Uncertainty‑weighted learning to reduce diagnostic uncertainty, evaluate AI models, and augment precise clinical decisions

Abstract

Deep neural networks (DNNs) have been widely applied to medical image analysis; however, unreliable uncertainty quantification (UQ) remains a major barrier to safe clinical deployment. In a previous work, we proposed methods for uncertainty-based model evaluation and sample-level risk estimation, and implemented the methods using variance-based uncertainty quantified from a heteroscedastic binary cross-entropy (HBCE) model. These methods will benefit from better-calibrated uncertainty estimates. In this work, we proposed uncertainty-weighted learning (UWL) as a training strategy to improve the calibration of predictive uncertainty in medical imaging classification models. UWL leverages uncertainty scores produced by the HBCE model to adaptively emphasize more uncertain samples during training. Experiments on a breast cancer mammography dataset for a binary classification task demonstrated that UWL improves uncertainty calibration without degrading classification performance, and yielded more accurate prediction risk estimates under our uncertainty-based evaluation methods. These results highlighted the practical value of uncertainty-aware training for reliable model assessment and patient-level decision support, and established UWL as an effective approach for enhancing the trustworthiness of medical AI systems.

Ähnliche Arbeiten