Diagnostic performance of artificial intelligence for histologic melanoma recognition compared to 18 international expert pathologists: Supplementary Material

The study was designed to compare the performance of classifiers based on image analyses by convolutionnal neural networks (CNNs) with that of 18 expert dermatopathologists in a binary classification task for pigmented skin lesions. Mendeley Supplementary Figure 1 shows a schematic representation of the testing approach. The CNNs were trained by cross-testing. Each iteration consists of five folds. Orange rectangles represent the folds used for testing and blue rectangles represent folds for training. For each iteration a CNN is trained and tested on the respective folds. Each trained CNN has a performance that is determined on the fold. To calculate the overall performance, the sum of all 5 performances is taken. This procedure was repeated 3 times and the results were combined to generate an ensemble. Mendeley Supplementary Figure 2 depicts the whole slide image (WSI) analysis. Tissue sections on whole slide images (top) were divided into tiles (middle). The CNN (a pre-trained ResNeXt50) assigned a malignancy score to every individual tile (bottom). Red tiles were classified as melanoma, blue tiles as nevus. Scores for all tiles on one image were averaged to a final malignancy score for the complete slide. Supplementary Table 1 shows the characteristics of the pigmented skin lesions included in the test set. Supplementary Table 2 provides an analysis of statistical differences between the performance of pathologists and CNN classifiers.本研究旨在对比基于卷积神经网络（Convolutional Neural Networks, CNNs）图像分析的分类器，与18名皮肤病理专家在色素性皮肤病变二分类任务中的表现。门德雷补充图1展示了该测试方法的示意图。本研究采用交叉验证方式训练卷积神经网络，每一轮迭代包含5个折。橙色矩形代表用于测试的折，蓝色矩形代表用于训练的折。在每一轮迭代中，卷积神经网络将在对应折上完成训练与测试，其性能将基于该折进行评估。为计算整体性能，需对5个折的性能结果求和。上述流程重复3次后，将所得结果整合以构建集成模型。门德雷补充图2展示了全玻片图像（Whole Slide Image, WSI）分析流程：全玻片图像（上图）中的组织切片会被分割为图像块（中图）。本研究采用预训练的ResNeXt50卷积神经网络为每个独立图像块分配恶性评分，其中红色图像块被分类为黑色素瘤，蓝色图像块被分类为痣。将单张全玻片图像上所有图像块的评分取平均，即可得到该玻片的最终恶性评分。补充表1展示了测试集所纳入的色素性皮肤病变的特征信息；补充表2则对病理专家与卷积神经网络分类器的性能差异进行了统计学分析。