医疗图像分割中的可靠不确定性的平均校准损失
IEEE transactions on medical imaging
|March 11, 2026
概括
我们引入了平均校准错误 (ACE) 损失,以提高医疗图像细分的深度学习模型的可靠性. 这种方法可以提高模型校准,而不会显著影响细分精度,有助于临床整合.
科学领域:
- 医学图像分析 医学图像分析
- 机器学习 机器学习
- 计算机视觉 计算机视觉
背景情况:
- 医学成像中的深度神经网络 (DNN) 经常表现出过度自信,降低了它们的可靠性和临床适用性.
- 校准,预测的信心与真实的准确性对齐,对于医疗保健中可靠的AI至关重要.
研究的目的:
- 开发和评估边际L1平均校准误差 (mL1-ACE) 的可微分配方,作为辅助损失函数.
- 改进用于医疗图像分割的深度神经网络中的像素智能校准.
- 为控制模型校准和细分精度之间的权衡提供一种方法.
主要方法:
- 建议微分 mL1-ACE 作为每图像辅助损失.
- 对比硬binning和软binning方法用于像素智能校准.
- 在 ACDC,AMOS,KiTS 和 BraTS 数据集上评估性能.
- 引入了数据集可靠性历史图,用于分析校准变化.
主要成果:
- 在数据集中,mL1-ACE显著降低了平均校准误差 (ACE) 和最大校准误差 (MCE).
- 固体的mL1-ACE保持了高的子相似系数 (DSC),同时改善了校准.
- 软体mL1-ACE显示了更大的校准改进,但有时会损害细分性能.
- 数据集可靠性历史图显示,预测的可靠性与真实准确性之间有更好的对齐.
结论:
- 可差分的mL1-ACE有效地增强了用于医疗图像细分的DNN校准.
- 硬包装和软包装之间的选择允许平衡校准收益和细分性能.
- 提出的方法为从业者提供了对可靠性的更好控制,促进AI工具的临床采用.
相关概念视频
Uncertainty in Measurement: Accuracy and Precision
112.9K
Scientists typically make repeated measurements of a quantity to ensure the quality of their findings and to evaluate both the precision and the accuracy of their results. Measurements are said to be precise if they yield very similar results when repeated in the same manner. A measurement is considered accurate if it yields a result that is very close to the true or the accepted value. Precise values agree with each other; accurate values agree with a true value.
112.9K
Instrument Calibration
983
Instrument calibration is essential for ensuring that instruments produce accurate and consistent results. It is vital in manufacturing, healthcare, testing laboratories, and scientific research. Calibration processes are specific to each instrument and help enhance data accuracy. Each instrument has a unique calibration process tailored to its design and function to improve data accuracy.
Analytical Balance Calibration
An analytical balance measures mass and requires regular calibration to...
Analytical Balance Calibration
An analytical balance measures mass and requires regular calibration to...
983


