Artificial intelligence using a compact residual neural network accurately detected pituitary tumors and distinguished them from meningiomas, gliomas, and normal brain magnetic resonance imaging findings, achieving 98% accuracy during validation and 94% accuracy on an independent test dataset, according to a research abstract presented at ENDO 2026, the Endocrine Society annual meeting, in Chicago. 

Pituitary tumors are a common cause of endocrine dysfunction and often require imaging to guide medical and surgical management. According to the investigators, differentiating pituitary tumors from meningiomas, aggressive gliomas, and normal anatomy can be difficult, particularly in resource-limited settings.  

Researchers evaluated whether ResNet18, a compact artificial intelligence (AI) architecture, could provide accurate tumor classification while maintaining computational efficiency suitable for cross-platform clinical deployment. 

The team curated brain magnetic resonance imaging (MRI) scans from multiple institutions that included pituitary tumors, meningiomas, gliomas, and non-tumor controls. Each image underwent independent review by at least two expert neuroradiologists to establish ground-truth labels. The dataset was divided into training, validation, and independent test cohorts using a 3:1:1 ratio.  

The researchers applied transfer learning with pretrained initialization followed by full-network fine-tuning. Model performance was evaluated using accuracy, precision, recall, F1 score, class-specific area under the receiver operating characteristic curve (AUROC), and loss behavior. External validation and deployment testing were also performed using a cross-platform research application. 

During validation, ResNet18 achieved an overall accuracy of 98%. Precision, recall, and F1 score were all approximately 0.98, demonstrating balanced performance across all three measures. The investigators also reported consistently high class-specific AUROC values for pituitary tumors, gliomas, meningiomas, and non-tumor controls, although individual AUROC values for each class were not provided in the abstract. 

Performance remained strong during independent testing. Overall accuracy was 94%, and the F1 score was 0.94, confirming the model's reproducible performance outside the training environment. The researchers also reported efficient inference that enabled rapid image analysis for cross-platform deployment.  

The model's efficiency was attributed to the ResNet18 architecture, which uses identity-based shortcut connections to stabilize gradient propagation while maintaining shallow network depth. This design supports efficient feature learning with reduced parameter burden and rapid convergence, according to the investigators.

"ResNet18 demonstrates that compact residual architectures can deliver high diagnostic accuracy for pituitary tumor detection and differential diagnosis from brain MRI without excessive computational demand," wrote the researchers. Its "efficiency, robustness, and external validation support its potential role as a scalable decision-support tool within neuroendocrine imaging workflows." 

Researchers Elangovan Krishnan, PhD, Aboorva K. Sudhakar, MD, and Gowrishankar Palaniswamy, MD, reported no financial conflicts of interest. No additional disclosures were reported.