Artificial intelligence and machine learning have seen considerable progress in detecting and managing kidney stones in recent years. Deep learning has been proven invaluable in medical applications, as it can detect and localize kidney stones in CT images automatically. The resulting technique’s high accuracy and efficiency improve diagnostics and reduce the possibility of human error. Further, transfer learning has also emerged as a viable method to reuse a trained model for a new task. Through transfer learning, deep learning model performance is optimized in kidney stone segmentation by leveraging the knowledge of similar tasks, thereby providing a time-efficient and resource-saving approach toward model development. Implementing this method can address issues of adapting 3D convolution networks to clinical data collected from various modalities to which it has not previously been exposed5. Being the beneficial addition of AI and ML techniques in detecting and managing kidney stones, they are influential in improving diagnostic accuracy, lessening human errors, and giving personalized therapeutic approaches. Furthermore, these techniques would also help predict the therapy’s success rate and outcome and reduce the dose as low-dose CT techniques become available. Nevertheless, these limiting factors are still considered, including the lack of consideration for stone composition and the impact of AI algorithms on predicting the predicted rate of success and treatment outcomes.
Segmentation methods
Accurately detecting and classifying kidney stones and other renal conditions from medical images requires segmentation. In recent research, deep learning-based segmentation techniques, along with other methodologies, have shown an increase in the performance of the segmentation. The orientation of this section was the most relevant study to our work, and it provided a solid reason for why they were included. One such study by Leube et al.6 demonstrates the usefulness of combining positron emission tomography (PET) with CT in kidney segmentation is one. Our hybrid approach improved segmentation accuracy, especially for kidneys with cysts or near adjacent organs. Adding PET data to the set of modalities enabled better delineation, demonstrating the utility of multi-modality imaging in renal diagnostics. In another study relevant to this, Junyu et al.7 used a semi-supervised deep learning method based upon cycle generative adversarial network (CycleGAN) for segmenting kidneys in magnetic resonance imaging (MRI). An approach yielded excellent results with a mean Dice score of 0.92 and a mean Jaccard score of 0.85 over 80 datasets. The techniques derived from MRI were shown to apply to CT images, demonstrating the flexibility of deep learning in segmentation tasks. In Gaikar et al.8, kidney segmentation was investigated using different MRI sequences with fivefold cross-validation on eight deep-learning model datasets. Their findings on T1-weighted non-Gd images emphasized that U-Net provided the most accurate segmentation with a dice similarity coefficient (DSC) of 89.34. A critical point was made by the study, which is that good segmentation accuracy relies substantially on model selection and tuning.
Felix et al.9 also exploited pixel-based identification techniques to detect kidney stone boundaries. Using their advanced segmentation method, they achieved an accuracy of 92.5%, showing that identifying precise boundaries between the stone and the kidney improves the clinical outcomes of patients with kidney stones. In a related study, Angshuman et al.10 found kidney stones in hospital and clinical ultrasound images. Their feature extraction method, a 3D U-Net model, obtained 96.82% accuracy. The research suggests that ultrasound imaging and advanced segmentation can improve kidney stone detection. The work of Li et al.11 extends the feasibility of deep semantic segmentation models in kidney segmentation and stone detection. These findings support the idea that deep learning can improve kidney-related diagnostics accuracy. Amiri et al.12 show how several image features, including renal radiation dose, irradiated renal volume, and 24-h urine volume, predict chronic kidney disease (CKD). They surmounted the obstacles of sparse data and generated 140 radiomic features, achieving 94% accuracy with Random Forest. Ma et al.13 proposed a heterogeneous modified artificial neural network (HMANN) to perform early CKD detection and kidney segmentation within the internet of medical things context. On ultrasound images, HMANN consists of a support vector machine (SVM) and multilayer perceptron (MLP) and has extremely high efficiency with speedy computation, high sensitivity, specificity, and an area under curve score of 92.2 for kidney segmentation. These studies point to the potential of advanced techniques for predicting and diagnosing CKD. While advances in segmentation processes have been made, thorny issues remain, namely the ability to detect stones accurately and perform segmentation efficiently. Existing methods lead to long execution times and prominent memory usage and need optimization for practical clinical applications. This can be confirmed by the previous studies indicating the need for optimization. Addressing these limitations is crucial for improving the reliability and efficiency of kidney disease diagnostics.
Classification methods
With machine learning and deep learning techniques, kidney diseases have been explicitly classified for kidney stone patients. The discussion of the most relevant studies that advance understanding of the most practical classification methods is focused on how these studies are relevant toward better diagnostic accuracy and potentially better patient outcomes. Pande and Agarwal14 are notable ones who have applied the concept to classify different renal diseases using a multi-classification methodology. The researchers identified kidney cysts, stones, and tumors with an accuracy rate of 82.52% using a pre-trained YOLOv8n-cls model. By using advanced deep learning models, the work shows that precise disease classification is possible, which is a starting ground for future research in that area.
Saif et al.15 combined CNNs and LSTM networks to predict CKD occurrence in a proposed ensemble model. Their model, however, showed marked improvements in classification performance and suggested that the combination of different architectures could improve the diagnostic capabilities of each model. While this is promising for population-based studies, it may lack the ability to give personalized patient care. Subedi et al.16 focused on differentiating between four significant categories in CT scan images: Cysts, normal states, tumors, and stones. They enhanced the capabilities of a modified Vision Transformer (ViT) model, achieving optimal performance with a training/testing split of 90:10. The ViT model shows that it can classify many different pathologies of medical imaging with equal veracity, and in doing so, presents a modern view of kidney classification. Ugur Kilic et al.17 present a CNN-based computer-aided diagnostic system to detect kidney stones in DUSX images automatically. Their study used a novel dataset comprising 630 DUSX images and combined the YOLOv4 model with the CBC technique for pre-processing, the dataset achieved a remarkable accuracy rate of 96.1%. The effectiveness of YOLOv4 in accurately detecting kidney stones highlights the importance of model selection for achieving high diagnostic performance, reinforcing the need to choose the right model to ensure optimal results.
Yao et al.18 showed deep and shallow combination features derived from 18F-FDG PET/CT scans to predict EGFR-sensitizing mutations in non-small cell lung cancer (NSCLC). The work of their fusion model improves the precise identification of genetic mutations for personalized treatment of cancer. According to Wang et al.19, the Neuroendoscopic Parafascicular Evacuation of Spontaneous Intracerebral Hemorrhage (NESICH) is an effective technique for minimally invasive hematoma removal. The multi-center study is promising preliminary outcomes in ICH management. In light of the work by Li et al.20,21, the screening methods for patients with and without the interfering medications for PA were investigated. Their findings point to the importance of maximizing PA diagnosis for the management of hypertension. Fan et al.22 showed that gut microbiota was compared in peritoneal dialysis patients with peritoneal fibrosis. Their work confirms a relationship between the microbiota composition and peritoneal function, which may cause dialysis-related problems. Xingguang et al.23 presented a robotic bronchoscope system for endovascular localization and biopsy on pulmonary lesions. This innovation improves accuracy and access to lung cancer diagnosis, a trend for robotics in medical procedures. Zhang et al.24 brought forth an image-guided para-cortical spinal tract approach for hematoma evacuation in spontaneous intracerebral hemorrhage (ICH) patients. Image-guided neurosurgical interventions can play an essential role in this method of neurosurgery as it minimizes brain damage and increases patient outcomes.
Kumar et al.25 developed a new deep-learning model using a fuzzy deep neural network to classify and predict kidney disease. Their accuracy of 98.78% achieved with this model far surpassed existing methods. This shows that deep learning can further improve kidney disease classification when enwrapped by fuzzy logic. Bingol suggested a novel deep-learning model with convolution and batch normalization layers upon diagnosing kidney disease26. Although the model has a low layer count, it achieves high success rates, and the results highlight the significance of architectural design in building practical diagnostic tools. In a complementary work, Altalbe et al.27 developed a deep-learning CNN model for classifying kidney disease. The approach consisted of transferring digital imaging and communications in medicine (DICOM) images to jpg files, with the features extracted by a CNN giving back results that were better than those of existing methods. Together, these studies demonstrate the ongoing advances in deep learning models for improved kidney disease detection and classification, with improved accuracy and potential clinical applications.
Ashafuddula et al.28 developed an intelligent diagnostic system for early-stage renal disease using novel healthcare data. An elegant feature of this study was that data integrity was thus prioritized so that no loss could occur, even if some values were missing. To speed up model training and better perform, I use dimension reduction to reduce the space of features. The study looked at four datasets and identified adaptive boosting, logistic regression, and passive-aggressive classifiers for CKD analysis in real-life data. In forecasting unseen therapeutic CKD data, particularly early-stage cases, reliable CKD classification was achieved with 96.48% accuracy using an ensemble characteristics-based classifier.
Jerlin et al.29 proposed a new CKD classification method based on a fruit fly optimization algorithm for feature selection and multi-kernel SVM for classification. This study obtained a classification accuracy of 98.5% for the CKD dataset, which exceeds the results of the existing hybrid kernel-based SVM, fuzzy min–max GSO neural network, and SVM methods. Zhang et al.28 proposed an innovative approach for detecting kidney lesions in CT scans. They increased the visibility of small lesions using morphological cascade CNNs. Key components included a modified six-layer feature pyramid network for varying-sized feature maps and a four-layer cascade region-based CNN with high-precision detection. These cascade RCNNs were also validated in experiments, which confirmed their superiority over previous methods. However, the study shows how deep learning models can help automate the detection and classification of renal illnesses, helping to diagnose renal illness faster and more accurately, positively impacting patient outcomes. This collection of studies shows the advances in deep learning models for detecting and classifying kidney diseases. Researchers improve diagnostic accuracy and efficiency by using various architectures and methodologies. Then, we leverage these advances and integrate our work with transfer learning models, including VGG16, ResNet50, CNNAlexnet, and InceptionV3, to improve image classification of kidney conditions in CT images.
Transfer learning methods
Malware detection using medical imaging as a motivating example, transfer learning has become a predominant strategy in medical imaging, allowing new tasks to reuse pre-trained models. This approach learns from neighboring tasks and optimizes deep learning model performance, which leads to a less time-consuming and, as a byproduct, less resource-consuming downstream model development process. Some recent studies, such as that by Sassanarakkit et al.5, have shown the state of the art of transfer learning in overcoming some difficulties with using a 3D convolution network in being adapted to different modality data. In 2023, Muneer Majid presented a computer-assisted diagnosis machine learning and fine-tuned transfer learning-based system for kidney tumor detection on CT images. We used extensive experiments to evaluate different deep learning models, with the conventional fine-tuned ResNet 101 and DenseNet 121 being the primary focus. At the same time, Mahalakshmi30 used an ensemble method that combines the classification results of three DCNNs to reduce variance and bias using Majority voting. An overall accuracy of 98.49% was achieved using Hyperparameter tuning using Gorilla Troops Optimizer. Furthermore, Badawy et al.31 also presented a renal disease classification framework using CT and histopathological images of kidneys through a CNN and transfer learning-based approach with superior accuracy compared to other approaches for renal disease classification. Therefore, the study’s objective was also to improve the classification of kidney diseases for accurate diagnosis and treatment.
Anari et al.32 suggested an explainable attention-based breast tumor segmentation model using UNet, ResNet, DenseNet, and EfficientNet. The resulting segmentation method is more accurate and retains the interpretability required for clinical decision-making. EfficientUNetViT is further developed by Anari et al.33 with EfficientUNet and a pretrained Vision Transformer for breast tumor segmentation. Transformer-based feature extraction is used with this method for segmentation efficiency and is shown to offer improved medical imaging performance. Sarshar et al.34 improved brain MRI classification using VGG16 and ResNet50 and the Multi-Verse Optimization (MVO) method. Their feature selection optimizes classification performance,they show how metaheuristic techniques can help in medical imaging. Kia et al.35,36 present a fusion model for MRI-based brain tumor identification using VGG16, MobileNet, EfficientNet, AlexNet, and ResNet50. Using this approach, we can effectively integrate several architectures to improve classification accuracy, providing a robust means to detect tumors. In 2024, Kia36 introduced an attention-guided deep learning model in the multi-criteria decision analysis for managing customer loyalty. One of the aspects that the study covers is the use of AI to analyze customer behaviour and optimize retention strategies.
Dalia Alzu’bi et al.37 utilized deep learning methods to study kidney tumor injuries and classify kidney tumor types. The researchers used VGG16, ResNet50, and two 2D-CNN models modified versions for analyzing features extracted from the renal CT scans. It was also a significant contribution because it used a unique dataset from King Abdullah University Hospital composed of 8,400 images of 120 adult patients with suspected kidney masses who underwent CT scans. An 80% training and 20% testing division was used on the dataset. Overall, the outcomes highlight the superiority of the 3 proposed 2D-CNN models, with 60, 96 and 97 percent detect accuracies for VGG16, ResNet50, and 2D-CNN, respectively. In a separate investigation, Lee and Aqil38 also studied variations in renal glomerular tissue diseases, focal segmental, normal, and sclerotic. They used allied and multivariate models to develop a technique to improve model accuracy, and excellent results were achieved using a combined model, which achieved an accuracy of as much as up to 97%. However, the processing time presented in the study was longer than the general model, but it was alleviated by using high-performance computing resources.
In addition, Parakh et al.39 evaluated the diagnostic efficacy of a cascaded deep learning system on unenhanced CT images of urinary stones. The impact of pre-train with labeled CT image in transfer learning is evaluated in their work, and they show that the GrayNet pre-trained model can obtain 95.4% in detecting urinary tract stones amongst all other pre-trained U-Net models. Transfer learning with enriched datasets was the focus of the study, which suggested the possibility of improving CNNs’ performance and generalization across scanners. Finally, as shown, AI/ML (deep learning and transfer learning) techniques can be beneficial in detecting and managing kidney stones. This research was created to advance systems for the early and accurate detection of kidney-related diseases, which are important to achieve better patient outcomes and healthcare efficiency. Combining these techniques can improve diagnostic accuracy, help reduce human errors, and prescribe individualized therapeutic strategies. Further work is needed to overcome the current limitations and unlock these approaches’ potential for the clinical management of kidney stone disease. The study shows the need for a continuing research effort to improve the performance of these methods for practical implementation in healthcare applications. The studies above show that transfer learning works well for kidney disease classification, yet they also have some shortcomings. This could be because some of these models require great computational resources and cannot be applied practically in the clinical context. At the same time, the models rely on pre-trained models that need to be carefully scrutinized for the features learned from the source datasets to be viable for the target task.
Recent advancements in kidney disease detection
With deep learning and transfer integration, hyperparameter optimization, kidney disease detection, and classification have been made much better. Recent studies have shown that these technologies improve diagnostic accuracy. For instance, Kumar et al.25 developed a fuzzy deep neural network model that predicted kidney diseases with high accuracy compared to other developed methods. Besides ensemble methods and hybrid models, results promise to improve classification accuracy. In their research40, Ashafuddula et al. show that an ensemble characteristics-based classifier is effective for CKD analysis and achieves an accuracy of 96.48% in early-stage cases.
Table 1 summarizes recent advancements in transfer learning and hyperparameter optimization, highlighting their contributions and limitations in kidney disease classification. Here, key developments like applying pre-trained models, fine-tuning techniques, and an ensemble learning approach are used to improve diagnostic accuracy. However, challenges like generalization problems, computational cost, and complexity remain in managing the model. However, ViT and CycleGANs have shown promise in additional methods, requiring large datasets and many computational resources. This comprehensive overview qualifies the evolving enigmatic nature of intelligence within medical imaging and dedicates further exploration, showcasing the imperative for further research to address current restricting factors and improve patient outcomes.
link