Diagnosis of chest X-Rays images with Deep Learning enhanced with discrete wavelet transform

Deep learning (DL) is increasingly used in medical imaging to improve image acquisition technology and aid in disease diagnosis and classification. In this work, we present an enhancement of DL applied in lung chest X-rays (CXR) imaging first to enhance lung image segmentation, and second to accurately classify images based on their status of normal, or with tuberculosis, viral pneumonia, or COVID-19 induced pneumonia. The images were obtained from public datasets: 247 from the Japanese Society of Radiological Technology and the Japanese Radiological Society; 138 from the Health and Human Services, Montgomery County (MC), Maryland, USA; 800 from Shenzhen and Montgomery; and 33,920 from Qatar University. For the segmentation, we replaced the Max-Pooling in U-Net++ with the Discrete Wavelet Transform (UNet++-Rand-DWT) integrated with Attention Gates. Prior to U-Net++-Rand-DWT, the images were normalized with the Contrast Limited Adaptive Histogram Equalization (CLAHE) and augmented with the Progressive Growing Generative Adversarial Networks (PGGAN) by generating synthetic images to balance the different image classes. For image classification, the DenseNet-201 model was enhanced with Rand-DWT (DenseNet-201-Rand-DWT). The results for the segmentation obtained with U-Net++-Rand-DWT demonstrated greater accuracy, both qualitatively and quantitatively, compared to U-Net++ with Max-Pooling and to four other models, as evaluated by accuracy, specificity, sensitivity, Dice coefficient and Jaccard Index. For pathology classification, DenseNet-201-Rand-DWT outperformed the traditional U-Net++ and U-Net, and the synthetically generated images with PGGAN significantly enhanced classification accuracy, as determined by precision and F1-score metrics. Furthermore, DenseNet-201-Rand-DWT demonstrated an enhanced performance in discriminating pneumonia and COVID-19 CXR images, even without data augmentation.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.