Glomerular Classification Using Convolutional Neural Networks Based on Defined Annotation Criteria and Concordance...

Area of Science:

• Nephrology
• Computer Science
• Medical Imaging

Background:

• Accurate diagnosis of renal pathologies is crucial for effective treatment.
• Classifying glomeruli is challenging for clinicians, necessitating computational support.
• This study introduces an automated system for glomerular image classification.

Purpose of the Study:

• To develop and evaluate a convolutional neural network (CNN) for automatic classification of glomerular images.
• To assess the performance of CNNs in identifying specific renal pathological features.
• To analyze classification errors and understand CNN decision-making processes.

Main Methods:

• Defined annotation criteria for 12 glomerular features.
• Evaluated inter-clinician agreement using kappa (κ) coefficient.
• Trained CNNs on 10,102 annotated images, targeting an average κ ≥ 0.4.
• Assessed CNN performance using receiver operating characteristic-area under the curve (ROC-AUC) and conducted error analysis with Grad-CAM.

Main Results:

• Inter-clinician agreement (κ) ranged from 0.28 to 0.50.
• CNNs achieved ROC-AUC values from 0.65 to 0.98.
• "Capillary collapse" (0.98) and "fibrous crescent" (0.91) showed high ROC-AUC.
• CNNs struggled with features of similar visual structures or simultaneous occurrence.

Conclusions:

• CNN performance is influenced by feature texture and co-occurrence frequency.
• To enhance classification accuracy, advanced methods like image segmentation are necessary.
• Further development is needed to improve automated renal pathology diagnosis.