Skip to main content
2024 AANS Annual Scientific Meeting Main banner
Back

Stereotactic/ Functional Rapid Fire Abstract

Stereotactic and Functional

Automated Pituitary Adenoma Segmentation for Radiosurgery with Deep Learning-Based Model

Saturday, May 4, 2024
3:43 PM – 3:45 PM CT
Location:  S405b
Disclosure(s):
Neevya Balasubramaniam, MD (c): No financial relationships to disclose
Introduction: Pituitary adenomas are prevalent sellar region lesions primarily treated with transsphenoidal endoscopic surgery. Stereotactic radiosurgery (SRS) is increasingly used for recurrent or complex cases. Precise delineation of target structures and organs at risk is required for SRS. Our work aims to demonstrate the ability of artificial intelligence to provide accurate segmentation and enhance the efficiency of the treatment planning workflow.

Methods: A retrospective study was conducted on patients undergoing SRS for pituitary adenomas between January 2010 and October 2022 at Na Homolce Hospital. The collected data were split to a training and validation dataset in 80:20 ratio. A model based on nnU-net, a golden standard medical image segmentation tool, was trained to predict accurate segmentation for tumor, optic nerve and normal pituitary gland. Accuracy of predicted segmentations was assessed by Dice similarity coefficient (DSC) quantitatively and qualitatively by human expert rating. The effect of tumor volume and hormonal activity status on the accuracy was assessed.

Results: The training and validation sets included 582 and 146 patients, respectively, with similar patient characteristics. The model achieved DSC between original and predicted segmentations of 83.1% (tumor), 62.9% (normal gland), and 78.0% (optic nerve). Expert evaluations found 41% of segmentations directly applicable for treatment planning, 31.5% needed minor adjustments, and 27.4% were unsuitable for clinical use. Larger tumor volume and non-functioning adenomas were associated with higher accuracy (p < 0.001, p < 0.001). Adding T2 weighted scans as additional inputs resulted in higher DSC for optic nerve and normal gland (p < 0.001, p = 0.040).

Conclusion : This research demonstrates the promising potential of deep learning in automating pituitary adenoma segmentation from MRI data. Notably, we observed strong performance in a subgroup of large, hormonally inactive macroadenomas, warranting further research in this venue. This study encourages clinicians to leverage their expertise in conjunction with machine learning, ultimately enhancing patient care in neurosurgical practice.