TransDeeplab: A Novel Architecture for Medical Image Segmentation

Abstract

Early and accurate diagnosis of brain tumors plays a critical role in improving patient prognosis. Magnetic Resonance Imaging (MRI) serves as the core diagnostic modality; however, the inherent heterogeneity of brain tumors and the complexity of their morphological boundaries pose significant challenges for traditional segmentation methods, limiting their clinical applicability. To address this issue, we propose the TransDeepLab model, aiming to develop an efficient and precise brain tumor segmentation algorithm. The model integrates DeepLabV1 and Vision Transformer (ViT) within an Encoder–Decoder U-shaped architecture, leveraging the strengths of local feature extraction and global context modeling. Moreover, we introduce two novel components: a Feature Information Exchange Module to enhance feature fusion and a Trainable Gated Selection Module to optimize feature utilization. Using the BraTS2021 dataset and implemented within the PyTorch framework, the proposed approach is rigorously evaluated against state-of-the-art segmentation methods in terms of accuracy, robustness, and computational efficiency. The study aims to provide an automated, high-precision brain tumor segmentation tool to advance intelligent healthcare in the field of neuro-oncology.