Automatic Diagnostic Scoring of Brain CT Images as a Clinical Decision Support for Cerebrovascular Thrombolytic Therapy

Stroke is a medical emergency and can cause permanent neurological damage, complications, and even death. It is the leading cause of adult disability in Taiwan right now. Recently, IV tPA has emerged as an effective treatment if the stroke onset is within 3 hours and less than 1/3 MCA (Middle Cerebral Artery) territory is involved. Due to the high availability of CT scanners in the emergency care setting, the non-contrast CT imaging has become the most useful imaging tool available to aid in the prediction of outcome and assessment of appropriateness of thrombolytic treatment. Unfortunately, multiple studies have shown significant interobserver and intraobserver variability when it comes to quantifying the ischemic changes. In response to this issue, the Alberta Stroke Program Early CT Score (ASPECTS) was proposed to be a more reliable and reproducible grading system of evaluating acute stroke in comparison with the aforementioned one-third rule. But up to now, the usage of ASPECTS scoring relies on subjective perception and judgment of the clinicians and thereby is not immune to the issues of personal bias and inconsistency over time. Such issues have motivated us to make attempts of objective assessment by means of automatic computer image analysis and scoring. Our proposed method starts with segmenting the relevant images automatically so as to extract the ten regions of interest (ROI) based on a standard atlas derived from a number of normal subjects recruited locally from Chang Gung Hospital. Instead of using the existing standard human brain atlas developed by Western medical community based on Caucasian cohorts, we will develop a standard atlas based on local cohort to improve anatomical accuracy and precision. The overall segmentation begins with a global approximate registration followed by a regional precise registration. The former step relies on spatial normalization of the target image under inspection by means of rigid affine transformation to mesh with the standard atlas. Such an approximate registration is further fine-tuned to achieve precise registration by means of model-based deformable transformation. Once registration is completed, the mean pixel intensity and variance for each region of interest are computed and compared with those of its counterpart located at the contralateral side of the brain. The vase majority of stroke occurs to one side of the brain, leaving the other side intact. Wilcoxon two sample rank sum test is utilized to test the hypothesis that there exists significant difference in pixel intensity between opposite sides which indicates the sign of abnormal hypoattenuation and causes a point to be deducted. Finally the grade of each region of interest is summed up to be the overall ASPECTS score. Multiple studies have shown inadequate interpretation of non-contrast CT images—which has become part of the standard of care-- by acute care physicians who lack the needed training and experience of reading neuroimages. This inadequacy can lead to serious problems such as excessive hemorrhage rate and other complications. The proposed automated ASPECTS scoring system is aimed at providing objective and consistent assessment of the extent of hypoattenuation so as to offer valuable clinical decision support in the critical period of emergency care. It is expected that the quality of care will be enhanced by maximizing the usage of available imaging data, and the overall cost of patient management will be reduced by administering the optimal treatment in a timing manner. In the future, we hope that the techniques developed in this study can be extended to other clinical applications to further improve the quality of medical care at a reduced cost.

Project ID:PB10002-0109
External Project ID:NSC100-2218-E182-002