BACKGROUND Deep learning has great potential in digital kidney pathology. However, its effectiveness depends heavily on the availability of extensively labeled datasets, which are often limited due to the specialized knowledge and time required for their creation. This limitation hinders the widespread application of deep learning for the analysis of kidney biopsy images. METHODS We applied self-distillation with no labels (DINO), a self-supervised learning method, to a dataset of 10,423 glomerular images obtained from 384 PAS-stained kidney biopsy slides. Glomerular features extracted from the DINO-pretrained backbone were visualized using principal component analysis (PCA). We then performed classification tasks by adding either k-nearest neighbor (kNN) classifiers or linear head layers to the DINO-pretrained or ImageNet-pretrained backbones. These models were trained on our labeled classification dataset. Performance was evaluated using metrics such as the area under the receiver operating characteristic curve (ROC-AUC). The classification tasks encompassed four disease categories (minimal change disease, mesangial proliferative glomerulonephritis, membranous nephropathy, and diabetic nephropathy) as well as clinical parameters such as hypertension, proteinuria, and hematuria. RESULTS PCA visualization revealed distinct principal components corresponding to different glomerular structures, demonstrating the capability of the DINO-pretrained backbone to capture morphological features. In disease classification, the DINO-pretrained transferred model (ROC-AUC = 0.93) outperformed the ImageNet-pretrained fine-tuned model (ROC-AUC = 0.89). When the labeled data were limited, the ImageNet-pretrained fine-tuned model's ROC-AUC dropped to 0.76 (95% confidence interval [CI], 0.72-0.80), whereas the DINO-pretrained transferred model maintained superior performance (ROC-AUC 0.88, 95% CI 0.86-0.90). The DINO-pretrained transferred model also exhibited higher AUCs for the classification of several clinical parameters. External validation using two independent datasets confirmed DINO pre-training's superiority, particularly when labeled data were limited. CONCLUSIONS The application of DINO to unlabeled PAS-stained glomerular images facilitated the extraction of histological features that can be effectively utilized for disease classification.

中文翻译：

---

用于从肾小球图像中提取特征和疾病分类的自我监督学习，只需最少的注释。


背景深度学习在数字肾脏病理学中具有巨大潜力。但是，它的有效性在很大程度上取决于广泛标记数据集的可用性，而这些数据集通常由于创建数据集所需的专业知识和时间而受到限制。这种限制阻碍了深度学习在肾活检图像分析中的广泛应用。方法 我们将无标签自蒸馏 （DINO） 一种自我监督学习方法应用于从 384 张 PAS 染色肾活检玻片中获得的 10,423 张肾小球图像的数据集。使用主成分分析 （PCA） 可视化从 DINO 预训练主干中提取的肾小球特征。然后，我们通过将 k 最近邻 （kNN） 分类器或线性头层添加到 DINO 预训练或 ImageNet 预训练的主干来执行分类任务。这些模型在我们的标记分类数据集上进行了训练。使用受试者工作特征曲线下面积 （ROC-AUC） 等指标评估性能。分类任务包括四类疾病 （微小病变肾病、系膜增生性肾小球肾炎、膜性肾病和糖尿病肾病） 以及高血压、蛋白尿和血尿等临床参数。结果 PCA 可视化揭示了对应于不同肾小球结构的不同主成分，证明了 DINO 预训练骨架捕获形态学特征的能力。在疾病分类方面，DINO 预训练的转移模型 （ROC-AUC = 0.93） 优于 ImageNet 预训练的微调模型 （ROC-AUC = 0.89）。当标记数据受到限制时，ImageNet 预训练微调模型的 ROC-AUC 下降到 0。76 （95% 置信区间 [CI]，0.72-0.80），而 DINO 预训练的转移模型保持了卓越的性能 （ROC-AUC 0.88,95% CI 0.86-0.90）。DINO 预训练的转移模型在几个临床参数的分类方面也表现出更高的 AUC。使用两个独立数据集的外部验证证实了 DINO 预训练的优势，尤其是在标记数据有限的情况下。结论 将 DINO 应用于未标记的 PAS 染色肾小球图像有助于提取可有效用于疾病分类的组织学特征。