Pakize Erdoğmuş

Convolutional Neural Networks (CNNs) used for image classification often have complex architectures involving large images, time-costly training processes, and a large number of layers and hyperparameters. Therefore, improving the accuracy of CNN is a challenging process that requires time, resources and specialized knowledge. In this study, to improve the performance of CNN models, experiments were conducted on the MNIST, EMNIST, and Fashion-MNIST datasets using different optimization algorithms and a loss function (Si-CL) from the literature. The findings of the study reveal the effects of loss functions and optimization algorithms on model performance in detail. The SGDM, Adam, RMSProp, RMSProp, AdaMax, AdaDelta and AdaGrad optimization algorithms used during the experiments are examined and the results show that the Adam algorithm performs the best in terms of both training accuracy and test accuracy. The SGDM algorithm was particularly effective at larger batch sizes and low learning rates, but required longer training times compared to the Adam algorithm. The Si-CL loss function used in the study performed better than the traditional cross entropy loss. The model trained with the Si-CL loss function achieved higher results in terms of both training and test accuracy, reduced training time and lower loss value. This allowed the model to learn faster and more efficiently.

Alzheimer's disease (AD) is a neurodegenerative process that leads to impairments in fine motor skills as well as cognitive decline. In this study, the analysis of handwriting kinematics is aimed as a non-invasive early diagnosis method. The DARWIN (Diagnosis AlzheimeR With handwritINg) dataset located in the UCI Machine Learning Repository was used in the study. Classification performance was evaluated using Logistic Regression (LR), Random Forest (RF), and Support Vector Machines (SVM) algorithms. The study methodology consists of two stages: In the first stage, models were tested using the 3, 5, 10-Fold Cross-Validation method with the entire feature set. In the second stage, SHAP (SHapley Additive exPlanations) based feature selection was applied to increase model explainability and overcome the high dimensionality problem, and the "Best Fold" training strategy was adopted. The obtained findings indicate that handwriting analysis is a strong biomarker in AD detection and SHAP-supported optimization significantly increases model success.

Detecting Alzheimer’s Disease (AD) at an early stage is vital because it enables prompt treatment and intervention, which can help slow disease progression and enhance patient prognosis. Given the increasing prevalence of AD globally, with an estimated 50 million people currently living with the condition and projected to triple by 2050, the development of accurate and efficient diagnostic tools is paramount. In this study, a novel architecture for the early diagnosis of AD by combining Convolutional Neural Networks (CNNs) or Vision Transformers (ViTs) with traditional Machine Learning (ML) algorithms was proposed. Utilizing MRI images as input, CNNs/ViTs serve as feature extractors, while demographic data is integrated to enhance diagnostic accuracy. Through extensive experimentation, our proposed model, which utilizes a CNN backbone optimized for MRI analysis as a feature extractor and LGBM as the classifier, achieved superior accuracy, reaching up to 96.83%. Statistical validation through confidence intervals and McNemar’s test further demonstrated the robustness and significant performance improvements of the proposed model compared to baseline methods. This study employs eXplainable AI techniques to visualize critical regions in MRI images that influence the model’s diagnostic decisions, promoting clinical transparency and trust in AI-assisted early diagnosis of AD. The novelty of this study lies in integrating deep feature extractors (CNNs/ViTs) with traditional ML classifiers, supported by interpretability through Grad-CAM and statistical validation, offering a transparent and accurate framework for early diagnosis of AD.

Face detection is a crucial component in many AI-driven applications such as surveillance, biometric authentication, and human-computer interaction. However, real-world conditions like low-resolution imagery present significant challenges that degrade detection performance. In this study, we systematically investigate the impact of input resolution on the accuracy and robustness of three prominent deep learning-based face detectors: YOLOv11, YOLOv12, and MTCNN. Using the WIDER FACE dataset, we conduct extensive evaluations across multiple image resolutions (160x160, 320x320, and 640x640) and assess each model's performance using metrics such as precision, recall, mAP50, mAP50-95, and inference time. Results indicate that YOLOv11 outperforms YOLOv12 and MTCNN in terms of detection accuracy, especially at higher resolutions, while YOLOv12 exhibits slightly better recall. MTCNN, although competitive in landmark localization, lags in real-time inference speed. Our findings provide actionable insights for selecting resolution-aware face detection models suitable for varying operational constraints.

Alzheimer’s Disease (AD) is one of the most, if not the most, devastating neurodegenerative diseases that are incurable and progressive. Early diagnosis of AD comes with many promises in terms of medicine, sociology, and economics. Despite the existence of numerous studies that aim for early diagnosis of AD, to the best of our knowledge, there is not a publicly available tool that lets end-users assess AD. To address this gap, we propose a Graphical User Interface (GUI) powered by Machine Learning (ML) that makes self-assessment of AD possible – without any input from medical experts. The developed GUI lets end-users enter various information considering both commonly used features for the diagnosis of AD and the questions available in the gold standard screening tool for the diagnosis of AD, namely the Mini-Mental State Exam. In addition to employing 11 traditional ML algorithms, we propose a novel 1-dimensional (1D) Convolutional Neural Network (CNN). All ML models were trained on a gold standard dataset that comprised 373 records from three subject classes as follows: (i) non-demented, (ii) demented, and (iii) converted. Once the end-user enters the required input through the developed GUI, the previously trained ML model assesses the diagnosis of AD through this input in a couple of seconds. According to the experimental results, the proposed novel 1D CNN outperformed the state-of-the-art by obtaining an accuracy as high as 95,3% on the used gold standard dataset.

In histopathological image analysis, the detection and count of mitotic cells are important biomarkers for determining the degree and aggressiveness of cancer prognosis. Manual detection of mitosis by pathologists is a lengthy and challenging process. With advancements in deep learning architectures, numerous automatic mitotic detection methods have been proposed. However, most mitotic detection methods have weak generalizability across image areas and lack reproducibility and validation in multicenter environments. To overcome these issues, a new automatic mitotic detection approach called the Mi-DETR, based on the DETR architecture, has been proposed. In the proposed Mi-DETR model, the backbone of the original DETR is replaced by CSPResNeXt. The aim of this is to strengthen the learning capacity in feature extraction and increase the variability of the learned features. In this way, information loss and unwanted gradient flow are avoided. In the decoder layer, unnecessary model parameters have been filtered out using a layer reduction strategy to improve model efficiency and reduce computational costs. Additionally, a more stable model has been obtained by using the CIoU loss function instead of the L1+GIoU loss function used in the DETR model. The publicly available ICPR14 and TUPAC16 breast histopathology datasets were used for training, validation, and testing in the experiments. The results provided more precise and compact bounding boxes close to clinically validated ground truth, demonstrating the accuracy and generalizability of the proposed model. As a result, the proposed Mi-DETR model achieved a 0.921 F1-Score on the ICPR14 dataset and a 0.950 F1-Score on the TUPAC16 dataset. The results obtained on both datasets demonstrate that the proposed model performs well enough to compete with state-of-the-art deep learning architectures.

While there are many verification studies on signature images using deep learning algorithms in the literature, there is a lack of studies on the classification of signature images. Signatures are used as a means of identification for banking, security controls, symmetry, certificates, and contracts. In this study, the aim was to design network architectures that work very fast in areas that require only signature images. For this purpose, a new Si-CNN network architecture with existing layers was designed. Afterwards, a new loss function and layer (Si-CL), a novel architecture using Si-CL as classification layer in Si-CNN to increase the performance of this architecture, was designed. This architecture was called Si-CNN+NC (New Classification). Si-CNN and Si-CNN+NC were trained with two datasets. The first dataset which was used for training is the “C-Signatures” (Classification Signatures) dataset, which was created to test these networks. The second dataset is the “Cedar” dataset, which is a benchmark dataset. The number of classes and sample numbers in the two datasets are symmetrical with each other. To compare the performance of the trained networks, four of the most well-known pre-trained networks, GoogleNet, DenseNet201, Inceptionv3, and ResNet50, were also trained with the two datasets with transfer learning. The findings of the study showed that the proposed network models can learn features from two different handwritten signature images and achieve higher accuracy than other benchmark models. The test success of the trained networks showed that the Si-CNN+NC network outperforms the others, in terms of both accuracy and speed. Finally, Si-CNN and Si-CNN+NC networks were trained with the gold standard dataset MNIST and showed superior performance. Due to its superior performance, Si-CNN and Si-CNN+NC can be used by signature experts as an aid in a variety of applications, including criminal detection and forgery.

Human handwriting is used to investigate human characteristics in various applications, including but not limited to biometric authentication, personality profiling, historical document analysis, and forensic investigations. Gender is one of the most distinguishing characteristics of human beings. From this point forth, we propose a novel end-to-end model based on Convolutional Neural Network (CNN) that automatically extracts features from a given handwritten sample, which contains both handwritten text and numerals unlike the related work that uses only handwritten text, and classifies its owner’s gender. In addition to proposing a novel model, we introduce a new dataset that consists of 530 gender-labeled Turkish handwritten samples since, to the best of our knowledge, there does not exist a public gender-labeled Turkish handwriting dataset. Following an exhaustive process of hyperparameter optimization, the proposed CNN featured the most optimal hyperparameters and was both trained and evaluated on this dataset. According to the experimental result, the proposed novel model obtained an accuracy as high as 74.46%, which overperformed the state-of-the-art baselines and is promising on such a task that even humans could not have achieved highly-accurate results for, as of yet.

Despite the fact that chaotic systems do not have very complex circuit structures, interest in chaotic systems has increased considerably in recent years due to their interesting dynamic properties. Thanks to the noise-like properties of chaotic oscillators and the ability to mask information signals, great efforts have been made in recent years to develop chaos-based TRNG structures. In this study, a new chaos-based dual entropy core TRNG with high operating frequency and high bit generation rate was realized using 3D Pehlivan-Wei Chaotic Oscillator (PWCO) structure designed utilizing RK-Butcher numerical algorithm on FPGA and ring oscillator structure. In the FPGA-based TRNG model of the system, 32-bit IQ-Math fixed-point number standard is used. The developed model is coded using VHDL. The designed TRNG unit was synthesized for Virtex-7 XC7VX485T-2FFG1761 chip produced by Xilinx. Then, the statistics of the parameters of FPGA chip resource usage and unit clock speed were examined. The data processing time of the TRNG unit was achieved by using the Xilinx ISE Design Tools 14.2 simulation program, with a high bit production rate of 437.043 Mbit/s. In addition, number sequences obtained from FPGA-based TRNG were subjected to the internationally valid statistical NIST 800-22 Test Suite and all the randomness tests of NIST 800-22 Test Suite were successful.