Immunohistochemistry (IHC) is a diagnostic technique used throughout pathology. A machine learning algorithm that could predict individual cell immunophenotype based on hematoxylin and eosin (H&E) staining would save money, time, and reduce tissue consumed. Prior approaches have lacked the spatial accuracy needed for cell-specific analytical tasks. Here IHC performed on destained H&E slides is used to create a neural network that is potentially capable of predicting individual cell immunophenotype. Twelve slides were stained with H&E and scanned to create digital whole slide images. The H&E slides were then destained, and stained with SOX10 IHC. The SOX10 IHC slides were scanned, and corresponding H&E and IHC digital images were registered. Color-thresholding and machine learning techniques were applied to the registered H&E and IHC images to segment 3,396,668 SOX10-negative cells and 306,166 SOX10-positive cells. The resulting segmentation was used to annotate the original H&E images, and a convolutional neural network was trained to predict SOX10 nuclear staining. Sixteen thousand three hundred and nine image patches were used to train the virtual IHC (vIHC) neural network, and 1,813 image patches were used to quantitatively evaluate it. The resulting vIHC neural network achieved an area under the curve of 0.9422 in a receiver operator characteristics analysis when sorting individual nuclei. The vIHC network was applied to additional images from clinical practice, and was evaluated qualitatively by a board-certified dermatopathologist. Further work is needed to make the process more efficient and accurate for clinical use. This proof-of-concept demonstrates the feasibility of creating neural network-driven vIHC assays.

中文翻译：

---

用于模拟免疫组织化学的机器学习算法：SOX10 虚拟 IHC 的开发和对原发性黑素细胞肿瘤的评估。

免疫组织化学 (IHC) 是一种贯穿病理学的诊断技术。一种可以根据苏木精和伊红 (H&E) 染色预测个体细胞免疫表型的机器学习算法将节省金钱、时间并减少组织消耗。先前的方法缺乏细胞特异性分析任务所需的空间精度。这里，在脱色的 H&E 载玻片上进行的 IHC 用于创建一个可能能够预测个体细胞免疫表型的神经网络。将 12 张载玻片用 H&E 染色并扫描以创建数字完整载玻片图像。然后将 H&E 载玻片脱色，并用 SOX10 IHC 染色。扫描 SOX10 IHC 载玻片，并配准相应的 H&E 和 IHC 数字图像。将颜色阈值和机器学习技术应用于注册的 H&E 和 IHC 图像，以分割 3,396,668 个 SOX10 阴性细胞和 306,166 个 SOX10 阳性细胞。由此产生的分割用于注释原始 H&E 图像，并训练卷积神经网络来预测 SOX10 核染色。使用 16309 个图像块来训练虚拟 IHC (vIHC) 神经网络，并使用 1,813 个图像块对其进行定量评估。在对单个细胞核进行排序时，在接受者操作员特征分析中，所得 vIHC 神经网络的曲线下面积达到了 0.9422。vIHC 网络应用于临床实践中的其他图像，并由委员会认证的皮肤病理学家进行定性评估。需要进一步的工作来使该过程在临床使用中更加高效和准确。这一概念验证证明了创建神经网络驱动的 vIHC 检测的可行性。