Benchmarking Semantic Segmentation Approaches for Polyp and Lesion Detection in Medical Imaging

In medical imaging, semantic segmentation is essential because it can precisely locate and isolate regions of interest, such lesions or tumours, from intricate anatomical systems. Deep learning (DL) has led to a substantial evolution in segmentation approaches, which have improved medical diagnostics' accuracy and efficiency. The performance of six cutting-edge models is assessed in this paper: LinkNet, U-Net with ResNet50, LinkNet with ResNet50, FPN Net, Simple U-Net, and SegNet-Transformer across four segmentation tasks—polyp detection, lung lesion segmentation from Computed Tomography (CT) scans, breast lesion detection from ultrasound images, and Segmenting brain tumours with magnetic resonance imaging (MRI). With results like 0.9618 for brain tumour segmentation and 0.9483 for lung lesion segmentation, SegNet-Transformer consistently obtained the highest Dice Coefficient across all tests, demonstrating superior ability in capturing intricate boundaries. Other models, such as U-Net with ResNet50, performed competitively with Dice scores of 0.9375 for lung lesion segmentation and 0.9472 for brain tumors but required slightly more memory and longer training times. This study underscores that model performance varies across different tasks, revealing specific strengths and weaknesses. Ultimately, while SegNet-Transformer demonstrates robustness, the choice of model should be informed by the unique requirements of each segmentation task, balancing accuracy, efficiency, and computational demands.