Use of Artificial Intelligence for Barrier Integrity Detection in Porcine Small Intestinal Epithelial Cells

Abstract

Given the importance of monitoring intestinal permeability and the significant healthcare cost associated with the gut barrier disruption, automatic detection of barrier disruption pattern in porcine intestinal epithelial cells (IPEC-J2) using automatically computer-aided detection models based on a deep convolutional neural network for early detection and interpretation in measuring intestinal permeability is an area of active research and adequate experimental models are required to further understand the grade of localization and disruption of IPEC-J2 tight junction proteins. In the present study, a deep learning-based ensemble model to build a classifier to automatically analyze and extract features from input images in order to accurately assess the grade of localization and disruption of tight junction proteins (TJ) in IPEC-J2 have been proposed. Different data augmentation techniques including horizontal and vertical flips, rotating, zooming, contrast adjustment and brightness enhancement with different parameters are employed to increase the dataset size and tackle the over-fitting problem. At first, the experiments began with evaluating the performance of 8 state-of-the-art deep CNN architectures namely, VGG-Net, InceptionV3, MobileNet, DenseNet, Xception, NAS-Net, InceptionResNetV2 and ResNet models on IPEC-J2 cell image classification. Transfer learning is a common strategy in training deep CNN models. Using this strategy, the weights that are already learned on a cross-domain dataset to initialize weights of deep CNN models can be transferred in this research. The final results showed that the deep CNN ensemble of InceptionV3 and DenseNet201 achieved the best result with an accurate detection rate of 99.22% than the individual InceptionV3 architecture (95.03%) and the individual DenseNet201 architecture (91.11%). The second-best ensemble architecture was the ensemble of InceptionV3 and MobileNet with an accuracy of 97.78% than the individual InceptionV3 architecture (95.03%) and the individual MobileNet architecture (95.82%.) Collectively, employing CNN models could be considered as an automatic visual inspection system for the recognition, grading of expression, localization and disruption of tight junction proteins in epithelial cells with less misdiagnosis (false positive or false negative) and error rate, and also reduce the heavy workload of manual diagnosis.