Deep learning-based multi-model approach on electron microscopy image of renal biopsy classification

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Nephrology

Background:

• Electron microscopy is crucial for diagnosing renal diseases, particularly identifying electron-dense granules in immune-mediated conditions.
• Deep learning (DL) shows promise for histologic image analysis, but its application to renal biopsy electron microscopy images is underexplored.

Purpose of the Study:

• To develop and evaluate a DL-based multi-model for automated detection of electron-dense granules in renal biopsy transmission electron microscopy (TEM) images.
• To assist in the diagnosis of immune-mediated renal diseases by analyzing complex electron microscopy findings.

Main Methods:

• Trained three DL models on 910 renal biopsy electron microscopy images to classify the presence of electron-dense granules.
• Developed an improved ResNet model with skip architecture and Support Vector Machine (SVM) classifier, utilizing transfer learning.
• Created a multi-model combining traditional and improved ResNet models to enhance classification accuracy.

Main Results:

• The DL-based multi-model achieved the highest accuracy, averaging approximately 88%.
• The improved ResNet+SVM model demonstrated significantly better performance (83% average accuracy) compared to the traditional ResNet model (58% average accuracy).

Conclusions:

• This research presents the first DL models for electron microscopy image classification of renal biopsies.
• Accurate identification of electron-dense granules is vital for diagnosing immune complex nephropathy.
• AI models can effectively assist in analyzing complex electron microscopy images for improved disease diagnosis.