A deep-learning approach at automated detection of electron-dense immune deposits in medical renal biopsies

Abstract

Introduction: Identification of electron-dense immune deposits in electron microscopy (EM) images is integral to the diagnosis of medical renal disease. Deep learning has the potential to augment this process, especially in areas with limited resources. Objectives: Our study explores the feasibility of applying deep learning to detect electron dense immune deposits in electron microscopy images from medical renal biopsies. Patients and Methods: EM images (N=900) from native and transplant kidney biopsies were processed into 4530 tiles (512 x 512 pixels). These tiles were reviewed and classified into one of three categories: deposits absent, deposits present, and indeterminate. This classification resulted in 1255 images with consensus agreement for deposits present and deposits absent. These 1255 images were then used to train a machine learning model, using 1006 images for training, and 249 images for testing. Results: The overall accuracy on the test data was a competitive 78%, and the F1 scores for deposits absent and present was 0.76 and 0.79, respectively. Conclusion: This study demonstrated the feasibility of creating and applying a machine learning model that performs competitively in identifying electron dense deposits in EM images.