Development of a deep learning-based histological evaluation model for critical-size bone defect healing in rats - an objective tool.

INTRODUCTION: Critical-size femoral defects in rats are a well-established model for preclinical bone regeneration research. Histological evaluation is essential for assessing healing but remains time-consuming and subject to observer variability. Machine learning, particularly convolutional neural networks (CNNs), offers potential for objective and scalable analysis of histological sections. MATERIALS AND METHODS: We developed a modified U-Net model to perform semantic segmentation and classification of bone healing stages based on Movat pentachrome-stained histological sections (n = 669). Five tissue classes (bone, cartilage, bone marrow, granulation tissue, background) were manually annotated to train the model. Data were split into training (64%), validation (16%), and test (20%) sets. The model then was used to segment and rank histological images. In addition, a subset of 20 independent test images was scored by four orthopedic experts, seven medical students, and the AI using a refined bone healing score ranging from -10 to +10. RESULTS: The model achieved high segmentation performance, particularly for bone and background. AI-generated healing scores showed strong correlation with expert ratings (Spearman r = 0.819, p < 0.0001) and similar accuracy to student ratings (mean absolute deviation: AI = 0.468 vs. students = 0.469; p = 0.5753). ICC analysis confirmed excellent agreement between AI and experts (ICC = 0.820) and revealed a significant difference favoring AI over students (bootstrap p = 0.0466). CONCLUSION: This study introduces a CNN-based model capable of expert-level performance in the histological assessment of bone healing. It offers a reproducible and time-efficient tool for future preclinical applications.