MAPGI: Accurate identification of anatomical landmarks and diseased tissue in gastrointestinal tract using deep learning

Abstract

Automatic detection of anatomical landmarks and diseases in medical images is a challenging task which could greatly aid medical diagnosis and reduce the cost and time of investigational procedures. Also, two particular challenges of digital image processing in medical applications are the sparsity of annotated medical images and the lack of uniformity across images and image classes. This paper presents methodologies for maximizing classification accuracy on a small medical image dataset, the Kvasir dataset, by performing robust image preprocessing and applying state-of-the-art deep learning. Images are classified as being or involving an anatomical landmark (pylorus, z-line, cecum), a diseased state (esophagitis, ulcerative colitis, polyps), or a medical procedure (dyed lifted polyps, dyed resection margins). A framework for modular and automatic preprocessing of gastrointestinal tract images (MAPGI) is proposed, which applies edge removal, contrast enhancement, filtering, color mapping and scaling to each image in the dataset. Gamma correction values are automatically calculated for individual images such that the mean pixel value for each image is normalized to 90 ± 1 in a 0–255 pixel value range. Three state-of-the-art neural networks architectures, Inception-ResNet-v2, Inception-v4, and NASNet, are trained on the Kvasir dataset, and their classification performance is juxtaposed on validation data. In each case, 85% of the images from the Kvasir dataset are used for training, while the other 15% are reserved for validation. The resulting accuracies achieved using Inception-v4, Inception-ResNet-v2, and NASNet were 0.9845, 0.9848, and 0.9735, respectively. In addition, Inception-v4 achieved an average of 0.938 precision, 0.939 recall, 0.991 specificity, 0.938 F1 score, and 0.929 Matthews correlation coefficient (MCC). Bootstrapping provided NASNet, the worst performing model, a lower bound of 0.9723 accuracy on the 95% confidence interval.