DeepCertainty: Deep Learning for Contextual Diagnostic Uncertainty Measurement in Radiology Reports

PE成为心血管相关死亡的第三大原因，超过50万例PE 每年在美国发生，导致大约20万人死亡和住院 超过25万名患者。快速、准确地诊断PE是确保最高 护理质量。每年，美国1.2亿急诊科(ED)患者中有1%-2%的人 肺栓塞的CT肺血管造影(CTPA)检查。转诊医生严重依赖CTPA 诊断或排除PE的报告。放射学报告的清晰度是最关键的品质之一， 美国放射学会强调了在放射报告中进行精确沟通的必要性。 然而，在放射学报告中有效地传达不确定性是具有挑战性的。转诊医生可以 解释放射科医生表达诊断信心的文字表达与预期的不同。差距 放射科医生的意向信息和转诊医生的解释之间不能完全一致 通过结构化报告或标准化词典解决。CTPA报告中不必要的模糊语言 可能会进一步加剧报告的模糊性，并可能导致对患者的不当治疗。因此，我们 目的开发一种基于深度学习的上下文感知(UN)确定性评估方法(DeepCertainty)， 这是端到端可培训、可校准、可泛化、可扩展和可解释的。它将允许罚款- 粒度不确定度测量和标准化，有助于一致和准确的诊断确定性 在CTPA报告中进行沟通，从而改善体育护理。这项研究将为今后的工作奠定基础 实施DeepCertainty并将其集成到临床工作流中，以提示实时低确定性警报 为了提高PE诊断报告的质量和清晰度，这将为ED提供更好的治疗决策 疑似PE患者。

PE becomes the third leading cause of cardiovascular-related death, and more than 500,000 cases of PE occur in the United States (US) every year, resulting in approximately 200,000 deaths and hospitalization of over 250,000 patients. Rapid and accurate diagnosis of PE are of paramount importance to ensure the highest quality of care. Every year 1-2% of the 120 million emergency department (ED) patients in the US undergo computed tomographic pulmonary angiography (CTPA) for PE. The referring physicians rely heavily on CTPA reports diagnosing or excluding PEs. Clarity of the radiology report is one of the most critical qualities, and the American College of Radiology has emphasized a need for precision communication in radiological reports. Yet communicating uncertainty effectively in radiology reports is challenging. Referring physicians may interpret radiologists’ textual expressions that convey diagnostic confidence differently than intended. The gap between radiologists’ intended message and the referring physicians’ interpretation can not be completely resolved through structured reporting or standardized lexicon. Unnecessary hedging language in CTPA reports may further worsen the reporting ambiguity and may lead to inappropriate treatment of patients. Therefore, we aim to develop a deep learning-based approach for context-aware (un)certainty assessment (DeepCertainty), which is end-to-end trainable, calibratable, generalizable, scalable, and explainable. It would allow for fine- grained uncertainty measurement and standardization, facilitate consistent and accurate diagnostic certainty communication in CTPA reports and thus improve PE care. This study will build the foundation for future implementation and integration of DeepCertainty into clinical workflows to prompt real-time low-certainty alerts for improving PE diagnostic reporting quality and clarity, which will inform better treatment decisions for ED patients with suspected PE.