TorqueSAM: unsupervised kidney CT analysis with localization and SAM-integrated torque clustering segmentation

Abstract

Chronic kidney disease is a growing public health hazard that often leads to permanent renal system failure if not diagnosed and treated by experts within a treatable time interval. In addition to that, the shortage of expert nephrologists in unprivileged areas slows down the early detection and prevention. However, AI driven solutions can play a vital role in ensuring early detection and prevention despite the shortage of experts in many cases. But most existing AI driven automated renal disease detection systems heavily rely on manually annotated data, which still needs the experts’ intervention to detect any chronic renal illness. To address this issue, this study proposes and tests a fully unsupervised method to detect one critical chronic renal disease: Renal stone, often interpreted as a stone that does not rely on any annotated data. In this regard, our proposed method, namely TorqueSAM, introduces a different approach to work with modern detection and segmentation models, along with a classification module to detect any Renal Stone without relying on any ground truth labels. Our approach is tested on a dataset of 6,454 axial and coronal CT scan images. Our study focuses on one alarming illness that is highly dominant and associated with long-term kidney failure. TorqueSAM shows that the unsupervised image segmentation and classification surpass the classical supervised approaches significantly in terms of reducing the annotation overhead, time, and resources. For renal stone, TorqueSAM showed dice scores of 0.5767, IoU of 0.4183, precision of 0.9296, recall of 0.4655, accuracy of 0.9992, and specificity of 0.9998. However, the classification results of the clustering is found to be 0.8232 and 0.8472 for ARI and NMI, respectively. Despite having some limitations in the segmentation, this research shows the potential to outperform the classical supervised techniques, prioritizing the optimization of the region of interest (ROI) to enhance the predictive accuracy while reducing annotation overhead, which allows TorqueSAM to be a scalable solution to mitigate the sufferings of renal stone detection.