%0 Journal Article %A Thomas Funck %A Kevin Larcher %A Paule-Joanne Toussaint %A Alan Evans %A Alexander Thiel %T Automating quality control for image processing pipelines with groupwise anomaly detection %D 2018 %J Journal of Nuclear Medicine %P 1728-1728 %V 59 %N supplement 1 %X 1728Objectives: The increasing availability of large brain imaging datasets makes automated analysis essential but requires rigorous quality control (QC) to ensure that each processing step has been performed as expected. We have implemented a novel technique for automated QC via a framework termed GRAD (GRoupwise Anomaly Detection) and evaluate this approach for the automated QC of PET analysis using APPIAN (Automated Pipeline for PET Image ANalysis), an open-source package available through github.com/APPIAN-PET/APPIAN or Docker Hub: tffunck/appian. The overall approach of GRAD is to define a set of QC metrics that quantify the performance of a given processing step and perform outlier detection on the empirical distribution these metrics. Here we assess GRAD for PET-MRI coregistration and tracer kinetic analysis. Methods: 26 [18-F]-flumazenil PET images were acquired on the Siemens HRRT scanner and reconstructed with 3D-FBP. Corresponding T1 MRI were acquired on a Siemens Magnetom TrioTim Syngo MR using an MPRAGE sequence (repetition time (TR) 2300 ms, echo time (TE) 2.98 ms, TI 9 ms and flip angle = 9° matrix size = 160 × 256 × 256).To test the performance of GRAD for coregistration, a set of correctly co-registered PET and MRI were created by running APPIAN. Each of the correctly co-registered PET images were systematically misaligned by applying rotations of 2, 4, 8, 12, 16, 20 degrees in the axial plane. 3 QC metrics were used to quantify the accuracy with which the PET image was coregistered with the MRI: cross-correlation (CC), mutual information (MI), and feature-space entropy (FSE), as well as the combination of all three measures (All). The performance of GRAD for tracer kinetic analysis was evaluated by introducing error into the calculation of BPnd for each of the images using the Logan plot method [1]. This was done by simulating the misspecification of the reference region for the Logan plot, such that the reference region was progressively contaminated with radiotracer concentration from regions with specific radiotracer binding. That is, a pseudo-reference time activity curve (TAC) was created : pseudo-reference TAC = true-reference TAC x (α) + contaminating TAC x (1-α), where α ∊ (0, 0.25, 0.5, 0.75, 1). The reference region was defined as average TAC of the white matter and the contaminating TAC was the average of the radiotracer concentration in the cortical grey matter. Outlier measures were detected using the local outlier factor (LOF)[2], kernel density estimation (KDE), isolation forest [3], and median average deviation (MAD). The area under the curve (AUC) of the ROC curves for each outlier measure was calculated to compare the performance of GRAD at various levels of error. Results: GRAD was able to detect moderate misregistrations in the PET and MRI images. Overall, the best combination of QC metrics and outlier measures for coregistration was to use of all metrics with KDE (AUC=0.8 for an 8° rotation error; Fig.1). In the case of tracer kinetic analysis, all outlier measures performed similarly for detecting outliers in the distribution of BPnd (AUC=~0.75 for a 40% contamination of the reference TAC). Conclusions: GRAD is a general framework for automated QC that is able to detect errors in common processing steps used in PET image analysis, specifically PET-MRI coregistration and tracer kinetic analysis. APPIAN implements GRAD to facilitate the reliable automation of research with PET imaging. Future work will involve extending GRAD to more processing steps, e.g., partial-volume correction, evaluating its performance with more radiotracers, and testing additional outlier measures. References[1] Logan, et al. 1996. J. Cereb. Blood Flow Metab. 16, 834-840.[2] Breunig, et al. 2000. Proc. 2000 Acm Sigmod Int. Conf. Manag. Data 1-12. .[3] Liu, F.T., Ting K.M., Zhou, Z.H. Data Mining, 2008. ICDM'08. Eighth IEEE Int Conf. 2008. 413-422. %U