Modern imaging systems produce massive amount of data that can be used to obtain critical 3D information on targeted measurement objects (nodules), including volumes and largest diameter. Statistical analyses and experimental designs of quantitative image analysis experiments are important to understand the measurement variance components due to the various significant factors in the complicated quantitative image measurement process, including PACS (Picture Archiving and Communication Systems), CAD (Computer Assisted Diagnosis using Statistical Learning) as well as scanner settings and human factors due to human reader variation.
In this project, we have gained valuable experience in the emerging area of developing metrology support for medical image decision making, and we have focused mainly in the area of high resolution CT imaging for lung nodules,using both phantoms and clinical data, in the context of RSNA's QIBA(Quantitative Image Biomarker Alliance) framework. We have found that size and shape both play important roles in the measurement uncertainty in phantom and clinical nodule sizing experiments.
The statistical aspects of the project involve developing a brute force graphical approach to analysis of new types of complex experimental data and successfully adapting existing statistical methods /models to the complex nature of translational medical image data. Our specific contributions include the followings.