Background:A priori identification of patients at risk of artificial intelligence (AI) failure in diagnosing cancer would contribute to the safer clinical integration of diagnostic algorithms.
Purpose:To evaluate AI prediction variability as an uncertainty quantification (UQ) metric for identifying cases at risk of AI failure in diagnosing cancer at MRI and CT across different cancer types, data sets, and algorithms.
Materials and Methods:Multicenter data sets and publicly available AI algorithms from three previous studies that evaluated detec-tion of pancreatic cancer on contrast-enhanced CT images, detection of prostate cancer on MRI scans, and prediction of pulmo-nary nodule malignancy on low-dose CT images were analyzed retrospectively. Each task's algorithm was extended to generate an uncertainty score based on ensemble prediction variability. AI accuracy percentage and partial area under the receiver operating characteristic curve (pAUC) were compared between certain and uncertain patient groups in a range of percentile thresholds (10%-90%) for the uncertainty score using permutation tests for statistical significance. The pulmonary nodule malignancy prediction algorithm was compared with 11 clinical readers for the certain group (CG) and uncertain group (UG).
Results:In total, 18 022 images were used for training and 838 images were used for testing. AI diagnostic accuracy was higher for the cases in the CG across all tasks (P < .001). At an 80% threshold of certain predictions, accuracy in the CG was 21%-29% higher than in the UG and 4%-6% higher than in the overall test data sets. The lesion-level pAUC in the CG was 0.25-0.39 higher than in the UG and 0.05-0.08 higher than in the overall test data sets (P < .001). For pulmonary nodule malignancy prediction, accuracy of AI was on par with clinicians for cases in the CG (AI results vs clinician results, 80% [95% CI: 76, 85] vs 78% [95% CI: 70, 87]; P = .07) but worse for cases in the UG (AI results vs clinician results, 50% [95% CI: 37, 64] vs 68% [95% CI: 60, 76]; P < .001).
Conclusion:An AI-prediction UQ metric consistently identified reduced performance of AI in cancer diagnosis.