Hybrid Feature-Based Classification of Glomeruli Biopsy Images Using Vision Transformers and Statistical Feature Augmentation

The identification of abnormalities such as glomerulosclerosis is one of the most important aspects of the glomeruli biopsy study that is used in the diagnosis of kidney illnesses. For the purpose of classifying glomeruli biopsy images into Normal and Sclerosed categories, this work implements a hybrid classification system. The dataset, which was obtained from Kaggle, was processed with Vision Transformers (ViTs) for the purpose of feature extraction without any additional training being required. To be more specific, one thousand deep features were extracted from the head layer of the Vision Transformer model that had been first trained. In order to improve the effectiveness of classification, twelve statistical characteristics, which included mean, minimum, maximum, entropy, kurtosis, skewness, and root mean square, were computed and added to the deep features that were retrieved. This resulted in a hybrid representation that contained 1,012 features. In the subsequent step, traditional machine learning classifiers were utilized for the purpose of image classification. Evaluation and comparison of the performance of these classifiers were carried out, with a particular emphasis placed on the enhancement that was accomplished by using statistical characteristics. The findings of the experiments show that the hybrid model that was developed performs better than the baseline deep features in terms of accuracy and resilience. This indicates that the hybrid model is a promising technique for the classification of glomeruli biopsy images.