Automated Meningioma Detection and Segmentation Using Deep Neural Networks

Introduction: Though tumor volume and growth rate assessment are a central part of surgical planning and surveillance in meningioma patients, these can be time consuming and potentially inaccurate task when done using the presently available manual or semiautomatic tumor segmentation methods. Machine learning approaches have the potential to allow for automated meningioma detection and segmentation on magnetic resonance imaging (MRI).

Methods: Using a dataset of 812 brain MRIs of meningioma patients, with a total of 940 tumors represented, we designed a deep learning (DL) algorithm based on a three-dimensional convolutional neural network (3D CNN) architecture with the goal of automatically detecting and accurately segmenting meningiomas. Accuracy compared with expert labels of the same MRIs was assessed using a Dice’s score. The algorithm’s potential impact on clinical workflow was assessed in a simulated clinical scenario, measuring time needed for accurate segmentation and volume estimation accuracy.

Results: The automated segmentation performed with a Dice’s score reaching a median of 91% and a mean of 85% for single and multiple tumors with volume >3cc (Fig. 1). Automated segmentation reduced time to create clinically usable segmentations by experts by an average of 65%. The predicted volumes generated by the algorithm showed a stronger correlation (0.98, p < 0.001) to real tumor volume values compared with traditional two- and three-dimensional estimation techniques (Fig. 2).

Conclusion: We developed a high-performing DL algorithm for meningioma segmentation with the potential of significantly impacting current clinical workflow. Further prospective investigation can assess both the accuracy of this tool and its utility in meningioma surveillance over time.

Fig. 1 Meningioma segmentation by expert and algorithm in a patient with two lesions.

Fig. 2 Performance of deep learning algorithm at different tumor volumes (top) compared with traditional two- and three-dimensional volumetric estimation methods (bottom).

No conflict of interest has been declared by the author(s).