الفهرس 
Chapter 1 IntroductionOnly 14 pages are availabe for public view
 |  | from | 166 |  |  |
| | | from | 166 | | |
Abstract
”Computer vision is a branch of artificial intelligence that focuses on processing and analyzing visual data. The goal of computer vision is to allow machines to distinguish between objects as people do, by simulating human perception and allowing machines to recognize objects. Although mathematical models cannot equal a person’s visual talents, they are nonetheless superior to their vision. This field offers a wide range of applications in industry and healthcare where accuracy and speed of decision-making are critical. AI-enabled solutions, particularly precision learning algorithms, offer numerous benefits, including (1) earlier disease detection, (2) more effective and less expensive treatment, (3) access to diagnosis in a variety of clinical settings, and (4) more accurate interpretation. For images, (5) the fastest diagnosis time. These benefits can be achieved in every specialty of medicine that generates a lot of visual data, from neuroscience to gastroenterology to dermatology. This thesis presents several proposed smart models to improve the ability to detect and predict diseases from data and images such as CT scans based on deep learning techniques and Optimization First, this thesis proposes an automatic malaria detection (MCNN) model based on CNN in this work to classify infected cases. MCNN focuses on detecting infected cells, which aids in the calculation of parasitemia in the blood or infection procedures. The weights were also adjusted, with the data set increased by some techniques to overcome a problem. Over fitting Using the malaria database prepared in Kaggle, the proposed model achieved 0.9929, 0.9848, 0.9859, 0.9924, 0.0152, 0.0141, 0.0071, 0.9890, 0.8994, and 0.9780 in terms of specificity, sensitivity, F1-score, and Matthew’s correlation coefficient, respectively. With 27,558 images, a comparison was made between the proposed model and some recent related works, as this comparison showed that the proposed model outperforms the comparative works in terms of evaluation measures. Second, this thesis proposes and implements an automatic Worried Deep Neural Network (WDNN) model based on both the Deep Learning Network (DNN) and Transfer Learning. Experimental results showed that the proposed WDNN model outperforms the rest of the related work that was used in this part of the thesis using three pre-training models : InceptionV3, ResNet50, and VGG19. Due to the lack of a COVID-19 dataset, data augmentation was used to increase the number of images in positive states, and then normalization was used to make all images the same size. Practical trials were conducted on the COVID-19 dataset collected from different cases, with a total of 2,623 as follows: 1,573 exercises, 524 evaluations, and 524 tests. To verify the correctness of the proposed model, The proposed model achieved 99.046, 98.684, 99.119, and 98.90 in terms of accuracy, precision, recall, and F-score, respectively. The results of the experiments indicated the ability of the proposed WDNN classification model to be used as an alternative to the current diagnostic tool. These results also showed that the proposed WDNN model outperforms the traditional machine learning methods and some of the CNN models used in the comparison process.
Third, this thesis presents a proposed automated prediction (CDNN) for COVID-19 using CT scan images using transfer learning and CNN algorithms. The COVID-X database used for clinical trials contained a set of 13,413 samples divided into two categories: 7,395 CT scan images of individuals with confirmed COVID-19 virus and 6,018 images of suspected cases. The Resnet model (50) has the best training results in specificity, precision, negative predictive value, false-positive rate, false-negative rate, accuracy, F1 Score, and Matthews Correlation Coefficient with values of 0.9880, 0.9892, 0.9891, 0.9882, and 0.0108, 0.0109, 0.0120, 0.9886, 0.9885, and 0.9772, respectively, using Optimizer SGD. Fourth: A Deep Skin Cancer Detection (DSC) model is proposed based on Knowledge Distillation technology and some optimizers. Experimental results based on the standard datasets used from ISIC show good performance in diagnosing melanoma. According to the results obtained from practical experiments, a sensitivity value of 99.16% was experimentally achieved, and a specificity value of 99.57% was achieved using the Adamax optimizer. Furthermore, the results help diagnose some cases of COVID-19 due to the similarity between skin cancer and the black fungus found in some COVID-19 survivors, particularly those with co-morbid conditions like skin cancer infection. Also, a comparison of several optimization algorithms known for their efficiency in improving computer vision algorithms was presented and applied. The International Skin Imaging Collaboration (ISIC) was used to compare these algorithms in detecting skin cancer.