الفهرس | Only 14 pages are availabe for public view |
Abstract Deep learning models based on Convolutional Neural Networks (CNNs) have gained much attention in the area of image denoising, especially with medical images for automated diagnosis. Because of the similarity of diseases and the limited number of medical images, obtaining images with high quality represents a big challenge in the diagnostic process. Denoising is one of the main problems in image processing. Although it reduces noise, it may degrade the original image quality if some information of interest is lost. This thesis is concerned with the quality enhancement of medical images (Xray and CT) for efficient automated medical diagnosis, especially for COVID-19 detection depending on a proposed model called Classification Autoencoder Denoising transfer learning (CADTra). The quality enhancement is performed through noise reduction. An autoencoder algorithm is used for noise reduction and reconstruction of the image with highly-defined features. Only three layers are used in the encoder and decoder to reduce the computational cost and speed up the process of denoising . The denoising stage is used as a pre-processing stage to enhance the classification process performance. The subsequent image classification process is implemented with different deep learning models: one trained from scratch called Four-layer CNN (FCNN), and 12 pre-trained models (AlexNet, LeNet-5, VGG16, Inception naïve v1, DenseNet121, DenseNet169, DenseNet201, ResNet50, ResNet152, VGG16, VGG19, and Xception) with the help of transfer learning. The objective of transfer learning is to make use of the pre-trained model weights with some sort of finetuning to avoid the need for large databases for training. The proposed framework that comprises both denoising and classification is superior in performance to that working on noisy images. It approaches in performance the one that works on noise-free images. A comparative study is introduced between the proposed and existing approaches for noise removal from medical images. The proposed approach is superior from the Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), accuracy, precision, recall, F1-score, confusion matrix, and Receiver Operating characteristic (ROC) curve perspectives. It achieves an average accuracy of 98.34% in binary classification of CT scans and 98.42% in multi-class classification of X-ray images. Hence, the proposed framework can be recommended for utilization in automated diagnosis systems for COVID-19 cases to limit the spread of the corona virus in the world. |