Automated Machine Learning Tool for Differentiating Pituitary Adenomas from Parasellar Meningiomas Using Magnetic Resonance Imaging

Introduction: Automated machine learning (AutoML) is a user-friendly artificial intelligence tool that simplifies image recognition model development. It shows promise in preoperative tumor evaluation, where overlapping radiographic features hinder diagnoses. This study evaluates AutoML’s ability to differentiate pituitary adenomas (PA), parasellar meningiomas (PSM), non-sellar meningiomas (NSM), and gliomas using preoperative MRIs.

Methods: A publicly available, multi-validated dataset of T1-weighted contrast-enhanced MRIs from 233 patients was used, comprising 3064 individual 6-mm slices (708 meningiomas, 930 PAs, 1426 gliomas). A skull base surgeon identified 29 PSM slices from the meningioma cohort. De-identified images were labeled and uploaded to Google Cloud VertexAI. A single-label image classification model was trained, validated, and tested (80%-10%-10% dataset split). Diagnostic metrics included area under the precision-recall curve (AUPRC), precision, and recall.

Results: Our AutoML model demonstrates high diagnostic accuracy in distinguishing PA, PSM, NSM, and gliomas. The aggregate AUPRC was 0.999, with 99.016% precision and 98.639% recall at a 0.5 confidence threshold. For PAs, AUPRC was 1.000, precision 100%, and recall 98.9%. For PSMs, AUPRC was 1.000, precision 75%, and recall 100%. NSM (AUPRC = 0.998, precision = 98.5%, recall = 98.5%) and gliomas (AUPRC = 0.998, precision = 99.3%, recall = 98.6%) performed similarly.

Conclusions: Our AutoML model accurately differentiates parasellar masses on preoperative MRIs, outperforming traditional ML methods. It is the first model in the literature to be specifically trained on a PSM dataset. Distinct from other models, its user-friendly design can facilitate seamless integration into community clinic settings. Our model is a scalable and accessible tool for streamlining neuroimaging diagnostics into an otolaryngology practice.