Deep Learning Quality Control for RECIST-Oriented Assessment: A Vision Transformer Predicts Inter-Reader Variability...

Background/Objectives: Inter-reader variation can alter unidimensional tumor measurements used in RECIST-oriented response assessment. Unlike systems that automate segmentation or target selection, this study aimed to predict lesion-specific measurement uncertainty itself. Methods: The development cohort comprised 463 lung, liver, and lymph-node lesions from 280 patients in Vol-PACT, with four segmentations per lesion. Inter-reader variability was defined as (maximum longest diameter-minimum longest diameter)/minimum longest diameter. A Swin Transformer was trained to regress this continuous score. A value > 0.20 was used as a pragmatic high-variability alert threshold, with high variability designated as the positive class; this threshold is not equivalent to RECIST progressive disease. Preliminary external evaluation used 21 NSCLC lesions with five reader contours, providing one reader mask at a time. Results: The validation and internal test AUCs were 0.74 (95% CI, 0.61-0.85) and 0.89 (95% CI, 0.80-0.95), respectively. On the internal test set, MAE was 0.10 (95% CI, 0.08-0.12), accuracy was 82% (95% CI, 74-90%), sensitivity 72% (95% CI, 59-85%), and specificity for lower-variability lesions was 92% (95% CI, 82-100%). Across the five external readers, AUCs ranged from 0.77 to 0.92; the wide confidence intervals reflect the small external sample. Conclusions: This proof-of-concept model may provide a measurement-quality signal that supports target selection or adjudication. Larger multi-institutional studies, patient-level resampling, calibration, ablation testing, and workflow validation are required before clinical use.