Automated gastrointestinal abnormalities detection from endoscopic images

Abstract

Impressive high performance reported in the use of machine learning on computer vision problems is often due to the availability of very large datasets with which deep neural networks can be trained. With inference from medical images, however, this is not the case and available data is often only a small fraction in size in comparison to benchmark natural scene recognition problems. To circumvent this problem, transfer learning is often applied, where a model trained on a large natural image corpus is adapted, or pre-trained, to model the medical problem. In this work, we consider transfer learning applied to a specific medical diagnostics problem, that of abnormality detection in the gastrointestinal tract of a human body using images obtained during endoscopy. We carry out a search over several image recognition architectures and adapt pretrained models to the endoscopy problem. Using the benchmark KVASIR dataset, we show that transfer learning is effective in outperforming previously reported results, at an accuracy of 98.5±0.27.