Topological Data Analysis for COVID-19 Classification from Lung CT-Scan Images

Abstract

Topological Data Analysis (TDA) has emerged as a promising framework for capturing structural patterns in medical images; however, existing approaches for COVID-19 CT classification, particularly those based on Riemannian representations of persistence diagrams, often involve complex non-Euclidean computations that limit scalability and integration with standard machine learning models. In this paper, we propose a TDA-based classification framework that utilizes persistence landscapes as a computationally efficient Euclidean representation of persistent homology. Each CT image is transformed into a Feature Point Cloud via an intensity-based covering and spatial clustering procedure, from which topological features (H0, H1, H2) are extracted using Vietoris–Rips filtration, along with an additional descriptor obtained from lower-star filtration. These features are vectorized into a fixed-length representation and used for classification. Experiments on a balanced dataset of 1000 CT images with a 70:20:10 train/validation/test split demonstrate that the proposed method achieves accuracies of 96% using Support Vector Machines and 98% using Random Forest. An ablation study further reveals that the H0 feature consistently provides the strongest discriminative performance, indicating that connectivity-based structures are more robust and informative than higher-dimensional topological features for distinguishing COVID-19 CT images, while maintaining computational efficiency and compatibility with classical machine learning approaches