Deep Neural Networks for Disease Classification from Endoscopic Imaging

The GI tract is prone to lots of various normal problems that, if overlooked early, can show into very serious disease, including cancer. Unfortunately, with endoscopy, examination is very traditional and highly dependent on physician presence for a high chance of human error from differing morphology. Deep learning, leveraging mainly CNNs, has been recently shown to help improving accuracy of certain diseases diagnosis using automatically analyzed images from endoscopy. The proposed architecture model was successfully validated by the U-Net architecture model with respect to accuracy of 98.80% and final validation loss of 0.0309, since this helped in detecting the abnormalities in the given GI application. This value is higher than the benchmark which is the benchmarks that suggest that ResNet101 models had only achieved 98.37% accuracy on the KVASIR datasets. We quantify the nature of such sensitivity and precision, and demonstrate that our model is a good tool for computer assisted diagnosis in GI endoscopy by leveraging transfer learning and sophisticated CNN architectures. By supporting CNN-based models to become integrated into clinic, this work simultaneously improves patient outcomes and reduces diagnostic time in the management of GI tract disease. (Abstract)