Deep Learning and Nonlinear Optimization-Assisted PDE Segmentation Framework for Accurate Brain Tumor Boundary Detection in MRI Scans

Abstract

The correct outline of the brain tumour in magnetic resonance imaging (MRI) is an essential process in the neurosurgical planning process, radiation therapy, and longitudinal monitoring. The framework suggested in this paper combines deep learning with nonlinear partial differential equation (PDE)–based optimization in high-precision tumour boundary segmentation of multi-modal MRI scans. A convolutional neural network (CNN) is initially trained to give an initial approximate tumour and neighboring edema tissues segmentation mask. Second, a level-set module is a PDE-based level-set module, which is driven by nonlinear energy minimization, to refine the boundary, hence imposing smoothness and retaining fine structural information. The optimization parameters are adjusted in an adaptive manner by a deep reinforcement learning optimizer to learn control policies of the PDE energy weights to allow capturing patient-specific variability. Extensive trial on publicly accessible multi-modal MRI brain tumor datasets reveal that the framework proposed achieves significant enhancement in Dice similarity coefficient (DSC) increase of 4.7percent as well as decreased 95th-percentile Hausdorff distance (HD95) relative to state-of-the-art deep-learning segmentation on its own. Corrupted input and missing modality robustness tests demonstrate little performance drop (DSC decline less than 1.8 percent). This shows that the combination of deep learning and nonlinear optimization can be used to successfully present accurate and clinically reliable brain tumour boundary detection.