aetherAI, Co., Ltd., Taipei, Taiwan.
Department of Anatomic Pathology, Chang Gung Memorial Hospital at Linkou, Taoyuan, Taiwan.
Biomed J. 2022 Aug;45(4):675-685. doi: 10.1016/j.bj.2021.08.011. Epub 2021 Sep 8.
Classification of glomerular diseases and identification of glomerular lesions require careful morphological examination by experienced nephropathologists, which is labor-intensive, time-consuming, and prone to interobserver variability. In this regard, recent advance in machine learning-based image analysis is promising.
We combined Mask Region-based Convolutional Neural Networks (Mask R-CNN) with an additional classification step to build a glomerulus detection model using human kidney biopsy samples. A Long Short-Term Memory (LSTM) recurrent neural network was applied for glomerular disease classification, and another two-stage model using ResNeXt-101 was constructed for glomerular lesion identification in cases of lupus nephritis.
The detection model showed state-of-the-art performance on variedly stained slides with F1 scores up to 0.944. The disease classification model showed good accuracies up to 0.940 on recognizing different glomerular diseases based on H&E whole slide images. The lesion identification model demonstrated high discriminating power with area under the receiver operating characteristic curve up to 0.947 for various glomerular lesions. Models showed good generalization on external testing datasets.
This study is the first-of-its-kind showing how each step of kidney biopsy interpretation carried out by nephropathologists can be captured and simulated by machine learning models. The models were integrated into a whole slide image viewing and annotating platform to enable nephropathologists to review, correct, and confirm the inference results. Further improvement on model performances and incorporating inputs from immunofluorescence, electron microscopy, and clinical data might realize actual clinical use.
肾小球疾病的分类和肾小球病变的识别需要经验丰富的肾脏病病理学家进行仔细的形态学检查，这是一项劳动密集型、耗时且容易受到观察者间差异影响的工作。在这方面，基于机器学习的图像分析的最新进展很有前景。
我们结合基于掩模区域的卷积神经网络（Mask R-CNN）和额外的分类步骤，使用人类肾活检样本构建了肾小球检测模型。长短期记忆（LSTM）递归神经网络用于肾小球疾病分类，另一个使用 ResNeXt-101 的两阶段模型用于狼疮性肾炎病例中的肾小球病变识别。
该检测模型在不同染色的幻灯片上表现出了最先进的性能，F1 分数高达 0.944。疾病分类模型在基于 H&E 全幻灯片图像识别不同肾小球疾病方面表现出良好的准确性，最高可达 0.940。病变识别模型表现出较高的鉴别能力，ROC 曲线下面积高达 0.947，可用于各种肾小球病变。模型在外部测试数据集上具有良好的泛化能力。
这项研究首次表明，肾脏病病理学家进行的肾活检解释的每个步骤都可以通过机器学习模型来捕获和模拟。这些模型被整合到一个全幻灯片图像查看和注释平台中，使肾脏病病理学家能够对推理结果进行审查、纠正和确认。进一步提高模型性能并整合免疫荧光、电子显微镜和临床数据的输入，可能实现实际的临床应用。
