Quantifying and Mitigating AI Diagnostic Discrepancies in MRI: A Dual-Model Machine Learning Framework

Background: Artificial intelligence (AI) has transformed medical imaging, particularly in magnetic resonance imaging (MRI), by enabling rapid and accurate pathology detection. However, diagnostic errors in AI systems remain insufficiently characterized, posing risks to clinical reliability. Objective: This study develops a dual-model machine learning (ML) framework to quantify, characterize, and mitigate AI diagnostic errors in MRI brain scans, enhancing reliability for clinical deployment. Methods: We analyzed ۲,۵۰۰ anonymized MRI brain scans from the BraTS ۲۰۲۱ and FastMRI datasets, covering gliomas, ischemic lesions, and multiple sclerosis (MS) plaques. A ۳D U-Net performed initial pathology segmentation, while a gradient-boosted decision tree (GBDT) model identified and categorized errors using features such as prediction confidence, lesion size, and scan quality. Metrics included Dice coefficient, precision, recall, area under the curve (AUC), and error-specific analyses. Results: The ۳D U-Net achieved a mean Dice score of ۰.۸۶ but exhibited a ۱۳.۲% false negative rate and ۷.۹% false positive rate, particularly for small lesions and periventricular white matter. The GBDT error detector achieved a ۰.۹۱ AUC, identifying ۸۲% of false negatives and revealing strong correlations between errors and low signal-to-noise ratio (SNR), small lesion size, and non-standard acquisition protocols. Conclusion: This dual-model approach effectively quantifies and mitigates AI diagnostic errors, offering a scalable quality control mechanism for clinical MRI workflows. Integrating such frameworks could enhance AI reliability and patient safety.

کلیدواژه ها:

Artificial Intelligence ، Machine Learning ، MRI ، Diagnostic Errors ، ۳D U-Net ، Error Detection ، Medical Imaging