Doctor of Philosophy (PhD)


Division of Computer Science & Engineering

Document Type



Image processing using machine learning (ML) are new technologies in The Kingdom of Saudi Arabia, which is working to reach the highest level in this technology. Saudi Arabia has begun the most ambitious and far-reaching reform plan in the country’s history, guided by the “Vision 2030” initiative where AI is key to solving problems of national interest. Image processing has recently developed into a crucial tool in several industries, including remote sensing and medicine. Focusing on their uses in the analysis of satellite data and medical imaging, this dissertation provides a framework for ML-based image processing for satellite-driven products and medical imagery. The objective is to evaluate the benefits, drawbacks, and practical applications of these two opposing techniques in order to inform choices made in actual situations. Traditionally, image understanding has mostly been done manually in both satellite imagery and medical imaging domains \cite{li}. Human analysts that are skilled at image interpretation and information extraction are used in this process. Flexible and adaptable in complicated settings, the human brain may infer patterns and traits that automated approaches would find challenging. However, it is labor- and time-intensive and susceptible to biases and human error. In the first chapter (environment part), predicting sand shift roads is a key problem of road topography in Saudi Arabia. In this work, we present a graph-based framework to predict sand-shift road segments and connections between them based on high-resolution satellite imagery. The framework uses a graph-theoretic framework wherein the nodes are points on existing sand-shift roads. Two “adjacent” nodes are joined by an edge if there exists a sand-shift road segment connecting them. The attributes of the nodes can be road-types (as determined by Open-StreetMap), speed limits (as available from the Saudi Arabian government). A graph-embedding based approach is used to predict new sand-shift road segments as well as connections between them. To determine from a high-resolution image if these two ``adjacent” nodes are connected or not, we will use: Node Embedding of Hypercolumn (NEH); Node Embedding of Road Type (NERT); Node Embedding of Speed Limit (NESL); Node Embedding of Hypercolumn and Road Type (NEHRT); Node Embedding of Hypercolumn and Speed Limit (NEHSL); Node Embedding of Road Type and Speed Limit (NERTSL); and Node Embedding of Hypercolumn and Road Type and Speed Limit(NEHRTSL)(Jure et al. 2018). Also, dealing with the Saudi Arabia desert road via AI is a new research area which means we will be proactive in this field. The second part of this study examines methods for automatically understanding medical imagery using artificial intelligence algorithms. A crucial and crippling aspect of chronic lung disorders is pulmonary fibrosis (lung cancer), which is characterized by an abnormal accumulation of collagen in the lungs. To better comprehend lung fibrosis, methods are needed for the direct observation of fibrillar collagen across the whole lung (in this study, we use the mouse model). Parametric second-harmonic generation (P-SHG) microscopy is an imaging framework for visualizing the structural modification of collagen in utero Air (control), In utero SHS (Baseline), In utero Air + Urethane (Cancer Treatment\_03), and utero SHS + Urethane (Cancer Treatment\_04). Our experiment was on old female mice (58 weeks old). In full murine lungs, we were able to see the whole network of fibrillar collagen using the P-SHG. The segmentation and feature extraction required for current lung tissue grading by computer-assisted image analytics usually relies on domain expertise in lung tissue pathology, and hence has not been automated. Through learning, from a huge collection of images, forty-six parameters of fibrillar collagen, we may be able to identify the amount of collagen in these images without the requirement for prepossessing. Also, we want to predict treatment responsiveness. To evaluate lung tissue, we explored how well chart models (K-means, T-SNE and UMAP) are created using image analysis and imaging informatics of the fibrillar collagen when compared to more traditional approaches. We demonstrate that Image analysis and imaging informatics of the fibrillar collagen may be used to create a prediction model for scoring the different stages of lung tissue fibrosis that is more accurate traditional methods. In this study, a new, powerful technique combining Artificial Intelligence and P-SHG imaging is introduced for 2D understanding of lung fibrosis with macroscopic imagery of lung disease. This technique also gives a new way to investigate the whole lung without the need for biased regional sampling. Using image analysis and imaging informatics of fibrillar collagen characteristics of the lung tissue may also be used for diagnosis of cancer. Our framwork enables understanding the processes driving lung tissue fibrosis and discovering viable therapeutics for this malignancy as well as predicting therapy responsiveness. In the future, we will keep examining more images, and we will use our framework on different gender and age (e.g., Male Mice Group and Younger Female Mice Group). This study seeks to offer useful insights into the efficacy of manual analysis and machine learning in satellite images and medical imaging applications through a comparative examination of case studies and experimental findings. It emphasizes the circumstances in which one methodology may perform better than the other as well as circumstances in which a mix of both approaches may produce better results. The results of this study should help decision-makers select the best appropriate image processing strategy for their unique requirements, resulting in increased accuracy and efficiency in a variety of industries that rely on image interpretation.



Committee Chair

Supratik Mukhopadhyay

Available for download on Tuesday, January 19, 2027