Unrecognized Biases and Validation Gaps in TraceOrg for Automated Autosomal Dominant Polycystic Kidney Disease Volumetry

The multicenter study by Xiong et al., published in JASN, introduced TraceOrg, a transformer-enhanced deep learning system for automated segmentation of kidneys, liver, and cysts in autosomal dominant polycystic kidney disease (ADPKD).1 Although the model achieved near–expert-level Dice similarity coefficients across diverse datasets, several underexplored insights and potential weaknesses warrant closer scrutiny. A central issue is the sequence-dependent volumetric bias among magnetic resonance imaging (MRI) pulse sequences. The mean absolute percent difference was 5.3% for T1-weighted imaging and 6.1% for T2-weighted imaging, decreasing to 4.9% when averaged.1 This variation matches the average annual increase in total kidney volume reported in ADPKD,1 indicating that inconsistent MRI sequence use across visits could mimic disease progression or stability. Such bias may shift the Mayo Imaging Class, thereby affecting therapeutic eligibility. Consistent with previous work showing that multisequence averaging improves reproducibility, harmonized or averaged MRI calibration should be incorporated into follow-up imaging to ensure longitudinal consistency.2 Segmentation performance also declined nonlinearly with cyst burden. Hepatic cyst Dice scores decreased from 0.90 for larger cysts above 50 ml to 0.76 for those below this threshold,1 indicating reduced sensitivity for early-stage or microcystic disease. This limitation is clinically significant because the small cyst burden is increasingly recognized as a precursor of hepatic involvement and differs by genotype and sex.3 Despite the high overall overlap metrics, large Hausdorff distances indicated local boundary deviations that could distort cyst morphology, potentially limiting the utility of emerging image-derived biomarkers related to cyst wall irregularity or hemorrhagic transformation. Furthermore, liver cyst reference masks were generated from model-assisted contours rather than fully independent manual annotations, introducing partial circular validation that may have inflated the reported accuracy. The absence of test-retest reproducibility and the limited demographic diversity of validation cohorts further constrain the clinical generalizability of these findings. Collectively, TraceOrg represents a major advance in automated ADPKD volumetry. However, robust clinical deployment will require standardized sequence calibration, independent ground truth validation, and longitudinal reproducibility testing to ensure accurate and interpretable patient-level monitoring.4