ABSTRACT
Background The number of glomeruli on a kidney biopsy slide followed by glomerular assessment constitute as standard components of a renal pathology report. The prevailing method for glomerular identification and assessment remains manual, labor intensive and non-standardized. In the era of digitized kidney biopsies, an automated method to identify, segment and count glomeruli is highly desirable.
Methods and results We developed an automated method to detect and segment the glomeruli within digitized kidney biopsy images by leveraging a deep learning architecture based on convolutional neural networks (CNN). A total of 275 trichrome-stained images (Average image size: 2560×1920×3 pixels, 1-2 unique images per patient, Scale: 0.85 μm/pixel) processed at 40× magnification from renal biopsies of 171 chronic kidney disease patients treated at the Boston Medical Center from 2009-2012 were analyzed. A sliding window operation was defined to crop each 40× image to smaller images of size 300×300×3 pixels. Each cropped image was then evaluated by clinical experts to identify the presence of a unique glomerulus, and each identified glomerulus was included in the training dataset (n = 751). About the same number of cropped images, containing the non-glomerular regions of the kidney biopsy, served as control cases. The CNN model was constructed as a binary classification problem to discriminate glomerular images from the non-glomerular ones (Performance on test data - Accuracy: 97.47±0.31%; Sensitivity: 96.43±1.89%; Specificity: 98.76±1.44%). Using the trained CNN model, another sliding window operator was developed to scan the digitized biopsies. A heatmap was generated to highlight regions of intensity that the CNN model classified as glomerular regions. Subsequently, two independent image processing strategies, one using steps such as image binarization and image erosion, and the other using image binarization, distance transform and watershed segmentation, were performed on the heatmaps to generate discriminatory signatures of the identified glomeruli. The final step involved automatically drawing a box around the higher intensity areas leading to output images with segmented glomerular regions (Performance on test data - Accuracy: 99.97±0.0086%; Sensitivity: 54.37±0.23%; Specificity: 99.99±0.009%).
Conclusion While used in the context of nephropathology, this study demonstrates the power of artificial intelligence to assess complex histologic structures and identify structural variations. Adoption of such methods of counting and classifying glomeruli using standard histological staining without disturbing the workflow can expedite the assessment of slides by the pathologists and serve as a first step toward more comprehensive automated analysis.