Deep Learning-based Eye Disease Detection, Severity Prediction and Localization: Automated analysis of fundus images using AI-based techniques has significant potential in improving healthcare services. We have developed a deep learning-based system that analyzes fundus images, predicts eye diseases, predicts the stage of diabetic retinopathy, and locates the regions on fundus images related to the eye diseases. The system have three main components: (i) The eye disease prediction model is trained with a dataset containing fundus images with cataracts, glaucoma, diabetic retinopathy, age-related macular degeneration, hypertension, and pathological myopia. (ii) The diabetic retinopathy severity prediction model is trained with fundus images obtained from patients with diabetic retinopathy labeled with the stage of the disease. (iii) The system locates the regions related to the eye diseases using the Grad-CAM algorithm, which identifies the regions that CNN-based architecture used to make its classification decision regarding eye diseases.
- This research is presented at Engineering Sciences and Research Student Congress, Ankara, Atılım University, 2024.
- The project has been selected as the winner among 31 groups at the 17th Project Fair and Competition held at ESTU on May 28, 2024.
Researcher: Göksu Turaç, Onur Haktan / Advisor: Sema Candemir
Machine Learning-Based Model Development in Mobile Health Applications and Device-Cloud Deployment: Mobile Health allows health applications to be portable and enables mobile phone users to instantly access health solutions. In recent years, models developed with machine learning (ML) approaches have played an important role in healthcare in diagnosing and monitoring various diseases. However, the integration and execution of ML models, especially those based on image processing and deep learning, on mobile devices can be challenging due to the devices`s limited processing capacity and limited storage resources. This study describes the step-by-step development of a ML-based model, and its integration into mobile devices and cloud environments. Predicting skin disease has been chosen as a use-case mHealth model due to its usability. A MobileNet architecture hase been developed due to its suitability for mobile applications. Transfer learning technique, data augmentation and different loss functions have been considered to improve the model`s performance. The trained mHealth model is integrated into both device and cloud. All technical details of the process and a comparision of device-cloud intergration are provided.
Researcher: Özge Çiçek / Advisor: Sema Candemir
Explainable Artificial Intelligence-based Diagnosis of Lung Diseases from Chest X-rays: This project explores the utilization of Explainable Artificial Intelligence (XAI) techniques in the diagnosis of lung diseases from chest X-rays (CXR). The study uses the National Institutes of Health chest X-ray dataset containing lung pathologies: Atelectasis, Consolidation, Infiltration, Pneumothorax, Edema, Emphysema, Fibrosis, Effusion, Pneumonia, Pleural thickening, Cardiomegaly, Nodule, Mass, and Hernia. We have been developing a multi-class classifier based on convolutional neural networks integrated with an XAI component. The model has been learning the distinctive features in CXR associated with different lung diseases. The XAI module pinpoints the locations of these distinctive regions on CXR, thereby enhancing the interpretability of diagnostic decisions.
This research is supported by Tubitak-2209A Research Project Support Programme for Undergraduate Students
Researchers: İlayda Su İrday, Emirhan Sarp / Advisor: Sema Candemir
Deep Learning Models in Cognitive Diseases: Today`s technology empowers computers to process large amounts of medical data, enabling the development of AI-based systems. Especially, deep learning-based algorihtms have demonstrated performance comparable to expert decision-making, even uncovering previously unknown correlations in medical data. This talk will focus on using AI algorithms to identify, and assess the prognosis of cognitive impairment, Alzheimer`s disease and mild cognitive impairment. The talk will also address the challenges in the development of deep learning models for cognitive disease detection and prognosis.
Researchers: Sema Candemir, Demet Özbabalık Adapınar
Self-Training Segmentation for Covid-19 Lesions: With the advancements in computer technology and its reflection in the field of health, deep learning-based techniques have revolutionized the diagnosis and treathment of many diseases. The success of deep learning-based techniques relies on large-scale annotated dataset. However, curating and annotating large amounts of medical data presents challenges related to patient privacy, data anonymization and domain expertise necessity for annotation. The purpose of this study is to enhance the segmentation performance on a dataset with limited annotation using a semi-supervised approach. Semi-supervised training combines labeled and unlabed data and enhances the model performance by inferring information from unlabeled dataset. The segmentation performance of the algorithm is being tested on CT sequences for successful Covid-19 lesion segmentation. This study is currently ongoing.
This research is financially supported by the Tubitak-BIDEB 2232 International fellowship program under the grant number 121C085.
Researchers: Umut Kaan Kavaklı, Tansu Temel, Mehmet Koç, Sema Candemir.