PulmoX-Net: a channel-attention enhanced deep learning model for multi-class pulmonary pathology classification in...

Area of Science:

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

Background:

• Respiratory diseases present a significant global health challenge.
• Chest radiography is a primary diagnostic tool for pulmonary abnormalities.
• Interpreting chest X-rays is complex due to overlapping patterns of various lung conditions.

Purpose of the Study:

• To develop an advanced deep learning model for multi-class chest X-ray classification.
• To enhance feature representation in chest X-ray analysis using attention mechanisms.
• To establish a benchmark for attention-enhanced classification of pulmonary diseases.

Main Methods:

• Developed PulmoX-Net, a hybrid convolutional neural network integrating Xception architecture with Squeeze-and-Excitation (SE) channel attention.
• Evaluated PulmoX-Net on the Kaggle 'X-ray Lung Diseases Images (9 classes)' dataset (6,743 images).
• Conducted supplementary analyses including per-class metrics, repeated testing, and clinician review of model interpretability using Grad-CAM.

Main Results:

• PulmoX-Net achieved high performance: 89.23% overall accuracy, 89.92% precision, 89.38% recall, and 88.97% F1-score.
• The model demonstrated a macro-average AUC of 0.9842.
• PulmoX-Net outperformed other evaluated convolutional neural network (CNN) backbones.

Conclusions:

• PulmoX-Net serves as a promising benchmark for attention-enhanced multi-class chest radiograph classification.
• The model's performance indicates its potential for improving the interpretation of pulmonary abnormalities.
• External validation on diverse clinical cohorts is necessary before widespread clinical adoption.