Explainable hybrid transformer for multi-classification of lung disease using chest X-rays

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Lung disease is a leading global cause of death.
• Thoracic X-rays are a cost-effective screening tool for lung disease.
• Deep learning models are increasingly vital for diagnosing lung diseases from X-ray images.

Purpose of the Study:

• To develop an explainable hybrid deep learning model for multi-lung disease classification using Chest X-ray images.
• To improve feature recognition and diagnostic accuracy through a novel network architecture.

Main Methods:

• Proposed LungMaxViT: a hybrid transformer combining CNN and SE blocks.
• Utilized transfer learning with pre-trained models (ResNet50, MobileNetV2, ViT, MaxViT) on two public X-ray datasets.
• Applied enhancement techniques (CLAHE, flipping, denoising) and Grad-CAM for explainability.

Main Results:

• LungMaxViT achieved 96.8% accuracy, 98.3% AUC, and 96.7% F1-score on a COVID-19 dataset.
• On the Chest X-ray 14 dataset, LungMaxViT reached 93.2% AUC and 70.7% F1-score.
• LungMaxViT demonstrated superior performance over classical pre-training models and other hybrid networks.

Conclusions:

• The proposed LungMaxViT model offers robust and generalizable detection of multiple lung lesions and COVID-19 from Chest X-rays.
• Explainable AI techniques like Grad-CAM confirm consistency between model predictions and clinical interpretation.
• LungMaxViT shows significant potential to assist clinicians in diagnosing lung diseases.