DeepQCT: Predicting fragility fracture from high-resolution peripheral quantitative CT using deep learning

Background Osteoporosis is prevalent in elderly women, which causes fragility fracture and hence increased mortality and morbidity. Predicting osteoporotic fracture risk is both clinically-beneficial and cost-effective. However, traditional tools using clinical factors and bone mineral density (BMD) fail to reflect bone microstructure. Here we aim to use high-resolution peripheral quantitative CT (HR-pQCT) images to construct deep-learning models which predict fragility fracture history in elderly Chinese women. Methods We used ChiVOS, a community-based national cohort of 2,664 Chinese elderly women. Demographic data, BMD, and HR-pQCT from 216 patients were used to construct three groups of models: BMD, pQCT-index, and DeepQCT. For DeepQCT, we used ResNet34 as classifier, and logistic regression for late fusion. Models were developed using 6-fold cross-validation in development set (90%, N=195), and tested in internal test set (10%, N=21). We applied unsupervised clustering on HR-pQCT indices to derive patient subgroups. Findings DeepQCT (best model AUC 0.86-0.94) was superior or similar to pQCT-index (best model AUC 0.8-0.93), which both outperformed BMD (best model AUC 0.54-0.78). Surprisingly, DeepQCT built from non-weight-bearing bones performed similarly to weight-bearing bones. Furthermore, two distinct patient groups were classified using HR-pQCT indices. The one with higher DeepQCT risk score showed lower volumetric BMD, bone more microarchitectural abnormalities, and had higher probability of osteoporosis and fragility fracture history. Interpretation DeepQCT scores and HR-pQCT-index permit early recognition of patients with high risk of fragility fracture. This established framework can be easily adapted for other diagnostic tasks using HR-pQCT scans, which promotes bone health management via digital medicine. Funding This research was supported by the National Natural Science Foundation of China (LC, 82100946; WX, 82270938), CAMS Innovation Fund for Medical Sciences (WX, 2021-I2M-1-002), National Key R&D Program of China (WX, 2021YFC2501700), National High Level Hospital Clinical Research Funding (WX, 2022-PUMCH-D-006), the Non-profit Central Research Institute Fund of Chinese Academy of Medical Sciences (LC, 2023-PT320-10), and Young Elite Scientists Sponsorship Program by BAST (LC, No.BYESS2023171). Part of the study was supported by Merck Sharp & Dohme China, Hangzhou, China. Evidence before this study Bone mineral density (BMD) from dual X-ray absorptiometry was firstly used to predict fragility fracture, but had low sensitivity. Tools like FRAX, QFracture, and Garvan, which also incorporated clinical factors into prediction models, showed improved performance. Models containing standard HR-pQCT indices (μFRAC) further surpassed most clinical tools. Nevertheless, direct learning from original HR-pQCT images is always desired to reduce labor and bias. Deep learning being the most common method for image-based learning, we searched PubMed for articles published up to Mar 25, 2024, using keywords “(‘fragility fracture’ OR ‘osteoporotic fracture’) and (‘prediction model’) and (‘HR-pQCT’ or ‘High-resolution peripheral quantitative CT’) and (‘deep-learning’ OR ‘deep learning’)”. Results showed that no study has built deep learning models from HR-pQCT for fragility fracture prediction. Added value of this study We developed DeepQCT from HR-pQCT of 216 elderly Chinese women from a national cohort (ChiVOS), which calculated risk scores using individual bone images and clinical features. BMD and pQCT-index models were compared to DeepQCT. We found both DeepQCT (best model AUC 0.86-0.94) and pQCT-index (best model AUC 0.8-0.93) outperformed BMD (best model AUC 0.54-0.78). DeepQCT using non-weight-bearing bones (ulna, fibula) performed similarly to weight-bearing bones (tibia, radius). Specifically, HR-pQCT revealed one patient subgroup with higher DeepQCT risk scores, which showed lower BMD and multiple bone microarchitectural abnormalities, associated with osteoporosis and fragility fracture history. Implications of all the available evidence DeepQCT is the first method which uses deep-learning to predict fragility fracture directly from HR-pQCT images. It is also the first to use single bones individually in prediction models, including non-weight-bearing bones, which are excluded in HR-pQCT-index computation. Of note, DeepQCT risk score is highly clinically relevant, as showed in bone density or microarchitectural features differences between patient subgroups. The non-inferior performance of DeepQCT compared to the manual annotation-dependent pQCT-index, supported its application to reduce labor and enhance efficiency. Performance of non-weight-bearing bones also challenges traditional perception of using load-bearing bones only in predicting osteoporotic conditions. Most importantly, the DeepQCT framework can be easily adapted for other tasks using HR-pQCT scans, which greatly expands application of digital medicine in bone mineral disease diagnosis or management. ### Competing Interest Statement The authors have declared no competing interest. ### Funding Statement This research was supported by the National Natural Science Foundation of China (LC, 82100946; WX, 82270938), CAMS Innovation Fund for Medical Sciences (WX, 2021-I2M-1-002), National Key R&D Program of China (WX, 2021YFC2501700), National High Level Hospital Clinical Research Funding (WX, 2022-PUMCH-D-006), the Non-profit Central Research Institute Fund of Chinese Academy of Medical Sciences (LC, 2023-PT320-10), and Young Elite Scientists Sponsorship Program by BAST (LC, No.BYESS2023171). Part of the study was supported by Merck Sharp & Dohme China, Hangzhou, China. ### Author Declarations I confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained. Yes The details of the IRB/oversight body that provided approval or exemption for the research described are given below: Ethics committee/IRB of Peking Union Medical College Hospital gave ethical approval for this work I confirm that all necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived, and that any patient/participant/sample identifiers included were not known to anyone (e.g., hospital staff, patients or participants themselves) outside the research group so cannot be used to identify individuals. Yes I understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance). Yes I have followed all appropriate research reporting guidelines, such as any relevant EQUATOR Network research reporting checklist(s) and other pertinent material, if applicable. Yes Desensitized demographic data, HR-pQCT indices, HR-pQCT images, and codes of the current study would be available upon reasonable request to corresponding author.