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

Area of Science:

• Radiology
• Medical Imaging Analysis
• Artificial Intelligence in Healthcare

Background:

• Inter-reader variability in tumor measurements impacts RECIST-based response assessment.
• Automated segmentation and target selection systems exist, but predicting measurement uncertainty is a novel approach.

Purpose of the Study:

• To develop and validate a model that predicts lesion-specific measurement uncertainty.
• To provide a signal for measurement quality to aid in target selection and adjudication.

Main Methods:

• A Swin Transformer model was trained on 463 lesions from the Vol-PACT cohort with four segmentations per lesion.
• Inter-reader variability was quantified as (max diameter - min diameter) / min diameter.
• A threshold of > 0.20 was used to define high variability; preliminary external validation was performed.

Main Results:

• Internal test set achieved an AUC of 0.89, MAE of 0.10, accuracy of 82%, sensitivity of 72%, and specificity of 92% for low variability.
• External validation AUCs ranged from 0.77 to 0.92 across five readers, with wide confidence intervals due to small sample size.

Conclusions:

• The proof-of-concept model shows potential as a measurement-quality signal for clinical use.
• Further multi-institutional studies, calibration, and workflow validation are necessary before clinical implementation.