A recent publication in The Lancet highlights a groundbreaking study that aims to standardize the diagnosis of glomerular nephritis (GN) through an innovative artificial intelligence-assisted model. This approach facilitates automated analysis of kidney biopsies, enhancing diagnostic accuracy and consistency.
The multicenter study, which involved data from over 6,682 patients across three medical facilities and diagnostic centers in China, marks a significant advancement in the field. This is the first research of its kind to leverage an AI-assisted model specifically for GN diagnosis, demonstrating improved standardization and efficiency while minimizing interobserver variability.
According to Dr. Fan Fan Hou, chief of the renal division at Nanfang Hospital in Guangzhou, the traditional diagnosis of GN relies heavily on the meticulous evaluation of histopathological images from kidney biopsies. This manual assessment is performed by skilled pathologists, who interpret various morphological, structural, and compositional changes in the glomeruli. However, this method can be subjective, labor-intensive, and time-consuming, often leading to inconsistencies in diagnosis.
The AI-assisted model analyzed 106,988 light microscopy images of glomeruli to classify different glomerular lesions. It also utilized immunofluorescence techniques to interpret immune marker patterns, extracting essential diagnostic features. The model”s performance was assessed through metrics such as F1-score, precision, recall, and overall accuracy.
Researchers conducted computational analyses using 6,682 retrospective and deidentified patient records from three medical centers. The internal validation cohorts comprised 1,235 and 312 patient groups from the Nanfang Hospital. Additionally, external validation was performed on two cohorts, which included 2,484 patients from the Jinyu Diagnostic Center and 2,652 from the Huayin Diagnostic Center.
In the internal validation cohort, various GN types were represented, with 72 patients diagnosed with focal segmental glomerulosclerosis (FSGS), 81 with IgA nephropathy (IgAN), 80 with minimal change disease (MCD), and 79 with membranous nephropathy (MN). External validation cohort I included 110 patients with FSGS, 914 with IgAN, 392 with MCD, and 1,067 with MN. Cohort II featured 183 patients with FSGS, 891 with IgAN, 222 with MCD, and 1,356 with MN.
The internal validation cohort reported F1-scores of 84.48%, with precision at 85.48% and recall at 85.52%. In external validation cohort I, the F1-score was 83.86%, precision was 83.86%, and recall reached 87.84%. For external validation cohort II, the F1-score stood at 85.45%, with precision at 83.12% and recall at 88.94%. Notably, the model excelled in classifying the four GN types, especially in external validation cohort I, where MN achieved a remarkable F1-score of 97.13%.
Despite the promising results, the researchers acknowledged limitations within their study, particularly regarding the need for further validation in diverse populations, as the external datasets may not fully represent different racial or ethnic backgrounds. They concluded that this study is a pioneering effort in developing an AI-assisted model for diagnosing GN using kidney biopsy images, backed by a sufficiently large sample size for both development and external validation.
