Deep Learning-Based Segmentation of Various Brain Lesions for Radiosurgery

Semantic segmentation of medical images with deep learning models is rapidly being developed. In this study, we benchmarked state-of-the-art algorithms on our clinical stereotactic radiosurgery dataset. The dataset consists 1688 patients various brain lesions (pituitary tumors, meningioma, schwannoma, metastases, arteriovenous malformation, and trigeminal neuralgia), divided the into a training set (1557 patients) test (131 patients). This study demonstrates strengths weaknesses deep-learning in fairly practical scenario. We compared model performances concerning their sampling method, architecture, choice loss functions, identifying suitable settings for applications shedding light possible improvements. Evidence from led us to conclude that could be promising assisting even if was high heterogeneity lesion types sizes.