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

Area of Science:

• Medical Imaging and Artificial Intelligence
• Deep Learning in Radiology
• Computer-Aided Diagnosis

Background:

• Accurate classification of chest X-rays is crucial for diagnosing various respiratory conditions, including different types of pneumonia and COVID-19.
• Traditional diagnostic methods can be time-consuming and may be subject to inter-observer variability.
• Deep learning models, particularly convolutional neural networks (CNNs), show promise for automating and improving the accuracy of medical image analysis.

Purpose of the Study:

• To compare the performance of two SE-enhanced CNN architectures, ResNet50V2-SE and InceptionV3-SE, for the automated classification of chest X-rays.
• To evaluate the models' ability to differentiate between normal radiographs, bacterial pneumonia, viral pneumonia, and COVID-19.
• To assess the generalizability and diagnostic potential of attention-augmented deep learning models in radiologic diagnosis.

Main Methods:

• Utilized a dataset of 9,208 posterior-anterior chest radiographs.
• Employed two CNN architectures: ResNet50V2 and InceptionV3, both enhanced with squeeze-and-excitation (SE) attention mechanisms.
• Trained and validated models on distinct datasets, applying identical preprocessing and fine-tuning techniques.

Main Results:

• ResNet50V2-SE achieved 98.18% test accuracy and an AUC of 0.9951; InceptionV3-SE achieved 97.86% accuracy and an AUC of 0.9949.
• Both models demonstrated high class-specific performance, with F1-scores ranging from 0.95 to 1.00 across all categories.
• McNemar's test revealed no statistically significant difference in classification performance between the two models (p>0.05).

Conclusions:

• SE-enhanced ResNet50V2 and InceptionV3 CNNs are highly accurate and generalizable for classifying chest X-rays into normal, bacterial pneumonia, viral pneumonia, and COVID-19 categories.
• These attention-augmented deep learning models show significant potential as effective decision-support tools for radiologists.
• Further validation on larger, diverse clinical datasets is recommended to confirm the clinical utility of these advanced diagnostic tools.