Automated Tumor Segmentation in Prostate Cancer MRI Using Hybrid SE-ResNet & Vision Transformer (ViT)

Prostate cancer is the most common type of cancer in men and a leading cause of cancer death in men. Precise localization of the tumor and detection during early stage are critical for effective and favorable treatment of the patient. To enhance the accuracy of the prostate cancer detection, in this work, we concentrate only on tumor segmentation on MRI using a combined deep learning model (SEViT). The dataset contains preprocessed MRI scans in DICOM format and these images have been augmented, pixel intensity normalized to provide model robustness. The suggested SE-ViT architecture enhances the model’s ability to collect significant tumor-related information, by adopting ResNet-34-based multi-scale SE (Squeeze-and-Excitation, SE) blocks for hierarchical feature extraction. Better understanding of tumor regions can be obtained by the Vision Transformer (ViT), which captures both global spatial dependencies. By incorporating transformer-based and convolutional representations, the model segments prostate tumors successfully as it is capable of differentiating affected regions from healthy tissue. Preprocessing of images, automatic formation of tumor mask, model training, validation and evaluation metrics such as Dice Score, IoU, Precision-Recall are all integrated in end-to-end system. The performance of SE-ViT as a state-of-the-art medical imaging tool is verified based on experimental studies in terms of the quality of tumour region segmentation.