Deep Learning, a subfield of Machine Learning, has enabled powerful tools for medical image analysis and diagnosis. These are mainly described as Deep Neural Networks (DNNs) and typically used for the identification of anatomy or pathology from medical images. Such tasks are subject to uncertainty due to the factors including the inherent imaging noise and artifacts, patient variability, unfulfilled modeling assumptions as well as the inter-annotator variability. However, DNNs do not provide well-calibrated and reliable uncertainty estimates regarding their outputs. My current research aims at developing methods for quantifying, improving and utilizing the uncertainty of DNNs in clinical settings.
Apart from the evaluation of uncertainty, interpretability of decisions is also crucial in medicine. Even though DNNs have been long criticized for being black-boxes, recent methods successfully uncover the inner-workings of DNNs and enable explanations for the decisions reached at as a result of the non-linear interactions of millions of artificial neurons. I am also interested in such explanations since they are as important as the uncertainty information.
At our institute, we are focusing on eye diseases, such as diabetic retinopathy and age-related macular degeneration, working with different modalities, e.g., retinal photography, 3D Optical Coherence Tomography (OCT), and collaborating with medical experts to validate our methods and bring them to life.
In the Past
Alzheimer’s disease is a major cause of dementia. Its diagnosis requires accurate biomarkers that are sensitive to disease stages. Neuroimaging techniques such as PET and MRI have been used as imaging biomarkers. However, it is virtually impossible to visually detect a slight decrease in regional cerebral blood flow or glucose metabolism in early stages of the disease. Moreover, visual inspection is susceptible to other factors like subjectivity and experience of the physician. On the other hand, voxel-based representations of neuroimagery can be used to perform both standardization and data-driven analysis. Computerized methods can also improve the speed of diagnosis with no compromise of accuracy and facilitate accurate diagnosis in cases where there is a lack of access to an experienced physician.
During my graduate studies, I focused on predictive modeling of Alzheimer’s disease using the variations of high-dimensional neuroimaging data. Simply, I regarded the induction of a classifier as the creation of a computational biomarker for disease staging. In this regard, I investigated clever utilizations of a wide range of machine learning algorithms in order to get the best of them, given the challenges of working with the scarce and high-dimensional neuroimaging data. I also strove for simplicity and comprehensibility in these predictive models, which is highly valuable for clinical decision-making.
- Data-driven Prognosis of Alzheimer’s Disease (Software Prototype), Patient Early Health Collaboration Project, 03/01/08 – 12/31/09, Funding from GE Healthcare, PIs: Vijay V. Raghavan and C. H. Chu, Laboratory for Internet Computing, CACS, UL Lafayette. Research Assistant: Murat Seçkin Ayhan
- Towards creating a Large and Scalable In-Memory Database Management System, State of Louisiana, Governor’s Information Technology Initiative Project, 07/01/07 – 07/31/09, PI: Vijay V. Raghavan, Laboratory for Internet Computing, CACS, UL Lafayette. Research Assistant: Murat Seçkin Ayhan
- Evliya Çelebi: A Middleware for Geographical Information, TÜBİTAK-SOBAG, 105K040, 2005, PI: Hayri Sever, Project completed at the Department of Computer Engineering, Baskent University, Research Assistant: Murat Seçkin Ayhan