Browsing by Author "Oltu, Burcu"

Now showing 1 - 9 of 9

Automated Tuberculosis Detection Using Pre-Trained CNN and SVM
(2021) Oltu, Burcu; Guney, Selda; Dengiz, Berna; Agildere, Muhtesem
Tuberculosis (TB) is a dreadfully contagious and life-threatening disease if left untreated. Therefore, early and accurate diagnosis is critical for treatment. Today, invasive, expensive, or time-consuming tests are performed for diagnosis. Unfortunately, accurate TB diagnosis is still a major challenge. In the proposed study, a decision support system that can automatically separate normal and TB chest X-ray (CXR) images is presented for objective and accurate diagnosis. In the presented methodology, first various data augmentation methods were applied to the data set, then pre-trained networks (VGG16, MobileNet), were employed as feature extractors from augmented CXR's. Afterward, the extracted features for all images were fed into a support vector machine classifier. In training process, 5-fold cross-validation was applied. As a result of this classification, it was concluded that TB can be diagnosed with an accuracy of 96,6% and an area under the ROC curve (AUC) of 0,99.
Classification of Heart Sound Recordings With Continuous Wavelet Transform Based Algorithm
(2018) Karaca, Busra Kubra; Oltu, Burcu; Kantar, Tugce; Kilic, Erkin; Aksahin, Mehmet Feyzi; Erdamar, Aykut
Cardiovascular diseases are the major cause of death in the world. Early diagnosis of heart diseases provide an effective treatment. Heart diseases can be diagnosed using data obtained from heart sounds. Heart sounds are listened by a physician with auscultation method and the disease diagnosis can vary depending on the physician's experience and hearing ability. For this reason, automatic detection of anomalies in heart sounds can give more objective results. In this study, features were obtained by processing phonocardiogram signals taken from Physionet database. The heart sounds are classified as normal and abnormal using these features and the k - nearest neighbor method. As a result, sensitivity, specificity and accuracy were determined as 100%, 96.1% and 98.2%, respectively.
Classification of Sleep Apnea by Photoplethysmography Signal
(2018) Aksahin, Mehmet Feyzi; Karaca, Busra Kubra; Oltu, Burcu
Sleep apnea is a very common respiratory disorder in the community that includes a range from upper airway obstruction to respiratory abnormalities and the absence of a breathing effort, which can lower people's standard of living and even cause death. Therefore, the sleep apnea needs to be diagnosed in a practical way and with high accuracy. The diagnosis of apnea is made by recording the physiological parameters of the patient with polysomnography (PSG) method and by examination of these parameters by specialist physicians, but it is a tedious and time consuming process. In order to simplify the apnea diagnosis process, phospletismography (PPG) signals are used instead of PSG records. PPG signals are suitable for diagnosis of apnea because they reflect changes in respiration. In the proposed study, a decision support system was developed to automatically diagnose apnea and to make apnea diagnosis easier and more objective using PPG signals. In the decision support system, the peaks of the PPG signal were determined and the heart rate variability (HRV) vector was generated depending on the time difference between these peaks. The mean and standard deviation values of the generated vector are determined as features for each epoch. The presence of the apnea at each epoch is classified using "Subspace K Nearest Neighbor (Subspace KNN)" and specified features. The "Subspace KNN" classifier was trained with 85% accuracy and then system was tested. As a result, sensitivity, accuracy and specificity rates were calculated as 91%, 95% and 90% respectively.
Derin öğrenme yöntemleri ile akciğer grafilerinde patoloji sınıflandırılması
(Başkent Üniversitesi Fen Bilimleri Enstitüsü, 2025) Oltu, Burcu
Akciğerler, solunum sisteminin temel organları olup yapısal bozukluklar, enfeksiyonlar veya çevresel etkenler nedeniyle işlev kaybına uğrayabilmektedir. COVID-19, tüberküloz, pnömoni, pulmoner fibröz, atelektazi, kardiyomegali ve pnömotoraks gibi akciğer hastalıkları erken evrede teşhis edilmediğinde ölümcül sonuçlara yol açabilmektedir. Bu nedenle akciğer hastalıklarının erken ve doğru şekilde teşhis edilmesi büyük önem taşımaktadır. Akciğer grafileri, düşük maliyet ve hızlı uygulanabilirlik gibi avantajlarıyla bu amaçla yaygın olarak kullanılmaktadır. Ancak örtüşen anatomik yapılar, sınırlı uzman sayısı ve yüksek görüntü hacmi gibi etkenler, bu görüntülerin yorumlanmasını zorlaştırmakta, günlük yaklaşık %3-5 oranında hata yapılmasına yol açmaktadır. Bu doğrultuda akciğer grafilerinin yorumlanabilmesi için radyologlara yardımcı olacak otomatik ve güvenilir teşhis sistemlerine duyulan ihtiyaç gün geçtikçe artmaktadır. Bu tez çalışmasında akciğer grafilerinin yüksek performansla sınıflandırılması amacıyla uçtan uca bir derin öğrenme modeli geliştirilmiş ve bu modelin başarısını artıracak özgün kayıp fonksiyonlarını önerilmiştir. Modelin omurgası olarak önceden eğitilmiş DenseNet201 mimarisi kullanılmıştır. DenseNet201’den çıkarılan öznitelik haritalarından daha zengin uzamsal bilgiler elde etmek amacıyla evrişimsel uzun kısa süreli bellek (Convolutional Long Short-Term Memory (ConvLSTM)) katmanı, kanal bazında özniteliklerin vurgulanması için sıkma-bırakma bloğu (Squeeze and Excitation (SE)) ve uzun menzilli bağımlılıkları yakalayan görü dönüştürücüler (Vision Transformer (ViT)) modele entegre edilmiştir. Ayrıca, küresel ortalama havuzlama (Global Average Pooling (GAP)) katmanı ile sınıflandırma için önemli uzamsal bilgilerin korunması sağlanmıştır. Bu bileşenlerin kombinasyonu ile önerilen model, yüksek sınıflandırma performansı sunmuştur. Çalışmada ayrıca, sınıflandırma başarısında önemli etkisi olan kayıp fonksiyonları derinlemesine incelenmiştir. Standart fonksiyonlara ek olarak hibrit ve dinamik yapılarda fonksiyonlar tasarlanmış, sınıflandırma eğilimlerine karşı ceza terimi içeren özgün bir fonksiyon önerilmiştir. Önerilen kayıp fonksiyonlarının avantajlarını bir araya getirerek sınıflandırma performansını artıracak farklı topluluk yaklaşımları uygulanmıştır. Bu fonksiyonların performansları sistematik olarak analiz edilmiş ve farklı modellerle de test edilerek modelden bağımsız başarı sağladığı gösterilmiştir. Başkent Üniversitesi Ankara Hastanesi Radyoloji Bölümü'nden elde edilen klinik veriler ile açık erişimli veri kümeleri birleştirilerek 7, 8, 9, 10, 12, 14 ve 15 sınıflı alt veri kümeleri oluşturulmuş ve her bir alt kümede önerilen model ve kayıp fonksiyonları detaylı şekilde test edilmiştir. Böylece hem veri kümesinden hem de sınıf sayısından bağımsız olarak modelin ve kayıp fonksiyonunun başarısı ortaya konmuştur. Elde edilen deneysel sonuçlar, önerilen yöntemin literatürdeki yaklaşımların çoğunu geride bıraktığını ve doğruluk, F1-skoru, AUC gibi performans metriklerinde yüksek başarı sergilediğini ortaya koymaktadır. Ayrıca modelin farklı senaryolarda tutarlı ve yüksek performans sergilemesi, güçlü bir genelleme yeteneğine sahip olduğunu kanıtlamaktadır. Sonuç olarak, bu tez kapsamında, akciğer grafilerinin otomatik olarak sınıflandırılmasını sağlayan yüksek performanslı ve tekrarlanabilir sonuçlar üreten yenilikçi bir model ve özgün kayıp fonksiyonları geliştirilmiştir. The lungs are the primary organs of the respiratory system and can lose functionality due to structural disorders, infections, or environmental factors. Lung diseases such as COVID-19, tuberculosis, pneumonia, pulmonary fibrosis, atelectasis, cardiomegaly, and pneumothorax may lead to fatal outcomes if not diagnosed at an early stage. Therefore, early and accurate diagnosis of lung diseases is crucial. Chest radiographs (CXRs) are widely used for this purpose due to their advantages, such as low cost, quick applicability, and low radiation dose. However, factors like overlapping anatomical structures, limited specialists, and high image volume make interpreting CXRs challenging, resulting in a daily error rate of 3–5%. Thus, the demand for automated and reliable diagnostic systems to assist radiologists in interpreting CXRs is increasing. In this thesis, an end-to-end deep learning model was developed for high-performance classification of CXRs, and novel loss functions were proposed to enhance the model’s performance. The pre-trained DenseNet201 architecture was used as the backbone. A Convolutional Long Short-Term Memory (ConvLSTM) layer to obtain richer spatial and temporal information from the feature maps obtained from DenseNet201, a Squeeze and Excitation (SE) block to emphasize channel-wise features, and Vision Transformers (ViT) to capture long-range dependencies were integrated into the model. In addition, a Global Average Pooling (GAP) layer was used to preserve important spatial information for classification. The combination of these components enabled high classification performance. The study also examined the often-overlooked role of loss functions in classification performance. Beyond standard functions, hybrid and dynamic structures were designed. A custom function with a penalty term was proposed to improve sensitivity to misclassification. Ensemble approaches combining the strengths of proposed loss functions were implemented. These loss functions were systematically analyzed and tested with various models, demonstrating architecture-independent success. To evaluate the performance and generalization ability of the proposed model and loss functions, a comprehensive dataset was created by combining clinical images from Başkent University Ankara Hospital with open-access datasets. This dataset was divided into subsets containing 7, 8, 9, 10, 12, 14, and 15 classes, and extensive testing was conducted on each subset. Results showed that the proposed method outperforms many existing approaches in the literature, achieving superior accuracy, F1-score, and AUC. The model’s consistent performance in various scenarios demonstrated strong generalization. In conclusion, this thesis presents a novel deep learning model and original loss functions that produce high-performance and reproducible results for the automatic classification of CXRs
Heart sound recording and automatic S1-S2 waves detecting system design
(2020) Aksahin, Mehmet Feyzi; Oltu, Burcu; Karaca, Busra Kubra
The second leading cause of death in the world is cardiovascular diseases. Diagnosis of vast majority of cardiovascular diseases is made by listening to heart sounds by specialists (auscultation method). However, since the method of auscultation depends on the experience and hearing ability of the specialist, obtained results can be subjective. Therefore, digitization and visualization of heart sounds enables accurate, rapid and economical diagnosis of cardiovascular diseases, especially heart valve diseases. For this purpose, a device prototype that collects the heart sound from human body and also amplifies, filters, displays and records collected data on digital environment was designed in the first part of this study. In order to test the working accuracy of the designed device, clinical applications were carried out with the permission of the ethics committee and as the result of this application 15 heart sound recordings from 5 different disease groups(mitral insufficiency, mitral-aortic insufficiency, mitral-tricuspid insufficiency, mitral-aortic tricuspid insufficiency and healthy heart sound recordings) were collected.and obtained recordings were examined. The most effective parameter for the diagnosis of heart valve diseases is the location of the S1-S2 heart sounds. For this reason, in the second part of the study, a medical decision support system was established to detect the S1-S2 locations to assist physicians in their diagnosis. In this context, heart sounds are first filtered by discrete wavelet transform. Then, the S1-S2 waves in the filtered signal are made evident by the teager energy operator and rule-based algorithm. As a result, S1-S2 locations in normal and pathological data were detected with 98.67% sensitivity, 97.69% specificity and 98.18% accuracy.
A novel electroencephalography based approach for Alzheimer's disease and mild cognitive impairment detection
(2021) Oltu, Burcu; Aksahin, Mehmet Feyzi; Kibaroglu, Seda; 0000-0002-3964-268X; AAJ-2956-2021
Background and objective: Alzheimer's disease (AD) is characterized by cognitive, behavioral and intellectual deficits. The term mild cognitive impairment (MCI) is used to describe individuals whose cognitive impairment departing from their expectations for the age that does not interfere with daily activities. To diagnose these disorders, a combination of time-consuming, expensive tests that has difficulties for the target population are evaluated, moreover, the evaluation may yield subjective results. In the presented study, a novel methodology is developed for the automatic detection of AD and MCI using EEG signals. Methods: This study analyzed the EEGs of 35 subjects (16 MCI, 8 AD, 11 healthy control) with the developed algorithm. The algorithm consists of 3 methods for analysis, discrete wavelet transform(DWT), power spectral density (PSD) and coherence. In the first approach, DWT is applied to the signals to obtain major EEG sub-bands, afterward, PSD of each sub-band is calculated using Burg's method. In the second approach, interhemispheric coherence values are calculated. The variance and amplitude summation of each sub-bands' PSD and the amplitude summation of the coherence values corresponding to the major sub-bands are determined as features. Bagged Trees is selected as a classifier among the other tested classification algorithms. Data set is used to train the classifier with 5-fold cross-validation. Results: As a result, accuracy, sensitivity, and specificity of 96.5%, 96.21%, 97.96% are achieved respectively. Conclusion: In this study, we have investigated whether EEG can provide efficient clues about the neuropathology of Alzheimer's Disease and mild cognitive impairment for early and accurate diagnosis. Accordingly, a decision support system that produces reproducible and objective results with high accuracy is developed.
Sağlıklı, hafif bilişsel bozukluk ve alzheimer hastalığı elektroensefalografi sinyallerinin sınıflandırılması
(Başkent Üniversitesi Fen Bilimleri Enstitüsü, 2020) Oltu, Burcu; Akşahin, Mehmet Feyzi
Demans terimi, beyin hücrelerinin ölümü veya hasar görmesi nedeniyle bilişsel ve davranışsal fonksiyonların bozulması ile karakterize edilen nörodejeneretif bozukluğu tanımlar. Demansın en yaygın görülen türü olan Alzheimer Hastalığı (AH) bilişsel, entelektüel eksiklikler ve davranış bozuklukları ile karakterize edilen progresif bir hastalıktır. Hafif bilişsel bozukluk (HBB) ise yaşları için beklenenin ötesinde bilişsel bozukluk gösteren; fakat AH tanı ölçütlerini karşılayacak kadar kuvvetli semptomlar göstermeyen bireyleri tanımlar. Her yıl yaklaşık %10-15 oranında HBB hastası, AH’ye ilerlemektedir. Bu hastalıkların teşhisi; bilişsel testlerin, fiziksel ve nörolojik muayenelerin, nörogörüntüleme yöntemlerinin, kan ve beyin omurilik sıvısı incelemelerinin sonuçlarının yorumlanmasıyla yapılır. Dolayısıyla tanı süreci zaman alan, hedef hasta popülasyonu için zorlukları olan ve subjektif sonuçlar verebilen bir süreçtir. Bunun yanında demansın erken teşhisi hastalara ve yakınlarına bu süreci planlamaları için zaman tanımaktadır. Aynı zamanda bu hastalıkların tedavisi olmasa da erken evrede başlanan tedavi süreçleri, hastalığın ilerlemesini yavaşlatabilir ve semptomların hafiflemesini sağlayabilir. Bu tez çalışmasında da AH ve HBB’nin tanısını sağlamak için, invaziv olmayan, ucuz, objektif sonuçlar verebilen bir yöntem olan elektroensefalografi (EEG) sinyallerinin kullanılabilirliği araştırılmıştır. Yapılan araştırmada Başkent Üniversitesi Ankara Hastanesi Nöroloji Kliniği’ndeki uzman doktorlar tarafından yapılan teşhis doğrultusunda gerçek hastalardan toplanan EEG sinyalleri incelenmiştir. Çalışmada EEG sinyalleri, ayrık dalgacık dönüşümü (ADD), güç spektral yoğunluğu (GSY), koherans, sürekli dalgacık dönüşümü (SDD) yöntemleri kullanılarak incelenmiştir. Yapılan incelemeler sonucunda farklı yöntemler için farklı öznitelik çıkarımları yapılmıştır. Ayrık dalgacık yöntemi ile yapılan analizlerde sinyaller 6 alt banda (delta (0-4 Hz), teta (4-8 Hz), alfa (8-12 Hz), beta1 (12-16 Hz), beta2 (16-32 Hz), gama (32-48 Hz)) ayrıştırılmış. Elde edilen 6 bandın ortalama, maksimum-minimum değerleri, varyansı ve standart sapması öznitelik seti olarak belirlenmiştir. Ayrıca elde edilen alt bantların GSY’sinin genlik toplamı ve varyansı frekans öznitelikleri olarak belirlenmiştir. Bunun yanında interhemisfer kanal çiftleri arasındaki koherans analizi sonucunda elde edilen normalize alt bant koheransları öznitelik olarak çıkartılmıştır. Son olarak sinyale uygulanan SDD sonucunda elde edilen, alt bant frekanslarına ait katsayıların ortalama, standart sapma ve kendi aralarındaki oranları kullanılarak farklı bir öznitelik seti oluşturulmuştur. Elde edilen öznitelik setlerinin AH, HBB ve sağlıklı kontrol EEG sinyallerini ayrıştırma başarısının ölçülmesi için destek vektör makineleri, k-en yakın komşu, karar ağaçları gibi sınıflandırıcı algoritmaları ile sınıflandırma çalışmaları yapılmıştır. Çalışma sonucunda literatürdeki doğruluk oranlarına göre yüksek ya da aynı seviyede sonuçlar verebilen ve başarılı bir şekilde teşhis koyabilen bir karar destek sistemi geliştirilmiştir. The term dementia refers to neurodegenerative disorder characterized by impaired cognitive and behavioral functions due to death or damage of brain cells. Alzheimer's Disease (AD), the most common form of dementia, is a progressive disorder characterized by cognitive, intellectual deficits and behavioral disorders. Mild cognitive impairment (MCI), on the other hand, refers to individuals with cognitive impairment beyond their expectations for age, but not strong enough to meet the diagnostic criteria for AD. Every year, approximately 10-15 percent of MCI patients progress to AD. Diagnosis of these diseases is made by interpreting the results of cognitive tests, physical and neurological examinations, neuroimaging methods, and blood and cerebrospinal fluid examinations. Therefore, the diagnostic process is time-consuming, has difficulties for the target patient population and can yield to subjective results. In addition, early diagnosis of dementia gives patients and their families time to plan this process. At the same time, even there is no treatment for these diseases, treatment processes initiated at an early stage may slow down the progression of the disease and alleviate the symptoms. In this thesis, we’ve investigated the feasibility of electroencephalography (EEG), a noninvasive, inexpensive, objective method for the diagnosis of AD and MCI. In this study, EEG signals collected from real patients were examined in accordance with the diagnosis made by specialists in Neurology Clinic of Baskent University Ankara Hospital. In this study, EEG signals were examined by using discrete wavelet transform (DWT), power spectral density (PSD), coherence, continuous wavelet transform (CWT) methods. As a result of the investigations, different features were extracted for each different method. In the analysis made by DWT, the signals are decomposed into 6 sub bands (delta (0-4 Hz), theta (4-8 Hz), alpha (8-12 Hz), beta1 (12-16 Hz), beta2 (16-32 Hz), gamma (32-48 Hz)). Mean, maximum-minimum values, variance and standard deviation of the 6 sub bands were determined as a feature set. Moreover, the amplitude sum and variance of the PSD of the sub bands obtained were determined as frequency features. In addition, the normalized sub band coherences, obtained as a result of the coherence analysis between the interhemispheric channel pairs, were extracted as a feature set. Finally, a different feature set was formed by using the mean, standard deviation and the ratio of the coefficients of the sub-band frequencies obtained from the CWT analysis of the signal. In order to measure the success of discrimination of AD, MCI and healthy control EEG signals, classification studies were performed with classifier algorithms such as support vector machines, k-nearest neighbor and decision trees. As a result of this study, a decision support system was developed that can diagnose successfully and can yield higher or equal results in the means of accuracy of the literature.
Sleep Apnea Detection Using Blood Pressure Signal
(2018) Aksahin, Mehmet Feyzi; Oltu, Burcu; Karaca, Busra Kubra
Sleep apnea is a common respiratory disease. Apnea affects sleep quality, reduces people's life standards, and it can result in death at advanced stage. Therefore the ability to detect the apnea quickly and accurately is important for the treatment of this disease. Apnea is diagnosed by specialists however this is a long and exhausting process. Accordingly, a decision support system that automatically diagnoses apnea has been developed to facilitate this process and make it more objective. The developed decision support system in this study is based on patient's blood pressure signals instead of traditional Polysomnography (PSG) records, which requires various physiological signals measured from the patients. In the examined blood pressure signals, the change that results from each heart beat was determined and heart rate variability (HRV) was calculated based on these changes. At the same time, maximum and minimum amplitude values were found for each change period and amplitude variability vector was created. The features for each epoch were determined using the generated amplitude variability vector and HRV data. Presence of apnea in each epoch is classified with determined features and with the use of "Quadratic SVM" classifier. The Quadratic SVM classifier was trained with 87.5% accuracy and then the system is tested. As a result 75.4% sensitivity and 75% positive predictive values were obtained.
A Systematic Review of Transfer Learning-Based Approaches for Diabetic Retinopathy Detection
(2023) Oltu, Burcu; Karaca, Busra Kubra; Erdem, Hamit; Ozgur, Atilla; 0000-0002-9237-8347; 0000-0003-1704-1581; AAD-6546-2019
Diabetic retinopathy, which is extreme visual blindness due to diabetes, has become an alarming issue worldwide. Early and accurate detection of DR is necessary to prevent the progression and reduce the risk of blindness. Recently, many approaches for DR detection have been proposed in the literature. Among them, deep neural networks (DNNs), especially Convolutional Neural Network (CNN) models, have become the most offered approach. However, designing and training new CNN architectures from scratch is a troublesome and labor-intensive task, particularly for medical images. Moreover, it requires training tremendous amounts of parameters. Therefore, transfer learning approaches as pre-trained models have become more prevalent in the last few years. Accordingly, in this study, 43 publications based on DNN and Transfer Learning approaches for DR detection between 2016 and 2021 are reviewed. The reviewed papers are summarized in 4 figures and 10 tables that present detailed information about 29 pre-trained CNN models, 13 DR data sets, and standard performance metrics.