Hybrid generative modeling with transformer-based classification for medical image-based silicosis detection

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Occupational Health

Background:

• Silicosis is a chronic, irreversible lung disease caused by silica inhalation, necessitating accurate diagnosis via chest X-rays (CXRs).
• Detecting silicosis from CXRs is challenging due to subtle texture variations and limited annotated data.

Purpose of the Study:

• To develop a novel hybrid deep learning framework for precise and automated silicosis detection from CXR images.
• To address limitations in current diagnostic methods, including inter-class texture changes and data scarcity.

Main Methods:

• A hybrid framework combining transformer-based classification, latent feature representation, metaheuristic optimization, and generative data augmentation.
• Image preprocessing with CLAHE, lung segmentation using K-means clustering, and data augmentation via Generative Adversarial Networks (GANs).
• Feature extraction using Variational Autoencoder and classification with a CNN-Enhanced Residual Multiscale Transformer Network (ConvRMT-Net), optimized with Frigate Bird Optimization.

Main Results:

• The proposed framework achieved superior accuracy, sensitivity, and specificity in silicosis detection compared to existing methods.
• Demonstrated an average improvement of 10.43% in diagnostic performance metrics.
• Successfully captured both local spatial patterns and global contextual interactions in CXR images.

Conclusions:

• The hybrid deep learning framework offers a promising solution for accurate and automated silicosis diagnosis.
• The integration of advanced AI techniques effectively overcomes challenges in CXR-based lung disease detection.
• This approach has the potential to improve clinical outcomes by enabling prompt and precise diagnosis of silicosis.