Glomerular disease classification and lesion identification by machine learning.

Area of Science:

• Nephropathology
• Medical Image Analysis
• Artificial Intelligence

Background:

• Kidney disease classification and lesion identification rely on manual, time-consuming, and variable expert analysis.
• Machine learning (ML)-based image analysis offers a promising alternative for objective and efficient interpretation.

Purpose of the Study:

• To develop and evaluate ML models for automated glomerulus detection, disease classification, and lesion identification in kidney biopsies.
• To simulate and potentially improve the accuracy and efficiency of nephropathologist interpretation.

Main Methods:

• A Mask Region-based Convolutional Neural Network (Mask R-CNN) was used for glomerulus detection.
• A Long Short-Term Memory (LSTM) recurrent neural network classified diseases, and a ResNeXt-101 model identified lesions in lupus nephritis.
• Models were trained and validated on human kidney biopsy samples.

Main Results:

• The glomerulus detection model achieved an F1 score of 0.944.
• The disease classification model reached an accuracy of 0.940 for H&E whole slide images.
• The lesion identification model showed an area under the ROC curve of 0.947, with good generalization to external datasets.

Conclusions:

• This study demonstrates the first ML models capable of simulating key steps in kidney biopsy interpretation.
• Integrated into a viewing platform, these models assist nephropathologists in reviewing and confirming AI-driven results.
• Future integration with other data types may lead to clinical application.