Deep Learning Augmented Osteoarthritis Grading Standardization.

Area of Science:

• Biomedical imaging
• Artificial intelligence in medicine
• Histopathology

Background:

• Manual grading of osteoarthritis (OA) from cartilage histology is subjective and prone to errors.
• Interobserver variability in manual grading leads to ambiguities in OA severity assessment.
• Deep learning (DL) offers a potential solution for objective and automated image classification.

Purpose of the Study:

• To assess the feasibility of training a deep neural network (DNN) for automated grading of knee OA severity using histology images.
• To develop and validate a DL model based on a modified Mankin scoring system for OA grading.
• To compare the performance of different DL architectures for cartilage histology image classification.

Main Methods:

• A simplified grading system based on Safranin-O staining, chondrocyte arrangement, and morphology was developed.
• Histology images were tiled, labeled, and grouped into four OA grades (0-3).
• Four DL architectures were evaluated using a fivefold cross-validation method, with DenseNet121 selected as the best model.

Main Results:

• DenseNet121 achieved a validation accuracy of approximately 84% and a Cohen's kappa score of 0.632.
• The model demonstrated excellent discriminatory ability with ROC-AUC values ranging from 0.89 to 0.99.
• Automated grades from the DL model aligned well with expert medical assessments.

Conclusions:

• Deep learning models can accurately interpret cartilage degradation from histology images for OA grading.
• This automated approach using histology images offers a fundamental and standardized method for OA assessment.
• The study demonstrates a paradigm shift towards AI-driven standardization in histology-based OA grading.