Performance Elevation Using Augmented Pivot Point Rotation for Kidney Stone Detection

Kidney stone detection has been one of the key issues of healthcare professionals in the past. The recent development of deep learning-based models for kidney stone detection has reduced the time and workload of radiologists by assisting in the classification of kidney stone images. The contribution focuses firstly on generating the annotation of the publicly available dataset consisting of 1799 Non-Contrast Computerized Tomography (NCCT) coronal images collected from GitHub. Without manipulating the ratio of the training and testing samples, annotation of the images were carried out with bounding box instances of normal and kidney stone. A competent algorithm generates a new dataset using the augmented pivot point rotation (APPR) to the bounding box. The original and augmented datasets are trained on Single shot Detector (SSD), You Only Live Once (YOLOv7), and Faster Region-based Convolution Neural Network (RCNN) with backbones such as ResNet50, MobileNetv2, and ResNet101, and the results are compared. The result gives a trade-off between the single-stage and two-stage object detection models. The precision of YOLOv7 is 0.986 and 0.966 for normal and kidney stones, respectively, but the precision of the Faster RCNN for normal and kidney stones is balanced. Faster-RCNN training parameters are more compared to YOLOv7, resulting in an increase in training time. YOLOv7 surpasses the outcomes compared to other models with a mAP@0.5:0.95 of 0.933. An average mAP@0.5:0.95 scores for all the models trained on the augmented dataset is intensified by 18.9%. YOLOv7 and Faster-RCNN with ResNet50 provide promising results after training on augmented data. It is concluded that YOLOv7 and Faster RCNN with ResNet50 are suitable for localizing kidney stones with the proposed augmentation technique.