Lung Disease Classification with Deep Learning Enhanced CNN Architecture in Chest X-Ray Imaging.
Authors
Affiliations (2)
Affiliations (2)
- Department of Medical Imaging and Radiation Sciences, 3001 12th Avenue North, Sherbrooke, Qc, J1H5N4, Canada.
- Department of Medical Imaging and Radiation Sciences, 3001 12th Avenue North, Sherbrooke, Qc, J1H5N4, Canada. [email protected].
Abstract
In this study, we introduce a robust method using a robust convolutional neural network (CNN) architecture for efficient segmentation and multi-classification of lung X-ray pathologies, by replacing the traditional max pooling with discrete wavelet transform (DWT), which provides more accurate down-sampling and enhances the detection of fine lung structure details. Integrated with an advanced U-Net + + model and Attention Gates (AG), our method significantly improves lung segmentation accuracy. For lung pathology classification, we integrated DWT in the DenseNet-201 model to differentiate normal lung images from images with tuberculosis, pneumonia, and coronavirus disease 2019 (COVID-19). To address the challenges of limited and variable data, we employed the technique progressive growing generative adversarial network (PGGAN) data augmentation to generate realistic, high-resolution chest X-ray (CXR) synthetic images. This approach not only enriches the training dataset but also provides a nuanced representation of lung pathologies, enhancing the robustness and comprehensiveness of our diagnostic system. Our robust approach demonstrated strong and consistent performance in both segmentation and classification tasks within lung X-ray imaging diagnostics. In segmentation, it achieved on the Japanese Society of Radiological Technology (JSRT) dataset, with metrics such as 99.1% accuracy and 97.2% Dice coefficient, outperforming established methods like U-Net and U-Net + + . For the classification, it demonstrated notable improvements over DenseNet-201, especially in precision with an increase of 2.4% when data augmentation techniques were employed. These advancements suggest a significant step forward in accuracy and reliability for CXR image analysis, affirming our method's superior adaptability and potential in handling diverse and augmented datasets.