Lung Disease Classification with Deep Learning Enhanced CNN Architecture in Chest X-Ray Imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Accurate segmentation and classification of lung pathologies from chest X-rays (CXRs) are crucial for diagnosis.
• Traditional methods face challenges with fine details and limited/variable datasets.
• Deep learning models show promise but require optimization for complex pathologies.

Purpose of the Study:

• To develop a robust convolutional neural network (CNN) for enhanced lung X-ray segmentation and multi-classification.
• To improve the detection of fine lung structures and pathological details.
• To address data limitations using advanced data augmentation techniques.

Main Methods:

• Replaced max pooling with discrete wavelet transform (DWT) in a U-Net++ model with Attention Gates (AG) for segmentation.
• Integrated DWT into DenseNet-201 for multi-classification of lung pathologies (tuberculosis, pneumonia, COVID-19).
• Employed progressive growing generative adversarial network (PGGAN) for data augmentation to generate realistic synthetic CXR images.

Main Results:

• Achieved 99.1% accuracy and 97.2% Dice coefficient in lung segmentation on the JSRT dataset, outperforming U-Net and U-Net++.
• Demonstrated improved classification precision by 2.4% over DenseNet-201 with PGGAN data augmentation.
• The integrated DWT approach enhanced detection of fine lung structures and pathological features.

Conclusions:

• The proposed DWT-integrated CNN method offers superior performance in lung X-ray segmentation and multi-classification.
• PGGAN data augmentation effectively enriches datasets, improving model robustness and diagnostic accuracy.
• This approach represents a significant advancement in automated CXR analysis for diverse lung pathologies.