School of Engineering Department of Computer Science and Engineering 107 AI in Medical Imaging: Automatic Stroke Analysis on Brain CT Scans. Supervisor: CHEN, Hao / CSE Student: LEE, Hsin-ning / COMP Course: UROP1100, Summer Ischemic stroke (IS) occurs when blood flow to the brain is impaired or reduced, resulting in many deaths or long-term disabilities worldwide. CT perfusion (CTP), the most widely accepted method for doctors to diagnose IS, involves taking a series of 3D X-ray images after injecting a contrast dye. These comprehensive brain scans show how dye migrates in the brain over time, allowing doctors to pinpoint places where blood clots are forming. As it is difficult to directly interpret raw CT images, perfusion parameter maps derived from the raw CT scans are required to precisely indicate the problematic areas. The conventional clinical process of deriving these maps is deflected by the arterial investment function (AIF) of each voxel's concentration-time curve (TCC) in a 4D CTP dataset. These parameter maps provide important information about cerebral blood flow by highlighting abnormalities in ischemic lesions. Despite the advantages of these maps, there are also potential drawbacks. CTP scans are sensitive to noise, artifacts, and other external interference during scanning. To overcome this, deep learning techniques using artificial intelligence have been developed to identify and analyze problematic areas to improve the retrieval of these maps, but few of them have adopted semi-supervised learning to leverage the use of unlabeled data. Our project introduces a semi-supervised framework that combines parameter map prediction and ischemic lesion segmentation into a shared model. Infarction core lesions prediction is achieved through aggregated results from a multi-tasking learning scheme that estimates robust and reliable parameter maps. Public and in-house datasets including labeled and unlabeled data were used to increase the reliability and generalizability of our results. Based on current developments, we performed several experiments in order to further improve the model’s efficiency. With the aim of providing satisfactory results against low temporal resolutions, we explore potential strategies for the efficient use of limited data. In conclusion, this method optimizes the use of parameter maps and improves the segmentation process. These advances have the potential to significantly improve the diagnosis and treatment of IS, offering better outcomes for patients with this condition.
RkJQdWJsaXNoZXIy NDk5Njg=