Comparative Analysis of Four Commercially Available Artificial Intelligence Software for Brain Volumetry

Background/Objectives: Automated brain volumetry is increasingly integrated into the work-up of several neurological disorders, yet between-software variability may limit the interchangeability of absolute measures. This study aimed to assess inter-software variability, agreement, and interchangeability among four clinically available AI-based brain volumetry tools. Methods: We conducted a single-center, retrospective comparison of four clinically available tools (Neurophet AQUA^®, icobrain dm^®, mdbrain^®, Pixyl^®) applied to identical 3D T1-weighted MRI examinations (IR-FSPGR, 1 mm isotropic) acquired at 3 T in 47 adults referred for tremor evaluation. Absolute volumes (mL) were extracted for hippocampus, temporal, parietal, and frontal lobes, lateral ventricles, and total brain volume; the occipital lobe was analyzed where reported (three tools). Region-wise complete-case analyses were performed. Repeated-measures ANOVA with Bonferroni post hoc tests assessed systematic differences, and intraclass correlation coefficients (two-way mixed, consistency) were reported as single- and average-measure with 95% CIs to quantify agreement. Results: Across regions, all tests were significant (p < 0.001), indicating method-dependent scale offsets typically on the order of ~20-40%. A consistent pattern emerged: icobrain dm^® yielded the largest absolute volumes, Neurophet AQUA^® the lowest, mdbrain^® tended to be lower for parietal measurements, and Pixyl^® was generally intermediate. Agreement was structure-dependent: excellent for lateral ventricles (single-measure ICC ≈ 0.92 [0.85-0.96]; average-measure ICC ≈ 0.98 [0.96-0.99]); good-to-excellent for temporal and parietal lobes (average-measure ICC ≈ 0.83-0.84); moderate-to-good for hippocampus and total brain volume (TBV) (average-measure ICC ≈ 0.79-0.81), with only moderate single-measure agreement for TBV (ICC ≈ 0.51); and lowest for the frontal lobe (single-measure ICC ≈ 0.43 [0.23-0.65]; average-measure ICC ≈ 0.75 [0.54-0.88]). In the occipital lobe (three tools), Neurophet AQUA^® yielded substantially smaller values than icobrain dm^® and mdbrain^® (p < 0.0001). Conclusions: Overall, absolute volumetric outputs are not directly interchangeable across platforms, but relative subject ranking is largely preserved; platform-consistent analysis is advisable for longitudinal monitoring, and cross-site studies should implement calibration/harmonization before pooling data.