Statistical image analysis of torso FDG-PET/CT images for automated cancer detection and quantification during chemotherapy

1890 Objectives Standardized uptake value (SUV) on Fludeoxyglucose (FDG) positron emission tomography (PET) is used for evaluating regional activities of glucose metabolism related to cancer development. Radiologists have to understand the SUV ranges or typical normal/abnormal values based on patients physiological accumulations. Ideas of statistical image analysis are widely accepted for brain function analysis in dementia diagnosis. We have developed a new scheme by applying the analysis idea to torso FDG-PET diagnosis. The purpose of this study was to verify the effectiveness of the statistical approach in torso FDG-PET imaging. Methods Institutional Review Board (IRB) approved cases (49 normal and 10 abnormal) were used in this study. The proposed scheme consisted of the following steps: (1) anatomical standardization of images based on organ recognition on CT images, (2) normal model construction, and (3) Z-score calculation). Bounding boxes surround organs region on CT images were determined automatically based on our organ recognition method (Fig.1). The normal model contains the mean and standard deviation (SD) after the standardized cases were summarized (Fig.2). The Z-score was obtained based on the mean and SD determined by the normal model by comparing with the standardized abnormal cases by voxel by voxel. To validate the Z-score index, we manually extracted 451 normal and 397 abnormal regions in liver and right lung images. SUV and Z-score in each region were obtained after the statistical analysis. We evaluated the discrimination performance based on receiver operating characteristics (ROC) analysis methods. Results The discrimination performance of SUV and Z-score for the liver and the lung ROIs were measured by using the area-under-the-ROC-curves (AUCs). Each AUC was over 0.97. When the ROIs were pooled, the AUCs of Z-score and SUV were 0.99 and 0.98, respectively. The discriminant performance of the Z-score was slightly better than that of SUV when the ROIs were pooled (Fig.3). The results suggested a possibility of a new quantitative determination method to support knowledges of SUV ranges as FDG accumulations in each organ. In addition, a combination of SUV and Z-score may provide a high accuracy of discriminations when readers interpret torso FDG-PET images. Conclusions Statistical image analysis of torso FDG-PET images may provide a new index for the interpretation. $$graphic_B1710888-1054-49D5-853E-597FD274B3B0$$ $$graphic_0FB7D015-E7E2-4606-8334-82747F2BD602$$ $$graphic_C27A9D73-7B4E-4BAD-BBCB-BB5891B0AA37$$