and an area under the receiver operating characteristic curve (AUC) of 0.934, outperforming the unimodal WSIs model (accuracy, 87.3%; AUC, 0.905; P < 0.05). On the temporal and external validation set, the clinically conditional model achieved an overall accuracy of 86.9% and an AUC of 0.901, also surpassing the unimodal model based only on WSIs (accuracy, 84.6%; AUC, 0.868; P < 0.05). Interpretability analyses showed that model attention was concentrated in morphologically meaningful regions and that the decision process was similar to recognized pathological features of individual PitNETs subtypes. These preliminary findings suggest that integrating clinical feature with histopathology may better simulate the comprehensive reasoning process of pathologists in routine diagnoses. This exploratory approach could potentially serve as a supplementary decision-support tool to facilitate ancillary test prioritization, optimize diagnostic workflow, and improve resource allocation, particularly in resource-constrained settings.