Deep Learning Based Classification Of Gastrointestinal Lesions Using Wireless Capsule Endoscopic Images

Abstract

Although wireless capsule endoscopy (WCE) has proven as an important minimally invasive diagnostic tool in the field of gastrointestinal (GI) diseases, it presents enormous difficulties when examining these images and the variety of lesions encountered.  The deep ensemble learning model for gastrointestinal lesion segmentation and detection are involved image preprocessing, lesion segmentation using deep learning, feature learning, and ensemble classification, which are designed to enhance the localization of lesions and diagnostic accuracy. For experimental evaluation we used the Kvasir-SEG dataset comprising of 1000 gastrointestinal polyp images, each accompanied by a segmentation mask. Three different types of CNN architectures (SegNet, U-Net and DeepLab) were trained to learn complementary spatial and contextual features independently. The segmentation results were then stacked to combine them together to improve the prediction accuracy with the help of a Random Forest Classifier (RFC). The individual models exhibited competitive segmentation performance with U-Net performing the highest validation accuracy (92.96%) and SegNet performing best with Dice coefficient (0.7210) and IoU (0.5654). The proposed ensemble framework showed a significant improvement in performance with 99.95% accuracy, 99.91% F1-score, 99.85% ROC-AUC, 99.99% precision, and 99.70% recall. Confusion Matrix analysis also substantiated good lesion identification with only a few false-positive and false-negative predictions. The results show the effectiveness of heterogeneous ensemble learning that merges the different strengths of various segmentation architectures and offers an effective method for automated analysis of gastrointestinal lesions. The proposed framework can be applied to decrease diagnostic burden and to aid computer-aided diagnosis systems for clinical applications of WCE.