Automatic Segmentation of Polyps using U-Net from Colonoscopy images

Deep learning is a sort of neural network that has more processing layers than regular neural networks and can abstract and predict input at higher levels. It has swiftly established itself as a leading machine learning tool for image and computer vision applications. Deep convolutional neural networks are also quite good in automatically assessing photos, according to current research trends. Using a huge collection of colonoscopy and endoscopy images, this paper presents an effective U-Net architecture based on pretrained deep CNNs. We integrate U-Net architecture with data augmentation and patch extraction from colonoscopy images to detect polyps. The U-Net model is trained and validated to detect polyps in a set of colonoscopic and endoscopic images. The system can achieve overall high polyp detection precision and sensitivity using the suggested U-Net framework and our tuned hyperparameters. After the U-Net model is trained and validated, we evaluated mean Intersection Over Union (IoU) score as 0.9897 and dice loss as 0.0523. With the help of obtained evaluation metrics, the IoU score and dice loss curves are plotted. Furthermore, the proposed U-Net architecture is scalable and adaptable, allowing it to be readily upgraded in the future to include other forms of disease detection and to include more deep learning models to improve its generalising capabilities.