ROC Analysis Of3D X-ray CT Performance For Lesion Detection

X-ray CT has been widely used in medical imaging and industrialnon-destructive testing fields. As CT imaging cannot accurately restore the structureof objects, misdiagnosis and false detection often happens. Since getting images isnot the ultimate goal of experiments in these applications, the task-based diagnosishas been a hot topic of research. As a comprehensive evaluation methodology, ROCmethod is well used to evaluate the performance of detection quantitatively inradiographic exams. In this paper, we introduce ROC method and its derivativemethods to evaluate X-ray CT systems andassess the detection performance, as wellas to give recommendations for system optimization.First of all, we address a brief introduction about the background ofthe workincluding the history, applications and possible future directions of ROC research.Secondly, we construct an ROC model and an LROC model to fit ROC curves.Calculation of the area under ROC curve and analysis of the deviation are followed.Through these steps, we complete the method to assess CT images. Then, we makesome deep investigate to assess the detection performance and system optimization,and give some suggestions about how to improve the ROC and LROC methods. Themain findings are in two aspects: the improvement of detection performance fromusing3D CT imaging and3D visualization, the optimization of system parameter fora single slice helical CT using the above methodology.In the aspect of that the capability of lesion detection using3D CT imagingtechnology can be improved by3D visualization, the main achievements are asfollows:1) We experimentally demonstrated that3D visualization had great advantage indefect diagnosis. Through data analysis we obtained that using multi slices could geta nice result in the case of lacking3D visualization. 2) We comprehensively studied how the performance of lesion detection can beimproved by3D visualization and multi slices. The most significant improvementoccurred while the AUC in case of using one single slice only is0.7-0.8.3) The trends of ROC and LROC results are similar. Whether the defectionlocation is fixed or not has a significant effect on diagnosis accuracy.In the aspect of Single Slice Helical CT (SSHCT) system optimization, weobtained following results:1) We found the relationship of imaging quality to the detector thickness. Thebest result acquired while the thickness of detector was1-1.25times pitches.2) We studied some other figure of merit (FOM) to compare with experimentalROC methods. Our results showed that SNR and HTC can be used to substitute theROC curve in some cases, but is quite different from what LROC results demonstrate.This tells us that a new FOM is needed if we want to correspond it with LROCmethod.3) We made a though study in the ROC model and LROC model. Thelimitations of the existing two models are addressed. Finally, we proposed somepossible future improvement.