Utilizing Deep Learning to Identify Electron-Dense Deposits in Renal Biopsy Electron Microscopy Images.

Introduction: Electron microscopy (EM) is a crucial technique for identifying and distinguishing the specific location of deposits within glomeruli. Manual classification of these deposit locations is not only time-consuming but also yields inconsistent results among pathologists. This study aimed to develop a deep learning-based platform to automatically classify the locations of electron-dense deposits in EM images of kidney biopsies.

Methods: We retrospectively collected 4,303 EM images at magnifications of ×4,000 to ×8,000 from 1,039 kidney biopsies performed on native kidneys at the Renal Pathology of King Medical Diagnostics Center in Guangzhou between June 1, 2022, and July 2, 2022. EM images were independently evaluated by pathologists Zhu and Luo for electron-dense deposits, categorized into mesangial, subepithelial, intramembranous, and subendothelial if present. These evaluations served as ground truth for model training and evaluation. Of these, 3,443 EM images were allocated to the training group and 860 to the test group. The ResNet18 architecture was selected for our task. To evaluate the model's classification capability, we created a binary classification model to identify the presence of deposits in EM images. Furthermore, we implemented a subnet classification network to predict the probability of mesangial, subepithelial, intramembranous, and subendothelial deposits. Four renal pathologists (two EM pathologists and two comprehensive renal pathologists) were invited to compare their agreement with the deep learning model. Comparing deep learning models against pathologists with Cohen's Kappa and accuracy.

Results: The deep learning model can accurately identify the presence of electron-dense deposits in EM images, with an area under the receiver operating characteristic curve (AUC) of 0.959 and an accuracy of 0.899. The classification subnet trained to identify mesangial, subepithelial, intramembranous, and subendothelial deposits yielded AUCs of 0.928, 0.987, 0.986, and 0.944, with accuracies of 0.880, 0.962, 0.959, and 0.883, respectively. Subepithelial and intramembranous deposits had near-perfect agreement, while mesangial and subendothelial deposits showed substantial agreement with the ground truth. The accuracy of deep learning models in assessing the presence and locations of deposits was lower than that of EM pathologists but higher than that of comprehensive renal pathologists. A web platform has been developed in which users can upload EM images of renal biopsies to receive probabilities regarding the four locations of deposits based on our algorithm.

Conclusion: We successfully developed a web platform for the automated assessment of the locations of electron-dense deposits in kidney biopsy EM images. The performance of this model surpasses that of experienced comprehensive renal pathologists, offering an efficient and reliable tool.