Comparative analysis of imaging diagnostic models for tubular basophilia and mineralization of kidney

Abstract

Abstract Background Now that it is possible to efficiently classify and save tissue images of laboratory animals using whole-slide imaging, many diagnostic models are being developed through transfer learning with Convolutional Neural Network (CNN). In this study, transfer learning was performed to gain toxicopathological knowledge using CNN models such as InceptionV3 and Xception. For the classification of tubular basophilia and mineralization, two representative background lesions that commonly occur in toxicological studies, accuracies of diagnosis were compared using MobileNetV2, Xception and InceptionV3. For the simultaneous detection of the two lesions, the accuracy was analysed using You Only Look Once version 4 (YOLOv4). Results The accuracy of the classification models was as follows: MobileNetV2 (epoch 50, accuracy: 98.57%) > Xception (epoch 70, accuracy: 97.47%) > InceptionV3 (epoch 70, accuracy: 89.62%). In the case of object detection, the accuracy of YOLOv4 was 98.62% at epoch 3000. Conclusions Among the classification models, MobileNetV2 had the best accuracy despite applying a lower epoch than InceptionV3 and Xception. The object detection model, YOLOv4, accurately and simultaneously diagnosed tubular basophilia and mineralization, with an accuracy of 98.62% at epoch 3000.

Keywords

• Artificial intelligence
• Diagnosis
• Classification models
• YOLOv4
• Tubular basophilia
• Mineralization