Version comparison of AccuBrain for automated brain volumetry regarding accuracy, reproducibility and application for diagnosis

Abstract Background Automated brain volumetry has been widely used to assess brain volumetric changes that may indicate clinical states and progression. Among the tools that implement automated brain volumetry, AccuBrain has been validated for its accuracy, reliability and clinical applications for the older version (IV1.2). Recently, AccuBrain underwent a substantial revision update (IV2.0) for segmentation algorithms. Here, we aim to compare the performance of the two versions of AccuBrain for future use from several aspects. Method Public datasets with 3D T1‐weighted scans were included for version comparisons, each with Alzheimer’s disease (AD) patients and normal control (NC) subjects that were matched in age and gender (Table 1). For the comparisons of the brain volumetric measures quantified from the same scans, we investigated the difference of hippocampal segmentation accuracy (using Dice similarity coefficient [DSC] as the major measurement). As AccuBrain generates a composite index (AD resemblance atrophy index, AD‐RAI) that indicates similarity with AD‐like brain atrophy pattern, we also compared the two versions for the diagnostic accuracy of AD vs. NC with AD‐RAI. Also, we examined the intra‐scanner reproducibility of the two versions for the scans acquired with short‐intervals using intraclass correlation coefficient (ICC). Result Regarding hippocampal segmentation, AccuBrain IV2.0 presented significantly higher accuracy than IV1.2 with manual segmentation as the reference, both in the entire cohort (DSC: 0.91 vs. 0.89, p<0.001) and in the subgroups (all with p<0.001) defined by diagnosis, field strength or manufacturer (Table 2 and Figure 1). Also, AccuBrain IV2.0 showed better diagnostic accuracy of AD (AUC: 0.977 vs. 0.921, p<0.001) than IV1.2 (Figure 2). Regarding the analyses of intra‐scanner reproducibility (Table 3 and Table 4), the point estimates of ICC values were all >0.95 for absolute volumes and AD‐RAI, and generally >0.90 for relative volumetric measures. These two versions generally presented no significant differences for ICC values for most brain volumetric measures in NC or AD. Conclusion AccuBrain IV2.0 presented better segmentation accuracy and diagnostic accuracy of AD, and similar intra‐scanner reproducibility compared with IV1.2. Both versions should be feasible for use due to the small magnitude of differences.