Machine learning in renal pathology.

Area of Science:

• Computational pathology
• Digital pathology
• Renal pathology

Background:

• Traditional kidney biopsy assessment methods (light microscopy, immunofluorescence, electron microscopy) are time-consuming, expensive, and prone to variability.
• Computational approaches offer potential solutions for screening and diagnosis, aiming to reduce pathologist workload and identify novel biomarkers.

Purpose of the Study:

• To evaluate the diagnostic performance of a biomarker feature extraction (BFE) model and three deep learning models (VGG16, VGG19, InceptionV3) for diagnosing glomerular diseases using only PAS-stained digital pathology images.

Main Methods:

• The BFE model extracted 233 features, reduced to 10 morphological and texture features for classification using linear discriminant analysis.
• Three pre-trained deep learning models (VGG16, VGG19, InceptionV3) were applied to digital pathology images.
• Data augmentation and Grad-CAM were utilized for deep learning models to enhance performance and interpretability.
• The study included 45 renal biopsies (371 glomeruli) from minimal change disease, membranous nephropathy, and thin-basement membrane nephropathy.

Main Results:

• The BFE model achieved a glomerular testing accuracy of 76.8%.
• Deep learning models showed higher validation accuracies, with VGG16 reaching 78.5%.
• The highest testing accuracies were VGG16 at the glomerular level (71.9%) and InceptionV3 at the patient level (73.3%).

Conclusions:

• Both traditional machine learning (BFE) and deep learning approaches demonstrate potential for aiding in kidney biopsy evaluation.
• These computational methods may offer efficient and accurate diagnostic tools for glomerular diseases.