Tez No	İndirme	Tez Künye	Durumu
797497		Holografik mikroskopinin hücresel analiz ve veri güvenliği performansının iyileştirilmesi / Improvement of cellular analysis and data security performance of holographic microscopy Yazar:MUHAMMED ALİ PALA Danışman: DOÇ. DR. MUSTAFA ZAHİD YILDIZ Yer Bilgisi: Sakarya Uygulamalı Bilimler Üniversitesi / Lisansüstü Eğitim Enstitüsü / Elektrik-Elektronik Mühendisliği Ana Bilim Dalı Konu:Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve Kontrol = Computer Engineering and Computer Science and Control ; Biyoteknoloji = Biotechnology ; Elektrik ve Elektronik Mühendisliği = Electrical and Electronics Engineering Dizin:	Onaylandı Doktora Türkçe 2023 207 s.
Mikro dünyadaki meydana gelen etkileşimler, doğada büyük değişimleri meydana getirmektedir. Geleneksel optik mikroskoplar, bu işleyişteki biyofiziksel süreçlerin anlaşılmasında önemli araçlardır. Yıllar içinde optik mikroskoplardaki gelişmeler, mikroskobik yapıların daha ayrıntılı olarak görselleştirilmesine olanak sağlasa da temel çalışma prensiplerinde bir değişim olmamıştır. Merceksiz Dijital Eksen Üstü Holografik Mikroskopi (DIHM), geleneksel mikroskoplardaki optik bileşenlerin görevini algoritmaların üstlendiği hesaplamalı bir mikroskopi tekniğidir. Bu sistemler düşük maliyet, geniş görüş alanı ve kolay entegrasyon gibi avantajları nedeniyle in-vitro ortamlarında gerçekleştirilen hücresel uygulamalarda geniş kullanım potansiyeline sahiptir. Yüksek seviyede bilgi içeren verilerin kaydedilmesini sağlayan bu sistemlerin başarısı, görüntüleme ve veri analizi süreçlerinin birlikte yürütülmesine bağlıdır. Fakat mevcut veri işleme yöntemlerinin kısıtlamaları, DIHM ile elde edilen görüntülerin analizinde önemli kısıtlamalara ve ayrıca verilerin güvenliğinin sağlanmasında eksikliklere neden olmaktadır. Bu tez çalışmasında, DIHM'nin in-vitro hücre deneylerinde analiz başarısının iyileştirilmesi ve veri güvenliğinin sağlanması amaçlanmıştır. Bu doğrultuda bir merceksiz mikroskopi sistemi tasarlanarak sistemin karakterizasyonu standart bir çözünürlük hedefiyle ortaya konulmuştur. Hologram görüntülerin analizi ve güvenliği için üç temel uygulama alanı üzerinde durulmuştur. İlk olarak, hücre slaytlarında bulunan hücrelerin sayım ve canlılık analizlerini yüksek hassasiyette yapabilmek amacıyla geleneksel görüntü işleme yöntemleriyle desteklenmiş, Evrişimli Sinir Ağları (CNN) tabanlı hibrit bir yöntem geliştirilmiştir. Örtüşen hücreler ve hücre dışı partiküller, sayım ve canlılık analizlerine dahil edilerek analiz hassasiyeti iyileştirilmiştir. İkinci olarak, hücrelerin bireysel olarak canlılık aktivitesinin belirlenmesi için fraktal boyut tabanlı bir öznitelik hesaplama tekniği geliştirilmiştir. Önerilen yöntem, hücrelerin görüntü içinde bulunduğu konum ve rotasyondan bağımsız olarak, tüm hücre morfolojisinin tek bir öznitelikte kimliklendirilmesine olanak sağlamaktadır. Canlılık aktivitesinin sınıflandırılması amacıyla Derin Sinir Ağları (DNN) kullanılmıştır. Son olarak, yüksek çözünürlüklü hologram verileri için güvenlik, sağlamlık ve algılanamazlık açısından yüksek performans gösteren yeni bir dijital damgalama yöntemi geliştirilmiştir. Rastgele sayı üreteçleriyle güvenlik kriteri sağlanarak, frekans uzayında hibrit bir damgalama şeması önerilmiştir. Tez kapsamında literatüre kazandırılan yöntemler, veri analizi ve güvenliği alanlarına oldukça değerli katkılar sağlamaktadır.
Interactions occurring in the micro world cause significant changes in nature. Conventional optical microscopes are essential for understanding this process's biophysical processes. Advances in optical microscopes have allowed the visualization of microscopic structures in more detail, but the basic working principles of microscopes have not changed. Due to the inherent limitations of lenses, which are the core components of conventional microscopes, these systems consist of highly complex and costly optical components. In recent years, developments in production technologies and computational capabilities have brought along innovative approaches in the field of microscopy. Lens-free digital in-line holographic microscopy (DIHM) is a computational microscopy technique in which algorithms take over the role of optical components in conventional microscopes. These systems have the potential to be used in cellular applications in in-vitro environments due to their advantages, such as low cost, wide field of view, and easy integration. The success of these systems, which enable the acquisition of high-level informational digital data, depends on the combination of imaging and data analysis processes. However, the most important obstacle to the widespread use of these systems in clinical and laboratory settings is the limitations of current analysis methods and data security. This thesis aims to improve the analysis success and ensure data security in in-vitro cell experiments of DIHM. Within the scope of the thesis, three application areas are mainly focused on. Methods have been developed for high-throughput analysis of cell line count and viability, high-accuracy determination of individual viability activity of cells, and secure watermarking of high-resolution digital holograms. In this direction, deep learning methods and image processing methods were used as tools. In the second part of the thesis, holography and deep learning methods were introduced and focused on the basic concepts. The methods used for forming, recording, and reconstructing diffraction patterns are explained mathematically. The criteria for the theoretical determination of the imaging area, axial and lateral resolution values in DIHM are included. Afterward, the hierarchical structure starting from the single-cell neuron model, which is the smallest processing unit of deep neural networks, and extending to multi-layered models is mentioned. Other machine learning methods used in the thesis are mentioned, and performance measurement metrics of classifiers are mentioned. The methods found in the literature regarding the use of deep learning in DIHM are mentioned, and the potential of the combination of the two fields is mentioned. In the third part of the thesis, the design parameters of the microscopy system are detailed and the proposed Convolutional Neural Networks (CNN) based hybrid method for high precision analysis of counting and viability rates of cancer cell lines is mentioned. CMOS imaging sensor and LED light source are used in the DIHM design. A microcomputer is integrated into the system to communicate these components. The angular spectrum method was used to reconstruct the holograms recorded in-line configuration. It has been demonstrated using the standard USAF 1951 resolution target that the designed system has a sufficient practical resolution for cellular imaging and analysis. The linear intensity and statistical analysis of reconstructed images of the resolution target demonstrate the characterization of DIHM. MCF-7 cell lines were grown under appropriate incubation conditions and stained with trypan blue protocol for viability analysis. Cell lines were injected into a glass at different concentrations, and their holograms were recorded with the designed system. Circular Hough Transform (CHT) was used to detect particles in the cell slide. By keeping the search space of CHT wide, it is possible to detect overlapping cells and cell-like non-cell particles. The images of the candidate particles whose positions were determined by CHT were segmented and divided into sub-images. CNN was used to classify the detected candidate particles. The topology and parameters of the designed model are detailed, and the model's classification performance is shown with various performance metrics with training, validation, and test data. At the output of the CNN model, the candidate particles were classified as living cells, dead cells, and non-cell particles. The class information of the cells obtained at the output of the CNN model is shown with the bounding box at the locations determined in the original image. According to the classification result, viability analyzes were performed with high sensitivity by determining the number of live and dead cells and the total number of cells. A graphic processor-based user interface has been designed in order to perform end-to-end, fast, and automatic imaging and analysis processes. The success of the proposed method in cell counting and viability analysis was compared with the data obtained from the ground truth system. It has been revealed that the proposed method can perform human-level analysis. In the fourth part of the thesis, a fractal dimension-based feature calculation technique was developed to determine the viability activity of cells. With the proposed methods, it is ensured that the complexity and self-similarity in the morphology of the cells can be evaluated as a criterion for determining the viability activity. In the proposed method, to determine the cells' viability regardless of their location and angle, each grey-level cell image was rotated at different angles, and their fractal dimensions were calculated. The image taken from each cell was rotated at different angles in this way, and 19 different fractal dimension values for each cell were calculated with the box-counting algorithm. The statistical significance of the method was revealed by finding the p-value of each feature as <0.05. Deep Neural Networks (DNN) and other machine learning methods were used to classify fractal features calculated with the proposed method. The success of the designed DNN model in classification during training and evaluation has been demonstrated by various performance metrics. It has been shown that the proposed method is independent of segmentation and other image pre-processing steps and can identify cells in a way that makes sense of the entire morphology of the cell. In addition, the proposed fractal dimension-based feature calculation method provides high accuracy in determining cell viability activity by eliminating the problems such as twin image and DC term that occur after the holographic reconstruction process. In the proposed method, fractal features were shown to be like a fingerprint in cell viability analyses, and it was seen that they gave reliable results in cell viability analyses. The presented method can be utilized in a wide range of laboratory applications in cancer research. Finally, for the security of high-resolution holograms and associated identifying data recorded with DIHM, a watermarking method has been developed that provides high imperceptibility and robustness. DIHM enables microscopic imaging at the sub-micron resolution, giga-pixel size. Ensuring the security of the data and related analysis results obtained from these systems, which have powerful capabilities in imaging and analysis processes, is necessary. Especially since holograms have a very high-resolution compared to other medical data, it becomes difficult to archive and transmit holograms securely and prevent illegal access. In order to overcome these problems, a new digital watermarking method has been developed with high performance in terms of security, robustness, and imperceptibility of high-resolution hologram data. The proposed method uses chaos-based random number generators to determine the watermark location and provide security. The suitability of the chaotic system for use in studies such as encryption, data hiding, and watermarking has been proven by dynamic analysis. NIST-800-22 and ENT statistical tests, which have the highest international standards, were used to measure the randomness performance of the generated numbers. A hybrid method has been developed for digital watermarking using Discrete Wavelet Transform (DWT) and Singular Value Decomposition (SVD). The application of the proposed watermarking method has been demonstrated on cancer cell lines and blood samples widely analyzed in the laboratory and clinic. The performance of the proposed new watermarking method has been proven by robustness and invisibility tests. The results show that the proposed method ensures the security of the data without affecting the image quality. Although the proposed method has been applied to medical images obtained in both clinical and laboratory conditions, it has the potential to be applied to many different high-resolution data.