Doktora Tezleri
Permanent URI for this collectionhttps://hdl.handle.net/20.500.12573/5800
Browse
5 results
Search Results
Doctoral Thesis Derin Öğrenme Tabanlı Kompozit Malzemelerin Ultrasonik Tomografi Görüntülerinden Kusurların Tespiti ve Sınıflandırılması(Abdullah Gül Üniversitesi, Fen Bilimleri Enstitüsü, 2024) Gülşen, Abdulkadir; Güngör, Burcu; Kolukısa, BurakThis thesis introduces novel methodologies for enhancing defect classification and characterization in advanced composite materials by leveraging state-of-the-art machine learning (ML), deep learning (DL), and federated learning (FL) techniques within ultrasonic and acoustic emission (AE) inspection environments. First, a new ultrasonic dataset (UNDT), comprising 1,150 images from 60 distinct composite materials, is introduced. Applying transfer learning methods to both the UNDT and a publicly available dataset demonstrates the efficacy of advanced neural architectures—such as DenseNet121 and VGG19—achieving accuracy rates up to 98.8% and 98.6%, respectively. Next, the scope is extended to AE-based health monitoring by introducing an ensemble feature selection methodology to identify features strongly correlated with damage modes. By selecting amplitude and peak frequency for labeling and subsequently applying unsupervised clustering, the analysis confirms that both traditional AE features (e.g., counts and energy) and less commonly employed features (e.g., partial powers) correlate with distinct defect types. Finally, a novel FL framework is introduced to address the scarcity of publicly available, real-world ultrasonic datasets. This decentralized approach preserves data privacy while maintaining performance levels comparable to centralized methods, ensuring scalability and confidentiality in diverse data environments. Overall, these contributions significantly advance the field of NDT, offering robust defect classification and characterization. In doing so, the findings not only improve the accuracy and reliability of material integrity assessments but also lay a durable foundation for more secure, collaborative, and efficient NDT systems.Doctoral Thesis Protein Yapı Tahmini için Derin Öğrenme Modellerinin Geliştirilmesi(Abdullah Gül Üniversitesi, Fen Bilimleri Enstitüsü, 2022) Görmez, Yasin; Aydın, ZaferThe three-dimensional structure of a protein provides important clues about the function of that protein. Although there have been many studies on protein structure prediction, the problem has still not been solved completely. As it is very difficult to predict the three-dimensional structure of a protein directly, predictions of structural properties of proteins such as secondary structure, solvent accessibility, and torsion angles are carried out first, which are later used as inputs to more elaborate structure estimation tasks. In this thesis, novel deep learning models have been developed by using convolutional neural networks (CNN), graph convolutional networks (GCN) and long-short-term memory (LSTM) recurrent neural networks to predict secondary structure, solvent accessibility and torsion angles of proteins. A rich feature set formed by using PSI-BLAST, HHBlits, physicochemical properties, structural profile matrices, AA index values, and graphs representing the relationship between amino acids were used as inputs to the models. In the first study, a deep learning model was developed by using CNN and GCN layers for secondary structure prediction. In the second study, LSTM layers were added to the first model, which was extended to make solvent accessibility and torsion angle predictions as well using the multi-task learning approach. In both studies, graphs were generated using neighborhood relations between amino acids. In the last study, a novel U-net-based model was designed for secondary structure prediction using CNN, GCN, and LSTM layers. The graph matrices used as input to GCN layers were obtained by using protein contact map prediction. All models were trained, optimized and tested on benchmark data sets. Improvements were obtained in accuracy as compared to the state-of-the-artDoctoral Thesis FDG-PET Görüntülerindeki Tümörlerin Makine ve Derin Öğrenme Tabanlı Analizi(Abdullah Gül Üniversitesi, Fen Bilimleri Enstitüsü, 2022) Ayyıldız, Oğuzhan; Yılmaz, BülentAnalysis of a tumor is essential in treatment planning and evaluation of treatment response. Positron Emission Tomography (PET) is a vital imaging device for clinical oncology in understanding the metabolic structure of the tumor. In this thesis, three separate studies investigating the application of machine, deep learning and statistical approaches on FDG-PET images from patients with non-small cell lung cancer (NSCLC) and pancreatic cancer. The first study aimed at performing a survey on subtype classification of NSCLC by using different texture features, feature selection methods and classifiers. Images from 92 patients and several clinical and metabolic features for each case were used in this study along with histopathological validation for the tumor subtype labeling. Stacking classifier resulted in 76% accuracy. The aim of our second study was to adapt an atrous (dilated) convolution-based tumor segmentation approach (DeepLabV3) on FDG-PET slices with maximum standard uptake value (SUVmax). MobileNet-v2 pretrained on ImageNet served as the backbone to DeepLabV3. The classification layer was interchanged with the Tversky loss layer which helped improve model's performance while the dataset was imbalanced. Images from 141 patients were employed and augmentation was performed in each training phase. Dice similarity index was obtained as 0.76 without preprocessing and 0.85 with preprocessing. The last study focused on determining the features to be used in the prognosis of pancreatic adenocarcinoma on FDG-PET images from 72 patients. Well-known texture, metabolic and physical features were extracted from tumor region that was determined with the help of random walk segmentation algorithm. On these features time-dependent ROC curve analysis was performed for 2-year overall survival (OS) prediction, and, in the univariable analyses, tumor size, energy, entropy, and strength were found to be significant predictors of OS. Keywords: PET/CT, NSCLC, Machine learning, Deep learning, Radiomics, Semantic segmentationDoctoral Thesis Zamansal Bilgiden Faydalanarak Videodan Orman Yangınlarının Erken Tespiti(Abdullah Gül Üniversitesi, Fen Bilimleri Enstitüsü, 2022) Taş, Merve; Taşdemir, Kasım; Aydın, ZaferForest fires are considered as the major threats to lives, properties and to the integrity of the ecosystem around the world. In most cases, the fire damage can be reduced, when the initial signs of the fire are detected in a timely manner. Since smoke is considered as the first visual sign of fire, detection of smoke is vital. Hence, a successfully designed smoke detection system is essentially critical in the early detection of smoke for outdoor environments. The existing smoke detection methods suffer from high false alarm rates and cannot accurately detect smoke in hazy environments. To address these problems, this thesis is focused on smoke detection model at an early stage that utilizes deep learning (DL) based techniques for outdoor locations. This work contributes mainly to four aspects of smoke detection: (1) new datasets preparation for three smoke detection tasks classification, detection-segmentation, and video classification, (2) utilizing transfer learning to detect the smoke on the relatively small dataset, (3) image dehazing process that includes removing the haze from the dataset images to enhance the system performance, (4) designing a novel hybrid video classification model by combining the two DL based video classification structures. This work will be a resourceful reference for researchers working in the fields of forest fire or smoke detection studies at an early stage. The experiments, research findings, and enhanced performance of the smoke detection system provide a source of information about smoke detection. Current studies can be utilized to further improve the design of efficient and reliable fire safety models. Keywords: Deep Learning, Spatio-Temporal Information, Forest Fire Early Detection, Smoke Detection, Image DehazingDoctoral Thesis Medikal Termal Görüntülerin Otomatik Olarak İşlenmesi ve Sınıflandırılması(Abdullah Gül Üniversitesi, Fen Bilimleri Enstitüsü, 2022) Özdil, Ahmet; Yılmaz, BülentThe aim of this dissertation is to develop computer aided methods for processing and evaluating medical infrared thermal images. Throughout this study three problems were evaluated. The first problem was to automatically classify the body part and pose in the thermal images. In this study there were four classes; upper-lower body parts with back-front views. The first step included the segmentation of the background with Otsu's thresholding method applying histogram equalization. Next, DarkNet-19 architecture was used to extract features from images and these features were reduced using PCA and t-SNE methods. Finally reduced feature sets were used for classification. The second problem was to automatically classify liver steatosis from using thermal images. In this study, the classification problem was tested on an anatomical region of interest from abdominal images corresponding to the liver. Deep learning and texture analysis methods were employed for feature extraction, and then the selected feature sets were used for classification. The third problem was to quantify thermograms of multiple sclerosis (MS) patients for better assessment of the disease and monitoring the therapy. Thermal images of two patients and a healthy control from lower limbs were evaluated during experiments, and localized quantification of the effect of MS on the feet of the patients using thermal images method was proposed. The proposed method was fully correlated with the evaluations of physician. It is shown that medical thermal imaging has high potential in many fields of medicine as a non-invasive method for pre-diagnosis and follow-up.