标题： 可信的图像到图像翻译：在无配对训练场景中评估不确定性校准 显示英文标题

标题： Trustworthy image-to-image translation: evaluating uncertainty calibration in unpaired training scenarios

摘要： 乳腺X线摄影筛查是检测乳腺癌的有效方法，有助于早期诊断。 然而，目前需要手动检查图像，给医疗系统带来了沉重的负担，促使人们渴望开发自动诊断协议。 基于深度神经网络的技术在一些研究中已被证明是有效的，但它们容易过拟合，导致泛化能力差和误诊的风险较大，阻碍了它们在临床环境中的广泛应用。 基于无配对神经风格迁移模型的数据增强方案已被提出，这些方案通过在没有配对训练数据（同一组织的不同图像风格）的情况下多样化训练图像特征的表示来提高泛化能力。 但这些模型同样容易出现各种病理问题，在没有真实标签/大型数据集的情况下（这在医学成像中很常见），评估它们的性能具有挑战性。 在这里，我们考虑了两种框架/架构：基于生成对抗网络的cycleGAN，以及最近开发的基于扩散的SynDiff。 我们在从三个公开访问的乳腺X线摄影数据集和一个非医学图像数据集中解析的图像块上训练它们，并评估其性能。 我们考虑使用不确定性量化来评估模型的可信度，并提出了一种在无配对训练场景中评估校准质量的方案。 这最终有助于在通常没有真实标签的领域中可信地使用图像到图像的翻译模型。 显示英文摘要

摘要： Mammographic screening is an effective method for detecting breast cancer, facilitating early diagnosis. However, the current need to manually inspect images places a heavy burden on healthcare systems, spurring a desire for automated diagnostic protocols. Techniques based on deep neural networks have been shown effective in some studies, but their tendency to overfit leaves considerable risk for poor generalisation and misdiagnosis, preventing their widespread adoption in clinical settings. Data augmentation schemes based on unpaired neural style transfer models have been proposed that improve generalisability by diversifying the representations of training image features in the absence of paired training data (images of the same tissue in either image style). But these models are similarly prone to various pathologies, and evaluating their performance is challenging without ground truths/large datasets (as is often the case in medical imaging). Here, we consider two frameworks/architectures: a GAN-based cycleGAN, and the more recently developed diffusion-based SynDiff. We evaluate their performance when trained on image patches parsed from three open access mammography datasets and one non-medical image dataset. We consider the use of uncertainty quantification to assess model trustworthiness, and propose a scheme to evaluate calibration quality in unpaired training scenarios. This ultimately helps facilitate the trustworthy use of image-to-image translation models in domains where ground truths are not typically available.