Evaluation of window parameters of Noncontrast cranial ct brain images for Hyperacute and acute ischemic stroke classification with deep learning

Abstract

Most of the recent study about deep learning in medical images have revolved the ability of deep learning models to interpretation of diagnostic result and anatomical recognition. However, deep learning can also be used to enhance a wide range of non-interpretive issues such as image enhancement that relevant to radiologists and patients. For ischemic stroke, a noncontrast cranial computer tomography (NCCT) is imaging technique used for diagnosis. The cerebral infarction in early stage on NCCT is difficult to notice because of limitation of image. Normally, the patients need to do a computer tomography perfusion (CTp) for identification the damage, but it takes time while the patients should be received the treatment quickly. Another problem about NCCT image, the range of intensity is very wide and sparse. It is needed to rescale in the suitable range for the classifier. In this paper, we aim to find the suitable window setting for classifying the Hyperacute and Acute phase of ischemic stroke in NCCT image without CTp by using Inception V3. The dataset is prepared in axial slices. Each slide is classified to either normal or lesion. Due to limitation of the training samples availability, transfer learning is applied for weight initialization of the model. The result indicates that the model can perform well with window level at 35 and window width at 95, 90.84% accuracy.