Limitations of principal component analysis in COVID-19 CT image classification

This study critically evaluates the methodological approach employed by Nayak et al. in COVID-19 CT scan classification using their Dense-Res-Inception Ensemble Network with PCA dimensionality reduction. We demonstrate fundamental limitations in applying PCA to nonlinear biological imaging data, where key mathematical assumptions—including linearity, meaningful correlations, and homoscedasticity—are frequently violated. Using the MNIST dataset as a benchmark, we comparatively assessed three dimensionality reduction techniques: PCA, High Variance Gene Selection, and Feature Agglomeration (FA). Results confirm that FA significantly outperforms PCA (92.79% vs 83.76% accuracy) by preserving crucial spatial relationships within image data. This performance disparity highlights the critical importance of methodological alignment with data characteristics in medical imaging analysis. We propose that nonlinear dimensionality reduction approaches better accommodate the complex relationships inherent in biological systems, potentially enhancing both computational efficiency and diagnostic reliability in clinical applications requiring rapid assessment.