Deep Learning for Quantitative Perfusion 99mTc-ECD SPECT in Clinically Confirmed Neuropsychiatric Systemic Lupus Erythematosus: Comparison with Established Software

Abstract

Abstract Purpose We compared a novel deep learning (DL)-based quantification software (BTXBrain) with an established platform (Q.Brain) for quantitative perfusion single photon emission computed tomography (SPECT) in patients with clinically confirmed neuropsychiatric systemic lupus erythematosus (NPSLE). Methods This retrospective exploratory inter-platform comparison study included patients with clinically established NPSLE who underwent brain SPECT with technetium-99m ethyl cysteinate dimer ( 99m Tc-ECD). SPECT datasets were reprocessed using BTXBrain and Q.Brain. Regional perfusion estimates were compared across predefined cortical territories using paired analyses, variance ratio assessments, and Kendall’s tau correlation analysis. Results A total of 540 paired regional and hemispheric evaluations from 30 NPSLE patients were successfully processed and analyzed. BTXBrain and Q.Brain unveiled systematic, region-specific divergence in perfusion metrics. BTXBrain yielded higher standardized uptake value ratio values than Q.Brain in frontal (mean difference 0.081, p < 0.001), posterior cingulate (0.090, p < 0.001), medial temporal (0.333, p < 0.001), and occipital regions (0.078, p < 0.001). Conversely, a reversed trend was observed in the right lateral parietal cortex (-0.056, p = 0.009). Variance ratio analysis confirmed that the DL-based approach introduces a significantly different data dispersion profile ( p < 0.05 across multiple macro-regions), while meaningful cross-platform correlation was restricted to specific territories, such as the left posterior cingulate cortex (τ = 0.44, p = 0.002). Conclusion BTXBrain and Q.Brain are not directly interchangeable in patients with NPSLE. Consequently, rather than a generic drop-in replacement, the routine clinical adoption of DL-driven SPECT quantification requires software-specific reference thresholds and further calibration before cross-platform use.