Arega Tewodros Weldebirhan, Bricq Stéphanie, Meriaudeau Fabrice
ImViA Laboratory, Université de Bourgogne, Dijon, France.
ICMUB, Université de Bourgogne, Dijon, France.
Comput Med Imaging Graph. 2025 Jan;119:102476. doi: 10.1016/j.compmedimag.2024.102476. Epub 2024 Dec 12.
In real-world scenarios, medical image segmentation models encounter input images that may deviate from the training images in various ways. These differences can arise from changes in image scanners and acquisition protocols, or even the images can come from a different modality or domain. When the model encounters these out-of-distribution (OOD) images, it can behave unpredictably. Therefore, it is important to develop a system that handles such out-of-distribution images to ensure the safe usage of the models in clinical practice. In this paper, we propose a post-hoc out-of-distribution (OOD) detection method that can be used with any pre-trained segmentation model. Our method utilizes multi-scale representations extracted from the encoder blocks of the segmentation model and employs Mahalanobis distance as a metric to measure the similarity between the input image and the in-distribution images. The segmentation model is pre-trained on a publicly available cardiac short-axis cine MRI dataset. The detection performance of the proposed method is evaluated on 13 different OOD datasets, which can be categorized as near, mild, and far OOD datasets based on their similarity to the in-distribution dataset. The results show that our method outperforms state-of-the-art feature space-based and uncertainty-based OOD detection methods across the various OOD datasets. Our method successfully detects near, mild, and far OOD images with high detection accuracy, showcasing the advantage of using the multi-scale and semantically rich representations of the encoder. In addition to the feature-based approach, we also propose a Dice coefficient-based OOD detection method, which demonstrates superior performance for adversarial OOD detection and shows a high correlation with segmentation quality. For the uncertainty-based method, despite having a strong correlation with the quality of the segmentation results in the near OOD datasets, they failed to detect mild and far OOD images, indicating the weakness of these methods when the images are more dissimilar. Future work will explore combining Mahalanobis distance and uncertainty scores for improved detection of challenging OOD images that are difficult to segment.
在实际应用场景中，医学图像分割模型会遇到与训练图像在多种方面存在差异的输入图像。这些差异可能源于图像扫描仪和采集协议的变化，甚至图像可能来自不同的模态或领域。当模型遇到这些分布外（OOD）图像时，其行为可能无法预测。因此，开发一个能够处理此类分布外图像的系统，以确保模型在临床实践中的安全使用非常重要。在本文中，我们提出了一种事后分布外（OOD）检测方法，该方法可与任何预训练的分割模型一起使用。我们的方法利用从分割模型的编码器块中提取的多尺度表示，并采用马氏距离作为度量来衡量输入图像与分布内图像之间的相似度。分割模型在一个公开可用的心脏短轴电影磁共振成像（MRI）数据集上进行预训练。所提出方法的检测性能在13个不同的OOD数据集上进行评估，这些数据集根据与分布内数据集的相似度可分为近、中、远OOD数据集。结果表明，我们的方法在各种OOD数据集上优于基于特征空间和基于不确定性的现有OOD检测方法。我们的方法以高检测准确率成功检测出近、中、远OOD图像，展示了使用编码器的多尺度和语义丰富表示的优势。除了基于特征的方法，我们还提出了一种基于骰子系数的OOD检测方法，该方法在对抗性OOD检测中表现出卓越性能，并且与分割质量具有高度相关性。对于基于不确定性的方法，尽管在近OOD数据集中与分割结果的质量有很强的相关性，但它们未能检测出中、远OOD图像，这表明当图像差异更大时这些方法存在局限性。未来的工作将探索结合马氏距离和不确定性分数，以改进对难以分割的具有挑战性的OOD图像的检测。
