AutoML Differentiation of PA and Parasellar Meningiomas Using Pre-Operative MRI

CLINICAL QUESTION

Can an automated machine learning (AutoML) model accurately differentiate pituitary macroadenomas (PA) from parasellar meningiomas (PSM) using pre-operative MRI?

BOTTOM LINE

A customized AutoML model demonstrated excellent diagnostic performance in distinguishing PA from PSM on contrast-enhanced MRI, with balanced sensitivity and specificity exceeding 97% and strong external validation, supporting its potential as a scalable clinical decision support tool.

BACKGROUND: Accurate pre-operative differentiation between PA and PSM is critical for surgical planning, as management strategies and operative complexity differ substantially. MRI interpretation can be challenging due to overlapping radiographic features, and reported diagnostic accuracy varies by expertise and institutional experience. Traditional machine learning approaches have shown promise but often require extensive preprocessing and technical expertise, limiting clinical adoption. AutoML offers a streamlined alternative that automates model development.

STUDY DESIGN: Retrospective model development and validation study using automated machine learning. A single-label image classification model was trained using contrast-enhanced T1-weighted MRI slices. Performance was evaluated using the area under the precision–recall curve (AUPRC), F1 score, sensitivity, specificity, and predictive values at predefined confidence thresholds.

SETTING: Single academic institution with external validation using a publicly available, multi-institutional imaging dataset.

SYNOPSIS: The institutional dataset included 1,628 MRI images from 116 patients (71 PA and 45 PSM). Images were uploaded without preprocessing to a Google Cloud Vertex AI AutoML platform and randomly split into training (80%), validation (10%), and testing (10%) sets. At a standard confidence threshold of 0.5, the model achieved an aggregate AUPRC of 0.997, with balanced sensitivity, specificity, and F1 scores of 97.55%. Performance remained robust across lower and higher confidence thresholds, demonstrating predictable trade-offs between sensitivity and specificity. Class-specific performance was similarly strong, with high accuracy for both tumor types. External validation using an independent public dataset confirmed generalizability, yielding near-perfect precision–recall performance. A secondary sub-analysis evaluated parasellar meningioma subtypes originating from the planum sphenoidale and tuberculum sellae; the model maintained high classification accuracy for both anatomical locations. The authors emphasize that AutoML outperformed or matched traditional machine learning approaches reported in prior literature while eliminating the need for manual segmentation, feature engineering, or computational expertise. Finally, the limitations of this study include reliance on a single-institution primary dataset and potential image-level clustering effects due to multiple images per patient. Nonetheless, external validation mitigates concerns regarding overfitting. In conclusion, the authors suggest that AutoML represents a promising, accessible approach for skull base tumor classification that could assist pre-operative planning and broaden clinical adoption of artificial intelligence tools.

CITATION: Sina EM, et al. Automated machine learning differentiation of pituitary macroadenomas and parasellar meningiomas using preoperative magnetic resonance imaging. Otolaryngol Head Neck Surg. 2025;173:1376-1384. doi:10.1002/ ohn.70034