Kidney diseases identification using deep and handcrafted models: a comparative study

Abstract

Kidney cysts, tumors, and stones pose significant health risks to individuals, potentially leading to kidney failure if not effectively managed. However, the shortage of nephrolo gists and kidney specialists globally presents a significant challenge in providing timely and accurate diagnoses. In this research, we aim to address this challenge by leveraging artificial intelligence (AI) techniques for the detection of kidney cysts, tumors, stones, and normal conditions this was based on early research that made a comparative study between cnn models and transformer variants. To expand that research Our study encompasses an invistigation involving diverse models, including CNN models, transformer variants, lightweight neural networks, and handcrafted features. We evaluate these models on a dataset of 12,446 unique images extracted from abdominal CT scans. To enhance the comprehensiveness of our investigation, we introduce several modifications to the original dataset, incorporating Gaussian noise, blur, illumination variations (both increase and decrease), and occlusion. These modifications aim to simulate real-world conditions and imaging artifacts, providing a more realistic and challenging dataset for evaluation purposes. Upon comparing the performance of these models, we observed interesting outcomes. In the original dataset, the BSIF handcrafted features exhibited superior accuracy, achieving an impressive accuracy of 99.51%. In the BLUR dataset, the lightweight neural networks outperformed all other models, achieving an accuracy of 99%. However, in the ILLU MINATION INCREASE dataset, we noted a decline in classifi-cation results when using Vision transformers models (EANet and SWIN). Conversely, in the OCCLUSION dataset, the lightweight neural networks surpassed all other models in terms of accuracy. In the GAUSSIAN dataset, the lightweight neural networks once again outperformed all other models, achieving an accuracy of 99%. Similarly, in the ILLUMI- NATION DECREASE dataset, the lightweight neural networks demonstrated superior accuracy