Analysis of transformer and CNN based approaches for classifying renal abnormality from image data

Abstract

There is a pressing need to revise the current diagnostic framework for renal abnor mality due to the projected increase in its global prevalence as about 10% of people worldwide are suffering from renal diseases. Recognizing the escalating trends of renal disease, proactive measures are warranted to overcome upcoming challenges in accurate diagnosis and management. Renal abnormalities, often symptomless and hard to diagnose, can be dangerous but curable if detected early. Therefore, machine learning and deep learning techniques can be instrumental if implemented correctly to determine this anomaly early in this modern time. Our approach for renal abnormality detection from image data incorporates the topologies of Con volutional Neural Networks and transformer-based image classification topologies, as well as data augmentation methods and precise hyperparameter tuning (learn ing rate, batch size, dropout rate, regularization strength, etc.); additionally, we proposed CNN-based and transformer-based architectures for renal abnormality de tection. Transformer-based deep learning methods are the latest trend in classify ing diseases from medical images; for this reason, we analyzed the performance of CNN-based architectures and transformer-based architectures. We build a hybrid binary class dataset of Computed Tomography(CT) scan renal images using pri mary data collected from Kidney Foundation Hospital & Research Institute, Dhaka, Bangladesh and secondary data from publicly available online source. Our approach is a sequence of steps that allows for the abnormality detection using state-of-the art classifiers ResNet50, Inception ResNetV2, InceptionV3 and VGG16 along with our proposed ResNet152 based custom model and ViT architecture-based custom model without manual intervention. Our experimental results showed that our pro posed transformer-based model achieved the highest accuracy of 99.99% while our proposed CNN model achieved an accuracy of 99.97%. Among the four pre-trained CNN models, ResNet50 scored the highest accuracy of 99.95%, and VGG16 scored 99.92%, InceptionResNetV2 was able to score 98.87%, while the lowest performance was shown by the InceptionV3 model, which was 96.87%. All four pre-trained mod els have demonstrated acceptable performance, and our proposed model was able to perform better than state-of-the-art prepared models.