Enhanced Brain Tumor Segmentation Using Transfer Learning- Based Residual U-Net Architecture

Abstract

Beyond assisting medical practitioners in tumor detection and quantification, advanced imaging techniques contribute to the development of more effective treatment and rehabilitation protocols. Contemporary MRI- based brain tumor segmentation methodologies have emphasized U-Net architectural designs to integrate high-level semantic features with low-level spatial information for enhanced segmentation precision. Traditional fully convolutional networks employed for this application demonstrate limitations in successful image reconstruction via the decoder pathway due to insufficient low-level feature propagationfrom the encoder components. Enhanced optimization of low-level feature transmission from encoding to decoding pathways is essential for improved image reconstruction capabilities. This research proposes a transfer learning-enhanced residual U-Net framework that integrates U-Net and VGG-16 architectures. VGG-16 integration within the encoder pathway enhances image reconstruction performance. Furthermore, residual pathways within skip connections are incorporated to emphasize critical feature characteristics while suppressing noisy and irrelevant feature responses. The model undergoes training using The Cancer Imaging Archive (TCIA) and BraTS 2018 datasets, demonstrating improved performance in segmenting small-scale brain tumors. The proposed methodology exhibits competitive performance compared to existing brain tumor segmentation approaches.