Hybrid Attention-Based CNN-BiLSTM Network for Aiding Covid-19 Diagnosis

Le résumé fourni par la source

Accurate interpretation of chest X-ray (CXR) is vital for early detection of diseases like pneumonia and COVID19. Though CNN perform well in medical image classification, they often miss subtle spatial features due to limited receptive fields and uniform region processing. To overcome these limitations, attention mechanisms have emerged as effective tools for enhancing feature representation by enabling the network to focus selectively on salient areas. Thus, this study presents a comparative study of three attention modules which are Squeeze-and-Excitation (SE), Bottleneck Attention Module (BAM), and Coordinate Attention (CA) with each integrated into a baseline CNN to improve its discriminative capability. Additionally, hybrid architecture is proposed by extending the best attention-integrated CNN with a Bidirectional Long ShortTerm Memory (BiLSTM) network, allowing the model to capture sequential spatial dependencies across the image. The classification analysis was conducted on a balanced dataset of 15000 CXR images spanning three diagnostic categories of normal, pneumonia, and COVID-19 cases. Results show all attention modules improved over the baseline CNN accuracy of 87.62 %, reaching up to 90.57 %, with CNN+BAM showing the largest gain. The hybrid CNN+BAM-BiLSTM model achieved the highest accuracy of 99.35 %, along with 99.35 % precision,$\mathbf{9 9. 4 2 \%}$recall, and$\mathbf{9 9. 3 9 \%}$F1-score. Grad-CAM visualizations further revealed that attention-integrated models produced more localized and clinically relevant activation maps.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.