Classification of glomerular hypercellularity using convolutional features and support vector machine

Area of Science:

• Nephrology
• Digital Pathology
• Biomedical Image Analysis

Background:

• Glomeruli are vital kidney structures for blood filtration.
• Glomerular lesions, like hypercellularity, impair kidney function and are common in kidney diseases.
• Automatic detection of glomerular hypercellularity can accelerate histological slide screening and improve clinical diagnosis.

Purpose of the Study:

• To develop and evaluate a novel deep learning approach for the classification of glomerular hypercellularity in human kidney images.
• To assess the performance of the proposed method in both binary (lesion vs. normal) and multi-classification (sub-lesion types) tasks.
• To compare the proposed method against established deep learning models and traditional approaches.

Main Methods:

• A novel convolutional neural network (CNN) architecture combined with a support vector machine was developed.
• The method was applied to a dataset of human kidney images (FIOCRUZ dataset).
• Performance was evaluated on binary classification and multi-classification of hypercellularity sub-lesions (mesangial, endocapilar, both).

Main Results:

• The proposed method achieved near-perfect average results in binary classification of glomerular hypercellularity.
• An average accuracy of 82% was reached for the multi-classification of hypercellularity sub-lesions.
• The novel approach outperformed Xception, ResNet50, InceptionV3, and a handcrafted feature-based method.

Conclusions:

• The developed deep learning method demonstrates high efficacy in classifying glomerular hypercellularity in human kidney images.
• This approach offers a significant advancement for accelerating the diagnosis of kidney diseases characterized by glomerular hypercellularity.
• This study represents the first application of deep learning to a dataset of glomerular hypercellularity in human kidney images.