100 Section - Quantitative Assessment of Bone Cell Populations in Human Trabecular Bone Histology Using Artificial Intelligence

Accurate characterization of bone cell populations is essential for studying bone remodeling, aging, and musculoskeletal recovery relevant to rehabilitation medicine. While deep learning–based nucleus detectors are increasingly used in histopathology, their baseline performance for bone-specific cell analysis remains unclear. This study aimed to evaluate the performance of a pretrained artificial intelligence (AI) model, HoVerNet, for detecting osteocyte and osteoblast nuclei in human trabecular bone histology, establishing a baseline for future adaptation.

Design:

Histological sections of human trabecular bone from femoral condyles obtained during hip arthroplasty were analyzed (n=15 donors; 8 female, 7 male; >300 sections). Osteocytes and osteoblasts were manually annotated by experts with pathologist quality control; osteoclasts were excluded. Bone tissue was identified within the annotated slides using a segmentation algorithm [1]. Nucleus detection by the HoVerNet algorithm [2] was evaluated within biologically motivated distance bands (≤20 µm and ≤40 µm from the bone surface). Detection performance was assessed by matching predicted nuclei to expert annotations within a fixed distance threshold. Quality control excluded sections with insufficient annotation–bone proximity.

References:  [1] DOI: 10.1016/j.pathol.2021.07.011; [2] DOI: 10.1016/j.media.2019.101563.

Results:

Ground-truth analysis showed that nearly all annotated bone cell nuclei were located within 40 µm of the bone region. 2 of 321 sections (0.6%) failed quality control, indicating robust annotation–segmentation alignment. For the 20 µm bone-surface band, pooled (micro-averaged) HoVerNet detection performance across quality-controlled sections yielded a precision of 0.112, recall of 0.614, and F1 score of 0.189. Slide-level median F1 was 0.241, reflecting variability in precision despite moderate recall.

Conclusion:

A pretrained AI-based nucleus detection algorithm achieves moderate sensitivity but low precision for bone cell detection in human trabecular bone histology, revealing substantial limitations of generic models for bone-specific applications. These findings underscore the need for musculoskeletal-specific cell classification approaches to support quantitative, clinically meaningful analysis relevant to regenerative medicine and precision rehabilitation.