الفهرس 
Chapter 1 : IntroductionOnly 14 pages are availabe for public view
 |  | from | 135 |  |  |
| | | from | 135 | | |
Abstract
Automated medical diagnosis is one of the main tasks of the research community in biomedical engineering. Medical image segmentation, fusion, interpolation and classification are required tasks for the overall automated diagnosis task. The operation principles differ among these modalities and performing these individual tasks. This thesis is mainly concerned with efficient medical image segmentation, fusion, interpolation and classification. For the segmented task different algorithms and techniques including watershed segmentation, threshold segmentation, K-means clustering, adaptive K- means clustering, active contour and Otsu are presented and their results are compared. In addition, a hybrid segmentation structure comprising watershed and threshold is proposed. The segmentation scenarios are investigated and tested through simulation experiments on images of different modalities including confocal microscopy retinal image, Magnetic Resonance (MR) and Computed Tomography (CT) images. Simulation results proved that the performance of the proposed segmentation methods achieves higher accuracy than traditional approaches. The hybrid segmentation method which one of the proposed approaches achieves accuracy 99.8 % for MR brain images. The active contour segmentation method achieves accuracy of the CT Coronavirus Disease 2019 pneumonia (COVID-19) images 99.59%.
In addition, the classification scenarios are introduced depending on deep learning concepts. These scenarios depend on Convolutional Neural Network (CNN) and Convolutional-Long Short Term Memory (Conv-LSTM). Results prove that the performance of the deep learning is better than traditional techniques known in this field. Using only CNN gave an accuracy of 94.26% for Diabetic Retinopathy (DR) and 95.8% for brain tumor, while the hybrid Conv-LSTM gave the maximum accuracy of 96.41% for DR, and 97.4% for brain tumor.
Image fusion concept is considered in this thesis. The objective of the fusion process is to merge images of different modalities for the same region in the human body. The fusion is implemented on this thesis on MR and CT images to collect more important features required for better diagnosis. An additional interpolation stage is implemented on the fusion results to obtain images with better resolution. In addition, medical image fusion is investigated consider the Non-Sub-Sampled Shearlet Transform (NSST) with helping of Modified Central Force Optimization (MCFO). Finally and based on the simulation results, a general framework comprising image fusion, interpolation, segmentation and classification can be recommended for successful automated diagnosis.