Detection of ischemic stroke on medical scans using deep learning methods

Stroke is a public health problem that causes millions of people to lose their lives and experience permanent physical and cognitive problems every year. It occurs as a result of a sudden interruption of blood flow in the brain. Diagnosing stroke cases in the first period when symptoms begin to appear and applying to the hospital, that is, in the acute phase, and starting treatment as soon as possible is an important factor for improving the individual's quality of life after treatment, as well as reducing the life risk. In addition to the need for expert opinion to detect stroke lesions, the fact that manual segmentation is a time-consuming process hinders the need for rapid diagnosis in critical hours. Detection of stroke lesions with computer-aided algorithms is a suitable solution to this need. Delays in the transition to treatment in stroke may cause loss of function in the patient and death of brain tissue, which may result in death. In recent years, developments in deep learning algorithms have become part of clinical solutions in medical screening analysis used in the diagnosis and treatment of diseases. Successful results in many recent studies are promising for developing highly accurate systems that can assist experts. This thesis study examines the effect of deep learning methods on a publicly available dataset (ISLES'22) to provide automatic segmentation of lesions on Magnetic Resonance Imaging (MRI) scans, which are commonly used in ischemic stroke cases. In the segmentation task, the performance of the methods combining the U-Net architecture, one of the deep learning models, and two different convolutional neural networks was measured on diffusion-weighted MR images (DWI). In experimental analyses, average scores of 0.85, 0.79, 0.86 and 0.87 were achieved on the test set for the key performance metrics F1-score, Dice similarity coefficient and Precision and Recall, respectively, by using models trained with MR images of 250 patients. The experimental study results showed that they overlap with the results obtained by participants in the ISLES'22 dataset, which had challenging stroke lesions due to small size and irregular shapes, and as a result, the models created with U-Net have the potential to successfully segment ischemic stroke lesions. It is evaluated that the proposed method will facilitate the diagnosis of stroke in clinical stages and can be an assistant tool that can be used in the decision-making process.