An Improved CovidConvLSTM model for pneumonia-COVID-19 detection and classification

TL;DR

This paper introduces an improved CovidConvLSTM model that combines advanced neural network components to enhance accuracy and reduce computational costs in pneumonia and COVID-19 detection from medical images.

Contribution

The paper presents a novel deep learning framework integrating RegNetX002, ConvLSTM, and SE blocks to address overfitting, performance degradation, and high computational complexity in COVID-19 detection.

Findings

Achieved 98.22% accuracy on CPN-CXRPA dataset.

Achieved 98.78% accuracy and F1 score on CXRI-P-C-CXR dataset.

Outperforms existing models in accuracy and efficiency.

Abstract

Recently, COVID-19 pandemic has rapidly evolved into a critical global health crisis, profoundly impacting daily life. As a result, CAD systems have gained significant interest for its massive computational capabilities, which facilitate the rapid analysis and interpretation of medical imaging. In particular, Deep Learning (DL )techniques have emerged as critical tools to assist radiologists and pulmonologists in distinguishing COVID-19 patients from other pneumonia types and healthy cases. Unfortunately, existing DL techniques face several challenges such as overfitting, performance degradation, feature irrelevance and redundancy, vanishing gradient problem, and high computational complexity. In this paper we address these challenges by introducing an enhanced Convolutional Neural Network algorithm that combines a bottleneck based model RegNetX002, ConvLstm layer, and Squeeze and Excitation block (SE). Specifically, the RegNetx002 and the ConvLstm layer are used for features map extraction and feature quality enhancement, while the attention mechanism SE block is employed to improve feature representation by highlighting important channel features and suppressing unimportant features. More importantly, The bottleneck module facilitates the extraction of more abstract features while lowering computational costs. Additionally, it incorporates residual connections that helps reducing the vanishing gradient problem. Balanced CPN-CXRPA and imbalanced CXRI-P/C-CXR datasets are used to assess the proposed model. Performance metrics such as accuracy and F1 score are used to evaluate the model efficiency. Using the CPN-CXRPA dataset, our model achieved an accuracy of 98.22%. For the CXRI-P-C-CXR dataset, it achieved 98.78% of both accuracy and F1 score. The experimental results show that this framework outperforms existing models in terms of performance and computational complexity.