A Deep Learning Framework for Personalized Brain Metastasis Prognosis and Diagnosis

Abstract

Brain Metastasis present important diagnostic and therapeutic problems, occurring in from 10% to 30% of cancer patients, significantly affecting cognitive function. Manual interpretation of MRI is a routine but time-consuming task that may be imprecise in the case of tiny or diverse tumors. This paper presents a novel approach for automatic brain metastasis segmentation on MRI data using a U-Net model. This method fuses several imaging modalities to refine the identification of BM. These include T1-weighted, T2-weighted, T1 contrast enhanced, and FLAIR. The dataset used for this purpose is the University of California San Francisco Brain Metastases Stereotactic Radiosurgery (UCSF-BMSR) MRI dataset. The model was trained, tested, and verified using this dataset. The U-Net model with dropout performed quite well with an overall accuracy of 99.75%; Dice, 64.49%; and IOU, 96.81%. In this paper, the proposed method is compared to two baseline models: Convolutional Neural Networks and Fully Convolutional Networks. The U-Net outperformed the baselines on all the important measures and demonstrated great potential for being applied clinically in real life. This finding puts a finger on the capability of greatly improved detection accuracy with the U-Net model, and thus, timely treatment decisions.