Automatic intestinal content classification using transfer learning architectures

Abstract

Investigation of anomalies in capsule endoscopy (CE) is affected by an impairment of the mucosal frames with bubbles, debris, intestinal fluid, foreign objects, and chyme (food) etc., which can lead to a higher false-positive rate during manual and computer-aided analysis. An automatic intestinal content classification can help in checking the reliability and efficacy of computer-aided anomaly detection for CE frames. This paper presents three transfer learning (TL) architectures namely VGG16, InceptionResNetV2, and ResNet50V2 for automatic intestinal content classification using 1,67,486 and 140 CE patches and frames. A comparative analysis of the TL architectures has been done through various evaluation metrics like accuracy, precision, recall, specificity, loss, area-under-curve (AUC) and F1-score, test set evaluation, and feature maps. ResNet50V2 performed best among the three architectures and achieved an accuracy, precision, recall, specificity, and F1-score up-to 94.15%, 94.73%, 93.17%, 95.08%, and 93.95% respectively for CE frames. All three architectures efficiently classified ‘dirty’ test set CE frames and outperformed in comparison to the existing state-of-the-art works.