Ph.D. Theses
Permanent URI for this collection
Browse
Browsing Ph.D. Theses by Author "Acar, Burak."
Now showing 1 - 3 of 3
Results Per Page
Sort Options
Item Content based medical image retrieval(Thesis (Ph.D.) - Bogazici University. Institute for Graduate Studies in Science and Engineering, 2017., 2017.) Marvasti, Neda Barzegar.; Acar, Burak.Fast technological developments of di erent medical imaging and data collection techniques increase the expectation of more accurate interpretations and diagnoses of radiologists. However, to carefully analyze the resulting big medical data, reliable and fast systems are needed. Content-based medical image retrieval (CBMIR) is a valuable technique to assist radiologists by identifying similar images in a large archive. However, due to the huge semantic gap between low-level image features and high-level semantic features, the challenge of retrieving similar images utilizing the high-level user speci ed semantic labels, which are closer to the users understandings and interpretations, has attracted great interest from various researches. In this dissertation, an iterative search and retrieval scheme to identify similar images from a database of 3-dimensional liver computed tomography (CT) images is proposed via utilizing the combination of lesion and liver related semantic features and patients' metadata. At each retrieval iteration, the lesion related concepts are annotated in a speci c order through a proposed computer aided medical image annotation (CMIA) scheme. The proposed radiologist-in-the-loop semi-automatic CMIA is based on a Bayesian tree structured model, linked to RadLex, to exploit the inter-dependencies between concepts to update the full annotation process and to guide the radiologist to input the most critical information at each iteration. Results show the e ectiveness of this modelbased interactive annotation scheme compared to the domain-blind models, as well as its advantage in the performance of the retrieval system, where a few number of manual annotations can signi cantly boost the retrieval accuracy. Moreover, better retrieval performance is achieved by incorporating a small contribution of the non-lesion data.Item Machine heath monitoring for cyber-physical systems(Thesis (Ph.D.) - Bogazici University. Institute for Graduate Studies in Science and Engineering, 2020., 2020.) Aydemir, Gürkan.; Acar, Burak.Estimating the failure time of the machinery that are used in the production is crucial to achieve an e cient maintenance in Industry 4:0 era. Remaining useful life (RUL) is the term that refers to the length of time in terms of the raw time intervals or usage that a machine will continue to operate before it requires a repair or replacement. Machine learning (ML), especially deep learning, provides industry practitioners with e cient tools for estimating the RUL. However, ML is far from being fully utilized, since domain knowledge is generally ignored in current studies. This thesis focuses on three main domain speci c problems in machine condition monitoring to improve the performance of ML based RUL estimation. First, RUL is ill-de ned during the healthy operation period of the machinery, hence enforcing ML with respect to a ctitious true RUL during these periods adversely a ects the overall RUL estimation accuracy. In this thesis, a system level anomaly detection triggered RUL estimation method is proposed to detect degradation onset point in sensor data to prevent ML models to estimate RUL in this period, and hence to increase the accuracy. Secondly, the operating conditions of the machines a ect their degradation pattern and related sensor measurements. Thus, the accuracy of ML based RUL estimation models decreases when the machinery operate in varying conditions. A siamese neural network based operating conditioninvariant feature extraction method is introduced to alleviate this problem. These two approaches are veri ed using a benchmark turbofan engine degradation data. Lastly, most of the ML models su er from lack of data in RUL estimation. If the data are high dimensional such as image, pro le, etc., the problem becomes more challenging. Two deep learning architectures are proposed to resolve curse of dimensionality in case of degradation data scarcity. E ciency of the proposed models is demonstrated with an infrared image data.Item Multi-modal tensor representations of brain networks(Thesis (Ph.D.) - Bogazici University. Institute for Graduate Studies in Science and Engineering, 2021., 2021.) Durusoy, Göktekin.; Acar, Burak.Considering the economic, social and psychological burdens of Alzheimer’s disease (AD), the most common form of dementia, it is essential to gain insight into the process and underlying mechanisms of the disease. Using structural and functional brain connectomes obtained by in-vivo MRI techniques as biomarkers is a promising approach. In this thesis, the B-Tensor structure that allows the representation of brain connectomes defined in structurally and functionally with a uni-modal and multi-modal fashion is presented. With the projection of structural connectomes onto known func tional networks, patients with AD and healthy control group are distinguished in a 7-dimensional space with a separation performance of over 90%. In addition, with the uni-modal and multi-modal tensor factorization methods, 47 patients with different levels of AD, are diagnosed with an accuracy of 77% - 100% in a 5-dimensional space. The results show that the multi-modal factorization technique performs better than the uni-modal one by successfully fusing the structural and functional networks which offer complementary information. While the neurological evaluations of the obtained sub-networks are highly consistent with previous literature, new findings regarding the progression of the disease are also recommended.