Weighted Metric Combination Predicts Radiologists' High Satisfaction with Thigh MRI AI Segmentations

Weighted metric combination predicted radiologist-assigned perfect satisfaction (Likert 5) for DL thigh MRI segmentations with 84.1% accuracy, 100% recall, AUC 0.824, outperforming any single metric.

• Post hoc analysis of 68 patients (mean age 29.4, 45% female) from a randomized trial; interreader agreement for Likert scores was near-perfect (κ=0.82).
• The DINOv2 model directly predicted Likert scores with lower error (MAE 0.2343) than MobileNetV2 (MAE 0.3488).
• Individual metrics (Dice, Jaccard) only weakly correlated with satisfaction (highest r=0.37), underscoring the need for combined metrics.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.