Deterministic Top-k Activation Sparse Encoder for Compact Feature Selection in Covid-19 Chest X-ray Classification

Abstract

Rapid identification of COVID-19 from chest X-ray images remains essential for scalable clinical screening systems. Traditional convolutional and transfer learning designs are often affected by high-dimensional redundancy, noise anatomical interference, and reconstruction-based sparse representation learning which enhance computational complexity. A Top-K Activation Sparse Encoder (TKASC) is proposed, a lightweight feature selection framework, specifically aimed at classification-based medical imaging pipelines. The suggested architecture implements activation-level sparsity, and only the K largest responses of neurons to an input sample are kept, this way eliminating decoder reconstruction dependency and guaranteeing a fixed budget of neurons to use across samples. The technique generates small latent representations, which inhibit irrelevant anatomical responses and still retain diagnostically meaningful features. Sparse representations are directly incorporated with the downstream classifiers without overhead reconstruction optimization. Multi-class chest X-ray experimental data show enhanced discriminative ability in comparison to CNN-based classifiers, ResNet-50, LSSVM and L1-regularized sparse autoencoder baselines. Performance results include 91.4% precision, 89.7% recall, 90.5% F1-score, 90.8% accuracy, and 92.7% AUC-ROC. Training and validation curves showed stable convergence behavior, which validates reliable generalization performance. Deterministic activation sparsity and decoder-free representation learning create an efficient feature compression scheme, which would be applicable in clinical screening conditions in real-time.