Comparing UNet, UNet++, FPN, PAN and Deeplabv3+ for Gastrointestinal Tract Disease Detection

Gastrointestinal illnesses is a global health concern that necessitate the use of modern diagnostic technologies. Using the Kvasir-Seg dataset, this study compares the U-Net++, FPN, DeepLabV3+, PAN, and U-Net architectures for endoscopic image processing. Each model has advantages and disadvantages. U-Net++ excels in accuracy but consumes a lot of resources. FPN strikes a compromise between precision and speed, but it sometimes struggle with little objects. DeepLabV3+ produces reliable findings but requires a lot of computing resources. PAN captures fine details but at a possibly slower rate, whilst U-Net is simple but may be limited with complex images. The study uses metrics such as Dice Loss and IoU to demonstrate deep learning's ability in segmenting gastrointestinal illnesses. UNet++ stands out with an IoU of 0.8586, F1 Score of 0.848 and a dice score of 0.0832, highlighting the diagnostic potential of deep learning. Gastrointestinal illnesses are a global health concern that necessitate improved dialysis. This study offers important insights into model selection for medical image analysis, which assist both researchers and practitioners.