Detection and counting of glomeruli in kidney tissue images using faster R-CNN model with different backbone architectures

In this study, the performance of different backbone architectures was comparatively evaluated using the Faster R-CNN architecture for the automatic detection and counting of glomeruli in histological kidney tissue images. Eight backbones-AlexNet, ResNet-50, ResNet-101, GoogLeNet (Inception-v3), VGG-16, DenseNet-121, EfficientNet-B0, and MobileNet-V2-were evaluated under the same dataset and training protocol. A total of 2500 histological images with a resolution of 500×500 pixels were divided into three groups: 70% training, 15% validation, and 15% testing. All models were trained using a transfer learning approach with pre-trained weights on the ImageNet dataset. Performance evaluation was performed using the metrics of precision, recall, F1-score, Mean Average Precision at 0,50 Intersection over Union (IoU) threshold (mAP50), and Mean Average Precision in the 0,50-0,95 IoU range (mAP50-95). According to the test set results, MobileNet-V2 stood out with the highest F1-score value of 0,973 in terms of overall performance. ResNet-50 was the most successful in the precision metric with 0,954, while DenseNet-121 showed the highest value in the recall metric with 0,994. In evaluating object detection quality, VGG-16 stood out with 0,992 in the mAP50 metric, while VGG-16 showed the best performance with 0,708 in the mAP50-95 metric.