DGUQA: Domain Generalization Uncertainty Informed Patient-Specific Quality Assurance

Deep Learning Automated Patient-Specific Quality Assurance PSQA endeavors to diminish the reliance on clinical resources. The accurate estimation of the dose difference metric, particularly the Gamma passing rate, is paramount in ensuring the safety and efficacy of radiation therapy plans. Although current research has yielded an overall performance on par with that of experts, it fails to address the local performance discrepancies of the model across diverse lesions, thereby highlighting a generalization challenge that undermines its credibility in real clinical settings.
This paper introduces DGUQA, based on the theory of domain generalization in deep learning. DGUQA employs an adversarial loss-based regularization to address the issue of generalization. Further, since the model is biased with the most common lesion organs, relying solely on a domain-generalized model would decrease overall performance. Therefore, in conjunction with safety requirements, we also model predictive uncertainty. The domain generalization model is used only when the uncertainty exceeds a certain threshold otherwise, a standard model is employed. Experiments demonstrate that DGUQA shows superiority in both generalization performance and overall effectiveness. DGUQA notably enhances the deep learning trustworthiness in the PSQA and has meaningful implications for the clinical significance of medical deep learning.