Efficiency of Transfer Learning for Abnormality Detection using Colonoscopy Images: A Critical Analysis

Abstract

Colonoscopy is a procedure which helps the gastroenterologists to examine the lower gastrointestinal (GI) region for any abnormal condition. Diagnosis of diseases in the lower GI area is challenging and complicated process which requires highly experienced professional for accurate examination. In recent times deep learning has gained popularity for faster and more accurate detection of diseases and can be an efficient assist to the gastroenterologists. In this work the effectiveness of pre-trained deep convolutional neural network (DCNN) models namely GoogleNet, ResNet-50 and ShuffleNet are analyzed using the concept of transfer learning for abnormality detection in lower GI. The DCNN architectures are trained and compared using the images related to lower GI region from HyperKvasir dataset. The work also demonstrates the effect of data augmentation in enhancing the performance of DCNN. The result shows that ResNet-50 architecture performs the best in classification of lower GI images as normal or abnormal with 94.08% of accuracy, MCC of 0.879, without data augmentation, as well as 95.93% of accuracy and MCC of 0.917 with data augmentation. This study can help researchers to further explore the strength of transfer learning and design their own deep networks.