Genetic Algorithm Optimized Vision Transformer Based Transfer Learning Approach for Early Prediction of Chronic Kidney Disease

Chronic Kidney Disease (CKD) is an ongoing health issue defined by the slow decline in kidney performance over a prolonged period. This condition can escalate to kidney failure, which is life-threatening without the absence of dialysis or a kidney transplant. The common key factors leading to CKD include diabetes, hypertension, Glomerulonephritis. Standard diagnostic practices tend to be laborious, require considerable resources, and are vulnerable to errors made by humans. With the ongoing advancements in artificial intelligence within the healthcare domain, machine learning and deep learning algorithms are playing a pivotal role in the accurate and effective detection of CKD. The purpose of the proposed research is to construct and authenticate a predictive model aimed at diagnosing chronic kidney disease. Image processing-based diagnostic approaches have shown a greater success rate compared to other detection methods. To tackle these challenges, our investigation proposes a new methodology that merges Genetic Algorithms (GA) with a Vision Transformer (ViT) model that employs Hierarchical Attention in a transfer learning framework, thereby improving both feature selection and classification accuracy. In this scholarly work, we have scrutinized the performance of two leading architectures, VGG16 and ResNet50, and have proposed an attention-centric approach that employs the Vision Transformer model optimized by evolutionary algorithm. The attention mechanism has the potential to grasp long-term dependencies in images. Extraction of complex features is to be done using ViT Model, preceded by several hyper parameters like number of epochs, learning rate, batch size, number of layers, layer size, number of attention heads, and attention window size should be optimized through a genetic algorithm to enhance performance and also feature selection is done by GA to get the optimized result. A dataset available to the public consists of four types of image data of kidneys: Cyst, Normal, Tumor, and Stone, which are used in the three architectures discussed above. In total, there are 12,446 images that are segmented for training, testing, and validation purposes. Experimental results demonstrate that our GA optimized ViT Model surpasses state-of-the-art traditional models, achieving a 98.05% F1-score, and the model also shows superiority in terms of trainable parameters