MRI-based deep learning approaches to classify prolactinomas and non-functioning pituitary adenomas

Details

Supervisors

Macq, Benoît

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil biomédical, à finalité spécialisée

Abstract

enContext : Prolactinomas and non-functioning pituitary adenomas are the two most common subtypes of pituitary adenomas encountered in clinical practice. However, their distinction can be difficult, particularly in the case of hyperprolactinemic macroadenomas. Magnetic resonance imaging (MRI) is the gold standard for evaluating sella turcica lesions but conventional morphological criteria alone are insufficient to reliably differentiate between the two entities. Objective : This study aims to determine whether T2-weighted coronal MRI images alone contain information that can distinguish prolactinomas from non-functioning pituitary adenomas labeled according to their endocrine status. Methods : A cohort of 119 patients from multiple hospitals was used. Binary labels were assigned based on serum prolactin levels. Following spatial pre-processing involving resampling and cropping of the sella turcica region, eleven deep learning model configurations (CNN architectures, BiomedCLIP-based models and DINOv3-based models) were evaluated on T2-weighted coronal slices. Slice-level predictions were aggregated at the patient level using three strategies (Majority voting, Gaussian weighting and middle slice prediction). Model selection and hyperparameter tuning were performed on a validation set. The final evaluation was conducted on a fixed test set of 21 patients who were never used during training. Results : All evaluated configurations achieved balanced accuracy (BA) at the patient level that exceeded chance. The highest performance was obtained with BiomedCLIP with LayerNorm fine-tuning (BA 0.71) with Gaussian weighting aggregation of the slice-level predictions on the test set. Average recalls of 0.57 and 0.86 were obtained, for this configuration, for non-functioning pituitary adenomas and prolactinomas, respectively. Foundation models consistently resulted in higher performance than custom CNNs whose higher balanced accuracy reached 0.58. Conclusion : This work demonstrates that T2-weighted coronal MRI provides a distinctive signal related to the functional identity of pituitary adenomas. However, the moderate performance observed is insufficient for direct clinical application. This work constitutes an exploratory proof of concept. It lays the groundwork for future research involving larger cohorts, additional MRI sequences and the pre-training of pituitary-specific models.