Convnext-PCA: A Parameter-Efficient Model for Accurate Kidney Abnormality Classification

This study proposes the integration of principal component analysis (PCA), convolution, dense residual network (DRN) for the image classification of healthy kidneys, renal cysts, stones, and tumors. PCA significantly reduces lowvariance features in the feature-map derived by the stack of convolution-blocks, without losing information content. Together, PCA and DRN reduce the information-loss and improve accuracy. PCA also improves computational efficiency. An algorithm has been presented. The model was tested on three standardized datasets comprising CT scan images, fully annotated by expert clinicians. Ablation studies with and without PCA confirm that the use o,f PCA improves the mean accuracy from 89.54% to 99.89%. Comparison with other leading models shows that reducing low-variance features improved the computational efficiency by more than 60%, while achieving 99.89% mean accuracy with 0.0007 standard deviation. Our model outperforms accuracy achieved by the leading models and exhibits a classification time of 2.9 seconds on a laptop for each test-case.