# Squeeze-and-Excitation Enhanced Convolutional Neural Networks for Multi-class Pneumonia Classification on Chest Radiographs

## TL;DR

This study shows that adding squeeze-and-excitation mechanisms to CNNs improves multi-class pneumonia classification on chest X-rays with high accuracy.

## Contribution

The novel use of squeeze-and-excitation modules with ResNet50V2 and InceptionV3 for multi-class pneumonia classification on chest radiographs.

## Key findings

- ResNet50V2-SE achieved 98.18% test accuracy and 0.9951 AUC for pneumonia classification.
- Both models showed high F1-scores across all classes, with no significant performance difference confirmed by McNemar's test.
- SE-enhanced CNNs demonstrated strong generalizability for differentiating pneumonia subtypes on chest X-rays.

## Abstract

This study compared two convolutional neural network (CNN) architectures, ResNet50V2 and InceptionV3, each enhanced with squeeze-and-excitation (SE) attention mechanisms, for automated classification of chest X-rays into normal, pneumonia-bacterial, pneumonia-viral, and COVID-19 categories. A total of 9,208 posterior-anterior chest radiographs were analyzed, divided into training, validation, and test datasets under identical preprocessing and fine-tuning conditions.

ResNet50V2-SE achieved a test accuracy of 98.18% and a macro-averaged area under the curve (AUC) of 0.9951, while InceptionV3-SE achieved 97.86% accuracy and an AUC of 0.9949. Class-specific evaluation showed that ResNet50V2-SE classified normal radiographs with perfect precision, recall, and F1-score (1.00). Pneumonia-bacterial images were classified with a precision, recall, and F1-score of 0.99, while pneumonia-viral and COVID-19 images reached F1-scores of 0.98 and 0.95, respectively. Comparable results were observed with InceptionV3-SE, achieving F1-scores of 1.00 for normal, 0.99 for pneumonia-bacterial, 0.97 for pneumonia-viral, and 0.95 for COVID-19.

McNemar's test was performed to compare model performance on the same test set. ResNet50V2-SE was correct in eight cases where InceptionV3-SE was incorrect, whereas InceptionV3-SE was correct in two cases where ResNet50V2-SE was incorrect. The resulting χ² statistic was 0.1250 (p>0.05), indicating no statistically significant difference in classification performance between the two models.

Both SE-enhanced CNN architectures demonstrated high accuracy and generalizability for differentiating among pneumonia subtypes on chest X-rays. These findings suggest that attention-augmented deep learning models may serve as effective decision-support tools in radiologic diagnosis, warranting further validation across larger and more diverse clinical datasets.