Improving the Estimation and Detection of Kidney Stone from Ultrasound Image using Despeckling and Deep learning algorithms

Abstract

Kidney stone illness has been a severe reoccurring issue observed more frequently all over the world. By the other name “Nephrolithiasis” a deposition of crystallized minerals, acids and proteins found in urine within the kidneys. These stones can be anywhere between <3mm up to >20 mm in length, causing hurt and injury to the body because of their size and shape. Diagnosis of kidney stones is done using a variety of imaging techniques, such as Computed Tomography (CT), X-ray and Ultrasound. Since, CT and X-ray methods are expensive, expose patients to radiation and due to limited access. Clinical centers in Ethiopia prefer ultrasound imaging approach for the patients. However, the inescapable roughness and speckles of ultrasonic images make it difficult, leading to erroneous diagnosis of smaller kidney stones and im pairing their treatment. In this study, a deep learning algorithm and despeckling techniques have been used to accurately estimate and detect kidney stones from ultrasound images. In total, a set of 135 normal and 495 kidney stone images have been acquired from Jimma University Medical Center and Worabe Comprehensive Specialized Hospital and used for training, validation and testing purposes. The acquired images were preprocessed and despeckled using image cropping and resizing, normalized, filtered using adaptive wiener filter, median filter and equalized using histogram equalization and the performance of each despeckling technique was evaluated. Also, augmentation techniques has been used for increasing the number of data sets after partitioning of dataset. Augmented train split data sets were trained on Yolo Nas, Yolov8, Yolov5 and Detectron2 Faster RCNN models. Each model trained first using pre-trained COCO data set rather than training from scratch. Then models were evaluated through valid and test datasets in the training and inference image respectively. By adjusting the hyper-parameters through trial and error of the neural network during training the execution of each model were examined and improved. Overall, the study anticipated a precise Nephrolithiasis diagnostic tool that provides reliable decision making support to the healthcare professionals ensuring patient safety. In this regard, our developed model using Yolo-Nas architecture has the highest accuracy and mAP scores of 94% and 86.7%, respec tively, in detecting the stones.