İstatistik Bölümüİstatistik Bölümühttps://hdl.handle.net/11655/202024-03-29T05:26:41Z2024-03-29T05:26:41ZTümörlü Beyin Hücreleri Tespitinde Öğrenme Aktarımıyla Derin Sinir Ağlarının UygulanmasıYağmur, Berkehttps://hdl.handle.net/11655/343352023-12-21T06:28:57Z2023-07-14T00:00:00ZTümörlü Beyin Hücreleri Tespitinde Öğrenme Aktarımıyla Derin Sinir Ağlarının Uygulanması
Yağmur, Berke
This thesis examines the effectiveness and applicability of deep learning models in medical image analysis. Specifically, it focuses on the classification of magnetic resonance images (MRI) containing tumors of various types. In the introduction section, the importance of medical image analysis, the rise of artificial intelligence and deep learning in this field, and the objective of this study are discussed in detail.
In the thesis, four popular deep learning models, namely ResNet, MobileNet, DenseNet, and InceptionV4, are utilized. Each model is trained using the transfer learning method with ImageNet weights and trained from scratch without ImageNet weights. The performance of the models is evaluated based on the accuracy of classifying various types of tumors.
Although the results are generally evalueted across all metrics, the final evaluation is done based on the F1-score due to the imbalanced distribution of the classes. It is observed that the InceptionV4 model has the highest overall success rate (%96 F1-score) when trained with ImageNet weights. While other models also exhibit comparably high accuracy rates, they face challenges in classifying certain types of tumors. Notably, the performance of all models dropped in certain classes such as Granuloma T2.
Additionally, the thesis emphasizes that transfer learning aids in the efficient and effective training of models. Moreover, it is noted that conducting model training for specific datasets and adjusting hyperparameters could enhance the performance of the models.
Ultimately, the thesis provides valuable insights into which deep learning model or models are most suitable for tumor classification in clinical applications, while understanding the limitations and challenges of these models can guide future research and developments.
This study highlights the importance of deep learning in medical image analysis and contributes to the field by paving the way for more accurate diagnoses and effective treatments for patients.
2023-07-14T00:00:00ZÇok Durumlu Modellerde Geçiş Olasılıklarının TahminiÇiftçi, Esrahttps://hdl.handle.net/11655/343182023-12-25T07:43:16Z2023-06-01T00:00:00ZÇok Durumlu Modellerde Geçiş Olasılıklarının Tahmini
Çiftçi, Esra
The time until an event occurs is examined in survival analysis. During the time that an individual or a unit under examination passes from the starting state to the final state, other situations may occur and transitions may take between them. In multi-state models, intermediate states are involved until the final state is. Multi-state models are an extended version of the two-state analysis models. The most commonly used model in the literature is the three-state disease-death model, which includes a transitional state between the first and last states.
This study introduces the multi-state models and their general model structures. The estimation approach proposed by Zaman et al. (2012) for transition probability and standard error estimates were adapted to a three-state illness-death model. Probability and standard error adjustments were made to the Aalen-Johansen (AJ), Presmoothed Aalen-Johansen (PAJ), Lifetime Data Analysis (LIDA), Landmark (LM), Presmoothed Landmark (PLM), Landmark Aalen-Johansen (LMAJ) and Presmoothed Landmark Aalen-Johansen (PLMAJ) estimators, and probability (p_11, p_12, p_13, p_22 ve p_23) and standard error values were calculated for all transitions between states. In addition, Fleming Harrington (FH) estimator used in two-state survival analysis was adapted to the three-state illness-death models, and adjustments were made to the FH estimators. Transition probabilities and standard error values were calculated for time points during the process for all estimators, and graphs were drawn for some s points determined for p_ij(s,t) and standard errors. The variability among estimators was demonstrated by comparing the results of AJ, PAJ, LIDA, LM, PLM, LMAJ, and PLMAJ, and the FH estimators calculated using these estimators.
Colon cancer, heart transplant, breast cancer and cervical cancer data sets have been studied to demonstrate the applicability of the methods. Lower standard error values have been obtained for the adjusted/corrected estimators for AJ, PAJ, LIDA, LM, PLM, LMAJ, and PLMAJ. While FH and adjusted FH estimators generally have higher standard error values than other estimators, the adjusted FH estimators have provided lower standard error values for p_11 in breast cancer data and p_23 in cervical cancer data. Better results were obtained with the adjusted estimators for all data sets.
2023-06-01T00:00:00ZEvrişimsel Sinir Ağları Kullanarak Sondaj Karot Sandıklarının İncelenmesi ve Tenör TahminiÇınar, Haydarhttps://hdl.handle.net/11655/343072023-12-26T11:20:02Z2023-01-01T00:00:00ZEvrişimsel Sinir Ağları Kullanarak Sondaj Karot Sandıklarının İncelenmesi ve Tenör Tahmini
Çınar, Haydar
Mining companies have allocated approximately 10 billion USD for the exploration and investigation of new mines in 2018. Drilling has been a long-standing method for the exploration of underground mineral deposits. By using core samples obtained from drilling, information about the type, value, and quantity of minerals composing the rock can be obtained. In the mining industry, resource estimation and mineral analysis are of critical importance for improving operational efficiency, reducing costs, and ensuring sustainability.
This study focuses on the examination of drilling core trays and ore grade estimation using Convolutional Neural Networks. An image segmentation and ore grade estimation application using the U-Net model has been developed for the evaluation of iron ore resources. The data used in this study were obtained from a database consisting of images extracted from drilling core trays. During the training process, U-Net models were trained to detect core samples and ore-bearing regions on the samples. The developed application performs ore grade calculations based on the core images and presents them as output.
The aim of this study is to highlight the potential of modern data analysis techniques in the mining sector by providing an alternative method in the field of resource estimation and data analysis. It is expected that the findings of this thesis will contribute significantly, primarily to our country's mining industry, and inspire future research in the development of resource estimation techniques and the enhancement of operational efficiency.
2023-01-01T00:00:00ZMakine Öğrenme Yaklaşımlarının Biyoinformatikte İlaç Geliştirme Probleminde KullanılmasıSemerci, Tuğçehttps://hdl.handle.net/11655/342912023-12-29T06:16:47Z2023-01-01T00:00:00ZMakine Öğrenme Yaklaşımlarının Biyoinformatikte İlaç Geliştirme Probleminde Kullanılması
Semerci, Tuğçe
Humans are at the center of the drug research and development process. It is aimed to
help the patient overcome his illness and improve his quality of life. In the drug
development process, innovative drugs are aimed to be effective, reliable and
treatments that will be offered to patients as soon as possible. However, the discovery
of a drug and putting it into the service of medicine requires time consuming and high
cost. In recent years, thanks to the development of information technologies and
bioinformatics-based applications, progress has been made in moving this process to
the clinical stage with less cost and quickly. In this thesis, it is aimed to detect
molecules that can be drug candidates for the treatment of Type-2 diabetes by using
DPP-4 inhibitors and with the help of machine learning approaches. The data obtained
from the ChEMBL database were analyzed with 10 machine learning algorithms and
artificial neural network model. In comparison of the performances of the models, the
Root Mean Square Error (RMSE) criteria were evaluated. As a result of the
application, it has been seen that the machine learning approaches that produce the
best predictions are Random Forest and a single layer feedforward neural network. It
has been observed that these two methods give predictive results close to each other. In the evaluation of the performances of the models, the Random Forest model was
chosen as the optimum model because it showed higher performance than the root
mean square error value, which is the most common criterion in the literature.
According to the results of this study, it has been seen that using the Random Forest
approach produces good results in detecting molecules that can be drug candidates for
the treatment of Type-2 diabetes.
2023-01-01T00:00:00Z