Deep learning based physician decision support system for diagnosis of stroke from brain CT images

Deep learning models, which detect anomalies, minimize human error, provide accurate results and reduce time spent, are utilized to a certain extent. Especially radiology and neuroradiology are suitable fields for the application of deep learning models due to the large amount of data generated. One of the most important areas in which medical diagnosis reliability is achieved by applying image processing rules in a meaningful way is medical imaging. Stroke is one of the diseases in the clinical management of neuroimaging plays an important role. Stroke, which is a major health problem all over the world, is one of the most important causes of disability and mortality in adults, causing physical, social, psychological, and economic destruction. Early diagnosis becomes difficult as the brain behaves like healthy brain functions in the early stages of stroke. Even specialist radiologists may overlook the findings at the initial stage of stroke. The time factor is very valuable in stroke. Therefore, in order for the treatments to be effective, these treatments should be applied as soon as possible with the onset of symptoms. Early treatment saves lives. Therefore, imaging is needed in stroke cases in order to understand what the cause of stroke (ischemic, hemorrhagic) is, to rule out bleeding, determine the infarct area, obtain information about the etiology, and plan treatment. Non-contrast CT is the primary imaging protocol used in the initial evaluation of patients with suspected stroke. Deep learning studies on determining the pathological type of stroke on Brain CT images obtained without contrast agent administration, which is the first step of rapid stroke protocols, are very limited. This thesis aims to show the impact of a deep learning-based new approach to automatic segmentation and classification for the detection and diagnosis of stroke lesions from non-contrast brain CT images is proposed. This study provides have been filtered out based on specific criteria for stroke detection, stroke classification, and segmentation of the pathological type of stroke from brain CT images using deep learning-based Convolutional Neural Networks published in the period 2015 to 2020. The obtained images consist of hemorrhagic and ischemic strokes as well as patients with normal CT image findings. Various imaging modalities can be used to detect stroke infarct areas.The accuracy rate for classification results is 94.75%. In the experiment, the AUC showed a success rate of 98.47%. It means there is a 98.47% probability that our model will be able to distinguish between the brain CT class involving stroke and the normal brain CT class. As a result of 300 epochs, each lasting 10s, our proposed segmentation model reached a validation mean IoU score of 83.32%. In this thesis, a physician decision support system is proposed for the detection of brain stroke as a result of the successful results obtained for classification and semantic segmentation problems.