Multimodal MRI Brain Tumor Segmentation Using a 3D LinkNet–ResNet152 Architecture

Abstract

Accurate segmentation of brain tumor regions from multimodal magnetic resonance imaging (MRI) is an important step in computer-assisted brain tumor analysis. However, the heterogeneous appearance and complex spatial structure of tumor regions make automated three-dimensional segmentation challenging. This study presents a multimodal three-dimensional LinkNet architecture with a ResNet152 encoder for semantic segmentation of brain tumor MRI volumes. The proposed framework utilizes multimodal MRI data from the BraTS 2020 dataset, comprising FLAIR, post-contrast T1-weighted, and T2-weighted modalities. The input volumes are intensity-normalized, spatially cropped, and combined to form a three-channel representation with a standardized volume size of 128 × 128 × 128 voxels. A LinkNet encoder-decoder architecture is employed, with ResNet152 serving as the feature extraction backbone and ImageNet-pretrained encoder weights used for initialization. The network performs four-class voxel-level segmentation corresponding to background, necrotic/non-enhancing tumor, peritumoral edema, and enhancing tumor. The trained model is evaluated using accuracy, precision, recall, F1-score, confusion matrix, and receiver operating characteristic analysis. The implementation also identifies tumor-containing slices and extracts region-based quantitative characteristics from the segmentation output. In addition, a Django-based interface is developed to facilitate MRI input, segmentation visualization, slice selection, and generation of a PDF-based analysis report. The results demonstrate the applicability of the proposed LinkNet–ResNet152 framework for automated multimodal brain tumor segmentation and subsequent tumor-region analysis, providing an integrated workflow from MRI preprocessing to segmentation visualization and reporting.