Assessment and Optimisation of Hermes Amyloid BRASS as a Quantitative Diagnostic Tool in Reporting ¹⁸F-florbetapir (Amyvid) Investigations

Purpose: Amyloid PET/CT plays an important role in investigation of patients with cognitive impairment and dementia, especially those with atypical and mixed clinical presentations. The aim of this work was to assess and optimise the ability of Hermes Amyloid BRASS software to identify positive and negative amyloid scans using default BRASS values and published literature values. Materials and

Methods: Data from 100 patients (41-88 years) injected with 370MBq 18F-florbetapir, and imaged for 20 minutes following a 40 minute uptake period, was evaluated using Hermes BRASS software, and compared against the final clinical report. Frames that included excessive motion were removed to create at least 10 minutes of data. To aid reporting of examinations the scans were classified as type A (typical features) or non-type A (atypical features) for both positive and negative scans according to imaging characteristics. The binary result (positive or negative) and sub-group classification for each scan was identified in the double reported original clinical report (independent review by two trained readers) and this was chosen as the reference standard. Of the 100 scans 79 were classified as Type A, and 21 non-Type A. All 100 studies were independently reviewed by three experienced reviewers and quantitatively evaluated in Amyvid BRASS (Hermes Medical Imaging Solutions). Regional amyloid uptake ratios (SUVr), mean cortical SUVr values (mcSUVr) and z-score (number of standard deviations from the healthy control SUVr in the BRASS template) were calculated relative to the cerebellum. Two separate classification metrics were used for positive scans: z>2 for calculated mcSUVr relative to the template mcSUVr and z>2 for ROI≥2 (two-region classification) for individual ROI SUVr values relative to template ROI SUVr values. ROC curves were used to define potential SUVr thresholds for BRASS to optimise true positive rate (TPR, sensitivity) and false positive rate (FPR, 1-specificity) and accuracy. Results: Accuracy for the three readers for all data was 90%, default BRASS settings yielded 85% and ROC optimised mcSUVr threshold yielded 85%. For type A only studies reader accuracy increased to 96%, BRASS default and ROC optimised mcSUVr thresholds both yielded 92%. The optimum point on the ROC curve corresponded to 1.18 for type A data, this is approximately equal to the BRASS template mcSUVr+2SD=1.17, and also equivalent to mcSUVr values derived by other means in published literature. Threshold mcSUVr = 1.11 is also frequently referred to in literature, for our dataset this threshold yielded accuracy = 87% i.e. poorer than ROC derived mcSUVr threshold found in this study. Equivalent accuracy (92%) can also be achieved by using the two-region method with ROC optimised ROI SUVr thresholds. For non-Type A scans reader accuracy fell to 70% (range 48-95%) and agreement ranged from very poor to poor. This suggests that readers require further support to assist with reporting non-type A studies. Mean cortical measures may be more difficult to apply to these scans due to presence of one or more focal regions of uptake, increased noise, motion artefacts and ventricular atrophy amongst other features. Moreover, a larger sample size is necessary to obtain definitive optimisation of mcSUVr and individual ROI SUVr. Conclusion: This study has shown a clear definition between inter-reader agreement and reporting accuracy of type A and non-type A florbetapir scans. Other studies which have investigated the added effect of quantification on a visual read have not considered this separation of studies. Separation of type A scans improves on measured values of inter-reader agreement and reader sensitivity, specificity and accuracy. In BRASS, mcSUVr threshold=1.17, or two-region classification with optimised ROI SUVr thresholds, improves sensitivity, specificity and accuracy of the software for type A scans compared to all data as a single group.