不同扫描仪构建的结直肠癌全切片数字病理图像中人工标注迁移的研究

目的研究结直肠癌人工智能病理诊断模型构建过程中，病理医师对数字切片癌组织的人工标注在不同扫描仪构建的全切片图像（WSI）中准确迁移的方法。 方法在本研究中，我们提出了一种基于图像配准的标注迁移方法，在来自不同扫描仪的WSI之间建立仿射映射。通过多分辨率最小化两个WSI缩略图之间的互信息来估计最佳仿射映射参数，以避免和改变扫描仪特定特性的影响，减少计算时间。我们使用了181张结直肠癌病理切片，使用两个品牌的扫描仪获得相应的WSI，对上述标注迁移方法进行测试。 结果181张HE切片的扫描结果表明，同一张切片由不同扫描仪构建的WSI在颜色、位置、大小等属性上都有不同的表现。使用我们提出的标注迁移方法，其中179张图像的人工标注成功地在不同扫描仪构建的WSI中迁移，其中125对使用单个CPU核心的计算时间不到1分钟。 结论我们提出了一种快速、准确的全自动的标注迁移方法，用于在不同扫描仪构建的WSI之间传递人工标注。在准备深度学习训练数据过程中，既可以避免病理医师对新图像的重新标注，也可以避免病理医师之间在标注上的差异。

Transfer of manual annotation in digital pathological images of colorectal cancer with different scanners

ObjectiveTo create an accurate transfer method of the digital slide of cancer tissue annotated by pathologist in the whole slide image (WSI) constructed by different scanners in the process of constructing the artificial intelligence pathological diagnosis model of colorectal cancer. MethodsIn this study, we proposed a annotation transfer method based on image registration to establish affine mapping between WSIs from different scanners. The best affine mapping parameters are estimated by minimizing the mutual information between the two WSI thumbnails to avoid and change the specific characteristics of the scanner and reduce the calculation time. We used 181 colorectal cancer pathological sections and two brands of scanners to obtain the corresponding WSI to test the above annotation transfer method. ResultsThe scanning results of 181 H&E slides showed that WSI constructed by different scanners in the same slide had different performance in color, position, size and other attributes. Using our proposed annotation transfer method, a total of 179 images were successfully transferred between WSIs constructed by different scanners, and 125 pairs of them took less than 1 minute to compute using a single CPU core. ConclusionWe propose a fast and accurate automatic annotation transfer method, which is used to transfer manual annotation between WSIs constructed by different scanners. In the process of preparing deep learning training data, we can not only avoid the new image reannotation by pathologists, but also avoid the difference in annotation between pathologists.