A Robust CNN Framework with Dual Feedback Feature Accumulation for Detecting Pneumonia Opacity from Chest X-ray Images

TL;DR

This paper introduces a robust deep learning framework with dual feedback feature accumulation for accurate and efficient pneumonia detection from chest X-ray images, outperforming traditional models in accuracy and parameter efficiency.

Contribution

The study proposes a novel CNN architecture with Process Convolution blocks and dual feedback for improved feature extraction and reduced model size, enhancing pneumonia detection accuracy.

Findings

Achieved 97.78% accuracy, 98.84% sensitivity, 95.04% specificity.

Model has significantly fewer parameters than traditional networks.

Lower false-negative rate indicates reliable pneumonia detection.

Abstract

Pneumonia is one of the most acute respiratory diseases having remarkably high prevalence and mortality rate. Chest X-ray (CXR) has been widely utilized for the diagnosis of this disease owing to its availability, diagnostic speed and accuracy. However, even for an expert radiologist, it is quite challenging to readily determine pneumonia opacity by examining CXRs. Therefore, this study has been structured to automate the pneumonia detection process by introducing a robust deep learning framework. The proposed network comprises of Process Convolution (Pro_Conv) blocks for feature accumulation inside Dual Feedback (DF) blocks to propagate the feature maps towards a viable detection. Experimental analysis showcase: (1) the proposed network proficiently distinguishes between normal and pneumonia opacity containing CXRs with the mean accuracy, sensitivity and specificity of 97.78%, 98.84% and 95.04%, respectively; (2) the network is constructed with significantly low parameters than the traditional ImageNets to reduce memory consumption for deployment in memory constrained mobile platforms; (3) the trade-off between accuracy and number of parameters of the model outperforms the considered classical networks by a remarkable margin; and (4) the false-negatives are lower than the false-positives (both of which are low in count) which prove the model's low-fatality prediction. Hence, the proposed network can be deployed for a rapid screening of pneumonia and can act as a great assistive tool for the radiologists in the diagnosis process.