自动CT胰腺细分的基准测试稳定性:通过循环中人优化实现人类级别的可靠性
Felipe Oviedo1, Felipe Lopez-Ramirez2, Florent Tixier2
1AI for Good Lab, Microsoft Corporation, Redmond, WA, 98052, United States.
Radiology advances
|December 15, 2025
概括
对于胰腺细分的深度学习模型有希望,但缺乏稳定性. 积极学习显著提高了可靠性,实现了接近人类的性能,并减少了工作量.
科学领域:
- 医学成像分析 医学成像分析
- 医疗保健中的人工智能
- 计算解剖学的计算解剖学
背景情况:
- 在CT扫描中用于胰腺细分的深度学习模型已经迅速发展.
- 当前的评估指标 (例如,子,表面指标) 并不能完全捕捉模型的稳定性,定义为在不同情况下一致的人类水平的性能.
- 细分强度对于临床应用,如早期检测和定量生物标志物分析至关重要.
研究的目的:
- 系统地评估与人类读者相比,对于胰腺细分的深度学习模型的稳定性.
- 调查积极学习策略对提高细分可靠性的有效性.
主要方法:
- 追溯分析903个CT扫描,其中100个健康的测试病例,每个病例有4个独立的人类细分.
- 引入一个分数值 (FT) 度量来量化与人类性能相对的稳定性.
- 评估各种深度学习模型,并实施积极学习方法,以对不确定的预测进行人为循环修订.
主要成果:
- 最好的3D U-Net模型实现了高重叠指标 (DSC: 0.88,NSD: 0.77),与人类读者 (DSC: 0.89,NSD: 0.75) 相比.
- 然而,分数值 (FT) 度量显示与人类表现相比,模型的持续变化.
- 积极学习与人类在循环中修改大大提高了稳定性 (FT 到 0.99),每例时间投资最小 (1.54 分钟),工作量减少了 23 倍.
结论:
- 自动胰腺细分可以减少工作负载,但在具有挑战性的情况下,由于不可预测的故障而受到限制.
- 积极学习策略对于提高模型可靠性和弥合人工智能和人类专家之间的绩效差距至关重要.
- 整合主动学习是迈向医疗图像细分的强大和临床部署AI的重要一步.
相关概念视频
Imaging Studies for Cardiovascular System V: CT
Cardiac computed tomography (CT) scanning is an advanced cardiac imaging technique that utilizes CT technology, with or without intravenous (IV) contrast, to produce accurate cross-sectional virtual slices of specific areas of the heart, coronary circulation, and major blood vessels such as the aorta, pulmonary veins, and arteries. The computer processes these slices to generate three-dimensional images. Multidetector CT (MDCT) is a rapid form of CT scanning that captures multiple slices...
Multiple Comparison Tests
Multiple comparison test, abbreviated as MCT, is a post hoc analysis generally performed after comparing multiple samples with one or more tests. An MCT will help identify a significantly different sample among multiple samples or a factor among multiple factors.
It would be easy to compare two samples using a significance alpha level of 0.05. In other words, there is only one sample pair to be compared. However, it would be difficult to identify a significantly different sample if the number...
It would be easy to compare two samples using a significance alpha level of 0.05. In other words, there is only one sample pair to be compared. However, it would be difficult to identify a significantly different sample if the number...


