Karaduman, Ismet; Pehlivanoglu, Meltem Kurt; Eker, Ayse Gul; Duru, Nevcihan; Dundar, Tolga Turan; Serdar Mutluer, Ahmet

Many of the decisions made and taken in the field of health are important for human life to be sufficient. In such an important area where human transactions are concerned, artificial intelligence-supported studies have a very important place, especially in recent years. This is the correct classification of four classes (Normal Tissue, Tumor Tissue, Hypervascularized Tissue and Background) separately with the 3D-CNN based deep learning model on the In-vivo Hyperspectral Human Brain Image dataset. For this purpose, firstly, three different data settings with 128 spectral bands were obtained from raw images, each containing 826 spectral bands, with dimension reduction methods. In this way, each data setting was made with the designs used in the 3D-CNN based deep learning model. As a result of the experimental studies, the highest F1 score of 99.69% was achieved with the 3D-CNN-based deep learning model.

AtıfScopus1

İncele

Ozturk, Candide; Eker, Ayse Gul; Pehlivanoglu, Meltem Kurt; Karaduman, Ismet; Olcer, Ercan; Duru, Nevcihan; Dundar, Tolga Turan; Mutluer, Ahmet Serdar; Dogan, Ihsan; Ericek, Omer Burak

Hyperspectral imaging (HSI) is an imaging method that differs from traditional methods by recording the light reflection and absorption properties of each point across a wide color range. HSI enables the non-invasive determination of chemical and structural differences in tissues with high accuracy. In the domain of medicine, hyperspectral imaging represents a novel methodology that facilitates novel approaches in brain surgery, including the distinction of tumors and blood vessels, cancer diagnosis, and the monitoring of tissue oxygen levels and blood flow during surgical procedures. Additionally, hyperspectral imaging has a wide range of applications in various fields, including agriculture, food, environment, geology, and defense.This review comprehensively examines current datasets, application areas, and computational approaches related to the use of hyperspectral imaging technology in brain surgery. HSI enables the precise differentiation of different brain tissues such as tumors, vessels, functional areas, and the spine by collecting data across a wide spectral band range for each pixel. Recent studies in the literature demonstrate that machine learning and deep learning-based methods can achieve high accuracy rates in tissue segmentation and the detection of pathological structures in brain surgery. Additionally, datasets developed at different centers and involving various patient groups provide an important foundation for enhancing the clinical efficacy of HSI.While HSI boasts significant advantages in terms of its non-invasiveness and real-time analysis potential, which benefit surgeons during intraoperative decision-making processes, challenges such as data size, computational costs, and device integration pose barriers to its routine use. In the future, it is anticipated that these limitations will be overcome through the development of efficient algorithms, AI-supported systems, and multi-center validation studies. In conclusion, HSI technology exhibits considerable potential to become a standard intraoperative instrument in neurosurgery, driven by advancing computational methods and increasing clinical validation.

AtıfScopus1

İncele

Eker, Ayse Gul; Kurt Pehlivanoglu, Meltem; Duru, Nevcihan; Dundar, Tolga Turan; Serdar Mutluer, Ahmet

Brain tumors are abnormal tissue masses formed as a result of uncontrolled cell proliferation in the brain, and surgery is the most common treatment method. During surgery, changes in the position of brain tissue are referred to as 'brain shift,' which prevents preoperative MR images from accurately reflecting the intraoperative situation and reduces surgical success. This study addresses the integration of MR and ultrasound (US) imaging methods to calculate brain shift and improve surgical accuracy in brain tumor surgeries. Different feature extraction algorithms such as SIFT, ORB, AKAZE, and BRISK were used to solve the alignment problem between MR and US images. Using an adaptive search strategy and 3D coordinate transformation, the amount of brain shift was calculated and real-time information was provided to surgeons. The use of the adaptive search strategy dynamically adjusted the search area according to the degree of brain shift, eliminating unnecessary regions and significantly increasing processing speed. Experimental results showed that the ORB algorithm performed best for small and moderate shifts (e.g., 3.78 ± 1.65 mm error for small shifts), while the SIFT algorithm was more successful for large shifts (9.91 ± 2.30 mm error). This method aims to enable surgeons to make more accurate decisions during surgery and to increase patient safety. In the future, it is planned to improve the accuracy of the method by using deep learning-based approaches and larger datasets.

İncele

Pehlivanoǧlu, Meltem Kurt; Varol, Onur; Aldemir, Osman; Uncu, Ata Emir; Duru, Nevcihan

The classification of dementia levels in Alzheimer's disease is divided based on various characteristics of the disease. In this study, for the purpose of classifying dementia levels in Alzheimer's disease using MRI images, various Convolutional Neural Network (CNN)-based deep learning models, as well as machine learning and Artificial Neural Network (ANN) models, were trained. Various image manipulation and augmentation techniques were applied to create new datasets, and model performances were observed on these datasets. As a different approach, numerical features were extracted from MRI images, and models were built using these extracted features. Fusion hybrid models were produced by combining the most successful models trained with the images and their numerical representations. Experimental results have proven that hybrid models outperform other models in terms of performance in the classification of Alzheimer's disease.

AtıfScopus1

İncele

Eker, Ayse Gul; Pehlivanoglu, Meltem Kurt; Ince, Ibrahim; Duru, Nevcihan

In brain tumors, which are formed by the abnormal growth of cells in brain tissue, early diagnosis, and accurate treatment are crucial. Brain tumor segmentation provides experts with the ability to accurately identify brain tumors and determine their location. Through segmentation, the type, size, and shape of tumors can be determined, enabling the application of appropriate treatment. In this study, the aim was to perform brain tumor segmentation on the Brain Tumor Figshare (BTF) dataset, which includes diverse images in terms of structural complexity, viewing angles, different device usage, noise, and bias field effects, without applying any preprocessing. Basic models such as U-Net and FCN were utilized, along with transfer learning-based approaches using VGG, XceptionNet, InceptionNet, and ResNet architectures as the underlying structures in these models. According to the experimental results, the 16-layer FCN architecture with VGG- 19 as the underlying structure achieved the highest score with a dice score of 0.9169.

AtıfScopus12

İncele

Pehlivanoglu, Meltem Kurt; Eker, Ayse Gul; Ince, Ibrahim; Karhan, Deniz; Duru, Nevcihan

Brain tumor classification categorizes brain tumors based on their various characteristics. It aids in determining suitable treatment strategies and predicting patient outcomes. Therefore, the objective of this study is to identify the optimal tumor classification model by utilizing various Convolutional Neural Network (CNN)-based deep learning models (including combined models) on a recently published tumor classification dataset. Furthermore, diverse image manipulation and augmentation techniques were applied and their impact on model performance was evaluated. The experimental results prove that the combined deep learning models outperform other approaches in terms of performance in multi-class brain tumor classification.

AtıfScopus3

İncele

Eker, Kadir; Pehlivanoglu, Meltem Kurt; Eker, Ayse Gul; Syakura, Muhammad Abdan; Duru, Nevcihan

Grammatical error correction (GEC) is the task of correcting grammatical, spelling, and word usage errors in texts. Detecting and correcting these errors makes it difficult for the reader to understand the text and negatively affects communication. Editors and proofreaders can identify and correct grammatical errors while reading texts. However, this manual correction takes a long time. With the development of technology, artificial intelligence-supported tools have been developed that automatically correct errors. In artificial intelligence-supported systems, this task is usually performed using natural language processing methods. In this study, a traditional neural network architecture (Attention BiLSTM) and a transformer-based language model (T5) were trained. Then, the success of these trained models and Large Language Models (LLMs) (ChatGPT, Google Bard, LanguageTool) in English grammatical error correction has been compared on the JFLEG dataset. According to experimental results, the pre-trained T5 language model beats others in all evaluation metrics for grammatical error correction.

AtıfScopus6

İncele

Eker, Kadir; Eker, Ayse Gul; Mandal, Dilek; Pehlivanoglu, Meltem Kurt; Duru, Nevcihan

Today, detection of malware is often done automatically with artificial intelligence supported systems. In this way, this task, which was previously done manually, can be performed more precisely and more quickly. Darknet traffic is an internet network with many different cybercrime threats. It is also fake traffic observed in empty address space. A globally valid set of Internet Protocol (IP) addresses that are not assigned to any host or device. In this study, a deep learning model and pervasive machine learning classification algorithms are compared using the CIC-Darknet2020 dataset to describe Darknet traffic. The 'Tor' and 'Vpn' classes in the dataset, 'Darknet', 'Non-Tor' and 'Non-Vpn' classes are also specified as harmless. Various adjustments have been made for the unbalanced dataset, which has been performed with multiple classification, various machine learning algorithms. Random Forest Algorithm, Decision Trees and Deep Learning Model were the algorithms with the most successful accuracy values with %97.3, %97.4 %98.2values respectively.

AtıfScopus2

İncele

Eker, Ayse Gul; Duru, Nevcihan; Eker, Kadir

EEG signals are one of the most basic methods used in identifying and analyzing brain activities. Visual representation of EEG signals can be achieved with spectrograms. Spectrograms represent a visual representation of a signal's signal strength over time. In this study, the signals in an EEG dataset containing 'positive', 'negative' and 'neutral' emotion classes were classified with a deep learning model, and then these signals were transformed into a spectrogram image in the dataset with convolutional network model and also with transfer learning (EfficientNet and XceptionNet). Multiple classification was performed with pre-trained models. The success value obtained by the classification of the EEG signals and the success of the visualization in this classification were measured and presented by comparison. While higher accuracy values were achieved in the classification of signals with the deep network model, in metrics such as precision and F1-score, the classification of images with the proposed convolutional network model achieved much higher performance.

AtıfScopus8

İncele