Deep Learning Model Predicts Pituitary Tumor Subtypes from Routine Slides

Researchers have developed a deep learning model that combines routinely available clinical features with hematoxylin-eosin-stained whole-slide images to predict five histological subtypes of pituitary neuroendocrine tumors, according to a study published in BMC Medical Informatics and Decision Making. On an 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 of 0.934, outperforming a unimodal model based only on whole-slide images, which reached 87.3% accuracy and an AUC of 0.905, a difference the authors report as statistically significant.

Histological subtype classification of pituitary neuroendocrine tumors matters because it informs prognostic assessment and therapeutic decision-making. Currently, however, subtype classification relies heavily on extensive immunohistochemistry, which the authors describe as time-consuming, costly, and tissue-consuming. The new model was designed to address this bottleneck by integrating information that pathologists already have on hand: the standard stained tissue sections and basic clinical variables.

To tailor the approach to the histomorphologic characteristics of these tumors, the team designed an automatic pathology-informed patch selection strategy that focuses on diagnostically relevant tumor regions within the slides. They also applied a masking strategy to simulate missing clinical features, a practical consideration for real-world deployment where not every patient record is complete. Model performance was evaluated using ensemble inference from five cross-validated models, and the models were assessed by accuracy, sensitivity, specificity, predictive values, and area under the receiver operating characteristic curve, each reported with 95% confidence intervals. Paired DeLong and McNemar tests were used to compare models, with a P value below 0.05 considered significant.

Importantly, the reported advantage held up beyond the internal test set. On a temporal and external validation set, the clinically conditional model achieved an overall accuracy of 86.9% and an AUC of 0.901, again surpassing the unimodal whole-slide-image model, which achieved 84.6% accuracy and an AUC of 0.868, with the comparison again reported as statistically significant. Temporal and external validation of this kind offers a stronger check on whether a model’s performance generalizes beyond the specific cases it was developed on.

The authors also examined what the model was paying attention to. Interpretability analyses showed that model attention was concentrated in morphologically meaningful regions, and that the decision process appeared similar to recognized pathological features of the individual pituitary neuroendocrine tumor subtypes. Attention heatmaps were visualized as part of this interpretability work, giving a window into whether the network was relying on biologically plausible cues rather than artifacts.

The authors frame these results as preliminary findings suggesting that integrating clinical features with histopathology may better simulate the comprehensive reasoning process of pathologists in routine diagnoses. They describe the exploratory approach as potentially serving as a supplementary decision-support tool to facilitate ancillary test prioritization, optimize diagnostic workflow, and improve resource allocation, particularly in resource-constrained settings. In other words, the model is positioned not as a replacement for immunohistochemistry or expert pathologists, but as a way to help decide which additional tests are worth prioritizing for a given case.

Several caveats are worth keeping in view. The article is being shared early as a peer-reviewed, accepted manuscript that is citable and carries a permanent DOI, but this version remains subject to further edits and will be replaced by the final version of record. The authors themselves label the conclusions preliminary and exploratory, and the validation, while including temporal and external data, reflects the performance reported in this single study. As with any diagnostic support tool, real-world value would depend on prospective evaluation in clinical workflows, which lies beyond what this study reports.

Within those limits, the study adds to a growing body of work applying deep learning to pituitary neuroendocrine tumor pathology, a field where related efforts have explored predicting tumor lineage and granulation patterns from stained histology. By showing that a modest amount of clinical information, combined with carefully selected image regions, can push subtype prediction accuracy above what images alone provide, the work suggests a practical direction for computational pathology in endocrine tumor diagnosis: models that mirror how pathologists actually reason, weighing both what the tissue looks like and what is already known about the patient.

Subject of Research: A deep learning model that predicts histological subtypes of pituitary neuroendocrine tumors from hematoxylin-eosin whole-slide images and clinical features

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

Article References: Zheng, A., Lin, R., Chen, C., Hua, X., Xiao, H., & Zhao, L. (2026). A clinically conditional deep learning model for predicting histological subtypes of pituitary neuroendocrine tumors from hematoxylin-eosin whole-slide images. BMC Medical Informatics and Decision Making. https://doi.org/10.1186/s12911-026-03918-y

Image Credits: AI Generated

DOI: 10.1186/s12911-026-03918-y

Keywords: deep learning, pituitary neuroendocrine tumors, whole-slide images, histological subtyping, computational pathology, immunohistochemistry, decision support, external validation, attention heatmaps, diagnostic workflow, clinically, conditional