我们研究了是否可以使用人工智能 (AI) 自动确定 Johnsen 评分的工具代替传统的 Johnsen 评分来支持病理学家的评估。平均准确率、准确率和召回率由 Google Cloud AutoML Vision 平台评估。我们获得了 275 名患者的睾丸组织，并且能够使用来自 264 名患者的苏木精和伊红 (H&E) 染色的玻璃显微镜载玻片。此外，我们切出部分组织病理学图像 (5.0 × 5.0 cm) 以扩大约翰森的特征区域与生精小管。我们定义了四个标签：Johnsen 评分 1-3、4-5、6-7 和 8-10，以在临床实践中区分 Johnsen 评分。所有图像都上传到 Google Cloud AutoML Vision 平台。我们获得了一个包含 7155 张放大 400 倍图像的数据集和一个包含 9822 张扩展图像的数据集，用于 5.0 × 5.0 厘米的切口。对于400倍放大图像数据集，该算法的平均准确率（阳性预测值）为82.6%，准确率为80.31%，召回率为60.96%。对于扩展图像数据集（5.0 × 5.0 cm），平均精度为 99.5%，精度为 96.29%，召回率为 96.23%。这是用于预测 Johnsen 分数的基于 AI 的算法的第一份报告。

"点击查看英文标题和摘要"

We examined whether a tool for determining Johnsen scores automatically using artificial intelligence (AI) could be used in place of traditional Johnsen scoring to support pathologists’ evaluations. Average precision, precision, and recall were assessed by the Google Cloud AutoML Vision platform. We obtained testicular tissues for 275 patients and were able to use haematoxylin and eosin (H&E)-stained glass microscope slides from 264 patients. In addition, we cut out of parts of the histopathology images (5.0 × 5.0 cm) for expansion of Johnsen’s characteristic areas with seminiferous tubules. We defined four labels: Johnsen score 1–3, 4–5, 6–7, and 8–10 to distinguish Johnsen scores in clinical practice. All images were uploaded to the Google Cloud AutoML Vision platform. We obtained a dataset of 7155 images at magnification 400× and a dataset of 9822 expansion images for the 5.0 × 5.0 cm cutouts. For the 400× magnification image dataset, the average precision (positive predictive value) of the algorithm was 82.6%, precision was 80.31%, and recall was 60.96%. For the expansion image dataset (5.0 × 5.0 cm), the average precision was 99.5%, precision was 96.29%, and recall was 96.23%. This is the first report of an AI-based algorithm for predicting Johnsen scores.