Abstract
1205
Objectives: AutoPERCIST is a software implementation of the analysis methods and framework described by the PERCIST 1.0 criteria. It automates most steps in a PERCIST analysis, reducing operator bias, user/computer interactions and time of analysis, thus providing an objective and quantitative assessment of disease presence and tumor burden to be used in response determination. It is written in the Java programming language. The aim of this study was to assess the amount of time necessary to accurately process and review a series of complex, whole body 18F-fluorodeoxyglucose positron emission tomography (FDG PET) images using the AutoPERCIST software.
Methods: FDG PET images of patients suspected of having melanoma were retrospectively identified and uploaded to a research workstation (Intel Xeon processor running Windows 7). The image field of view (FOV) was either of the whole body or from the vertex to the thighs. Each PET image dataset was consecutively loaded into AutoPERCIST and reviewed/analyzed by a trained researcher (Figure 1 shows a representative view of the software with a loaded image). The time to load, review, and extract necessary parameters (peak standardized uptake value, maximum standard uptake value, total lesion glycolysis, and metabolic tumor volume) was recorded. AutoPERCIST results were compared with clinical image reports to determine if any lesions were missed or falsely identified.
Results: A total of 75 PET datasets of patients with suspected melanoma were identified and evaluated using AutoPERCIST. Disease was identified using AutoPERCIST in 51/75 cases. The mean total time to load, review, and extract parameters from all images was 4.96 minutes. Loading an image required a mean time of 1.32 minutes. The mean time to review an image and extract parameters was 3.64 minutes. Images without disease were analyzed significantly faster than those with disease (2.81 minutes vs. 5.97 minutes, respectively; p < 0.0001). Lesions were falsely identified on 2/24 negative images, corresponding to a 91.7% negative predictive value. Disease was unable to be detected on 6/51 positive images, corresponding to an 88.2% positive predictive value. Sensitivity and specificity were 95.7% and 78.6%, respectively.
Conclusions: Computer-Intelligence assisted image analysis is a rapidly advancing area of study, augmenting tedious manual analyses. This work shows the AutoPERCIST software can be deployed with clinically acceptable image analysis times, while maintaining a high degree of diagnostic accuracy. Images without disease, in particular, were analyzed at a high rate. A computer-assisted diagnostic system, such as AutoPERCIST, that could identify negative scans and provide quantitative assessments could vastly improve clinical PET reading workflows. More powerful computing could further accelerate the pace of image analysis for quantitative PET.
