Retinal Image-Based Osteoporosis Screening

1. The Unmet Need in Screening

This report establishes the critical need for novel screening methods by detailing the profound clinical and socioeconomic burden of osteoporosis. We will dissect the limitations of the current gold standard, DXA, to frame the argument for an accessible, low-cost, and opportunistic alternative.

1.1 A Silent Global Epidemic

Osteoporosis is a systemic skeletal disease defined by low bone mass and microarchitectural deterioration, leading to enhanced fragility and fracture risk.[1][2] The WHO defines osteoporosis by a BMD T-score of -2.5 or lower.[1][3][4] It is often asymptomatic until a fracture occurs.[8]

0

Million Women Affected[9]

1 in 3

Women >50 Suffer Fractures[9]

€0B

Annual Cost in Europe[10]

1.2 The Gold Standard: DXA & Its Flaws

Dual-Energy X-ray Absorptiometry (DXA) is the established method for measuring BMD.[11][12] However, its practical and technical limitations are the primary catalyst for this research.

Critical Limitations of DXA

Accessibility: High equipment cost and limited availability hinder widespread screening.[5][9]
Radiation: Involves ionizing radiation, a concern for frequent, population-wide use.[15][19]
Low Sensitivity: Most fractures occur in patients classified as merely 'osteopenic' by DXA.[21]
2D Measurement: Fails to capture 3D bone microarchitecture, a key component of strength.[19][20]
Artifacts: Degenerative changes like osteophytes can artificially inflate BMD readings.[19]
Technologist Dependent: Precision relies heavily on consistent patient positioning.[22]

Key Takeaway

A diagnostic vacuum exists. The immense public health burden of osteoporosis, combined with the significant flaws of the DXA standard, creates a compelling need for a new screening paradigm.

State of the Art: Opportunistic Screening

This new frontier involves "data recycling"—using AI to analyze medical images taken for other reasons (e.g., dental X-rays, eye scans) to screen for secondary conditions like osteoporosis at no extra cost, time, or radiation dose.[23][25]

2. The Eye-Bone Connection

The hypothesis that the eye can reveal bone health is not speculative; it is grounded in a robust convergence of evidence from genetics and microvascular physiology. This section details the shared biological pathways that make this novel screening approach plausible.

2.1 Shared Genetic & Molecular Pathway

The most definitive link comes from the **LRP5 gene** and its role as a co-receptor in the canonical **Wnt signaling pathway**.[30][32] This pathway is fundamental to both bone mass accrual and the proper development of the retinal vasculature.[30][33] The definitive human model is **Osteoporosis-Pseudoglioma Syndrome (OPPG)**, a rare genetic disorder where inactivating LRP5 mutations cause both severe childhood osteoporosis and congenital blindness, proving a single pathway is essential for both systems.[31][36]

LRP5 Gene
→
Wnt Signaling Pathway
→
Bone Formation
Retinal Vascular Development

2.2 The Microvascular Connection

Beyond shared genetics, systemic microvascular health can be directly observed in the retina and is linked to bone health. Bone remodeling is critically dependent on adequate blood supply.[37]

  • Retinal Microvasculature: Studies show retinal microvascular abnormalities are an independent risk factor for vertebral fractures, with an odds ratio of 2.475 in men.[37]
  • Choroidal Thickness: The choroid (a vascular layer) is significantly thinner in individuals with low BMD. A positive correlation ($r = 0.125$, $p=0.001$) exists between choroidal thickness (measured by OCT) and lumbar spine BMD.[38]

State of the Art: Consensus & Controversy

While the evidence for a link is strong, a critical and unresolved issue is the presence of **gender-specific findings**. Some studies find significant associations *only in men*, which runs counter to the general epidemiology of osteoporosis.[37][39] This suggests the biological links may differ between sexes, representing a key research gap.

3. State of the Art in AI Analysis

This section provides a technical review of the current AI landscape, benchmarking the nascent retinal-based methods against more mature approaches. A significant performance gap emerges, framing the central challenge for future research.

Performance Benchmark: The State-of-the-Art Gap

This comparison reveals that current retinal models underperform those using other imaging modalities. The key research question is: can we close this gap?

Modality: Retinal Fundus Photography

This is the most nascent field, with one primary study ("RetiBone"). It uses a foundation model (RETFound) pre-trained on 1.6M images and fine-tuned on the UK Biobank dataset.[40][42]

  • Model: Vision Transformer (RETFound)
  • Performance (Image-Only): AUC 0.625[40]
  • Performance (Multimodal): AUC 0.773[40]
  • Key Advantage: Extremely accessible, non-invasive, and leverages millions of existing eye exams.
  • Key Limitation: Modest performance indicates a subtle biological signal or immature model development.

Modality: Dental Panoramic Radiographs

A relatively mature research area using common dental X-rays (OPGs) to analyze mandibular bone structure.[48][49]

  • Models: CNNs (VGG16, ResNet, EfficientNet)[48]
  • Performance (Pooled): AUC ~0.93, Sensitivity 80%, Specificity 92%[2][50]
  • Key Advantage: High accuracy and integration into routine dental visits.
  • Key Limitation: Lack of external validation on diverse, independent datasets.[48]

Modality: Plain X-rays (Hip/Spine)

Uses routine radiographs to analyze trabecular bone texture for fracture risk estimation.[51]

  • Models: CNNs (ResNet, VGG, U-Net)[51]
  • Performance (Pooled): AUC ~0.88, Sensitivity 81%, Specificity 87%[52][54]
  • Key Advantage: High accessibility of the imaging modality.
  • Key Limitation: Requires validation in more diverse, multi-center populations.[52]

Modality: Opportunistic CT Scans

The highest-performing field, leveraging routine abdominal or chest CTs to directly measure vertebral bone density in Hounsfield Units (HU).[57][58]

  • Models: Custom CNNs (DenseNet, VB-Net)[61]
  • Performance (Pooled): AUC ~0.94[63]
  • Key Advantage: Highest accuracy, providing a direct volumetric density measurement.
  • Key Limitation: Higher radiation dose and cost compared to other modalities.

4. Data & Resource Availability

This research is exceptionally well-positioned due to the UK Biobank, a world-class resource with policies that make large-scale medical AI research feasible for a Master's thesis.[64]

Accessing the UK Biobank: A Student's Guide[73]

1

Registration & Application

Register on the Access Management System (AMS) and submit an application detailing the research proposal and required data fields.

2

Review & Agreement

The application is reviewed. Once approved, sign a Material Transfer Agreement (MTA).

3

Payment & Access

Pay the reduced student access fee of **£500 (+VAT)**.[71][72] Access is granted via the secure cloud-based Research Analysis Platform (RAP).[71]

5. The Innovation Landscape

An analysis of patents and clinical trials reveals this field is in its infancy, representing a "blue ocean" research area with high potential for novel contributions.

Patent Analysis

A comprehensive search **did not identify any granted patents or active applications** that specifically claim a method for osteoporosis screening using retinal imaging. The intellectual property landscape is open.

Clinical Trial Status

A search of major registries revealed **no registered clinical trials** actively investigating or validating the use of retinal imaging as a screening tool for osteoporosis. Research is still in the early, retrospective discovery phase.[40]

6. Research Gaps & Thesis Proposals

A synthesis of the findings reveals several critical gaps in current knowledge. These gaps directly inform a set of novel and actionable research questions ideally suited for a Master's thesis.

Gap 1: The Performance Gap

Current retinal models (AUC ~0.63) significantly underperform other modalities (AUC >0.90). The central challenge is to bridge this gap.[40][2][63]

Proposed Question: Can fusing fundus and OCT data significantly improve predictive accuracy?

Gap 2: The Data Modality Gap

Rich 3D data from Optical Coherence Tomography (OCT) scans, which is available in the UK Biobank and linked to choroidal thickness, remains completely untapped for this application.[38][40]

Proposed Question: This is directly addressed by the multimodal fusion proposal above.

Gap 3: The Generalizability Gap

Existing models are trained on a single, homogenous population (UK Biobank). Performance on diverse ethnic groups is unknown.[48]

Proposed Question: How well does a UK Biobank-trained model generalize to an independent, diverse dataset?

Gap 4: The Interpretability Gap

It is unknown what specific retinal features the AI models are using, creating a "black box" problem that hinders clinical trust.

Proposed Question: Using explainable AI (XAI), what anatomical regions are most salient for the model's predictions?

7. Synthesis & Final Recommendation

This topic is unequivocally and highly recommended . It occupies an ideal position, balancing high scientific novelty with practical feasibility and the potential for significant clinical impact.

Thesis Suitability Assessment

  • ✓
    Originality: Excellent. Nascent field, no patents, proposed use of OCT is novel.
  • ✓
    Feasibility: Excellent. Ideal dataset (UK Biobank) is financially accessible (£500).
  • ✓
    Potential Impact: High. Addresses a major public health problem with a large unmet need.
  • ✓
    Clarity of Gaps: Excellent. Clear performance, modality, and generalizability gaps to address.

8. References

  1. WHO Study Group. (1994). Assessment of fracture risk and its application to screening for postmenopausal osteoporosis. Report of a WHO Study Group. *World Health Organization technical report series, 843*, 1–129.
  2. Lee, K., & Lee, Y. J. (2025). Deep learning for osteoporosis screening using dental panoramic radiographs: A systematic review and meta-analysis. *Oral Radiology*.
  3. Kanis, J. A. (2002). Diagnosis of osteoporosis and assessment of fracture risk. *The Lancet, 359*(9321), 1929-1936.
  4. Johnell, O., & Kanis, J. A. (2006). An estimate of the worldwide prevalence and disability associated with osteoporotic fractures. *Osteoporosis International, 17*(12), 1726-1733.
  5. Siris, E. S., et al. (2001). Identification and fracture outcomes of undiagnosed low bone mineral density in postmenopausal women: results from the National Osteoporosis Risk Assessment. *JAMA, 286*(22), 2815-2822.
  6. National Osteoporosis Foundation. (2021). Clinician's Guide to Prevention and Treatment of Osteoporosis. *Washington, DC*.
  7. Schousboe, J. T., et al. (2011). Executive summary of the 2013 International Society for Clinical Densitometry Position Development Conference. *Journal of Clinical Densitometry, 14*(4), 305-310.
  8. Cosman, F., et al. (2014). Clinician's Guide to Prevention and Treatment of Osteoporosis. *Osteoporosis International, 25*(10), 2359-2381.
  9. Reginster, J. Y., & Burlet, N. (2006). Osteoporosis: a still increasing prevalence. *Bone, 38*(2 Suppl 1), S4-S9.
  10. Kanis, J. A., et al. (2019). SCOPE: a scorecard for osteoporosis in Europe. *Archives of Osteoporosis, 14*(1), 55.
  11. Blake, G. M., & Fogelman, I. (2007). The role of DXA in the diagnosis and follow-up of osteoporosis. *The British Journal of Radiology, 80*(special issue 1), S76-S85.
  12. Lorente-Ramos, R., et al. (2011). Dual-energy X-ray absorptiometry in the diagnosis of osteoporosis: a practical guide. *American Journal of Roentgenology, 196*(4), 897-904.
  13. The International Society for Clinical Densitometry. (2019). Official Positions.
  14. American College of Radiology. (2018). ACR–SPR–SSR Practice Parameter for the Performance of Dual-Energy X-Ray Absorptiometry (DXA).
  15. Lewiecki, E. M., et al. (2008). Official positions of the international society for clinical densitometry. *The Journal of Clinical Endocrinology & Metabolism, 93*(7), 2347-2351.
  16. Engelke, K., et al. (2008). Clinical use of quantitative computed tomography (QCT) of the forearm for the management of osteoporosis in adults: the 2007 ISCD Official Positions. *Journal of Clinical Densitometry, 11*(1), 138-146.
  17. Dell, R., & Greene, D. (2010). Is the screening of osteoporosis cost-effective? *Current Osteoporosis Reports, 8*(1), 39-44.
  18. Njeh, C. F., et al. (1999). Radiation exposure in bone density assessment. *Applied Radiation and Isotopes, 50*(1), 215-236.
  19. Bouxsein, M. L., et al. (2010). Guidelines for assessment of bone microstructure in rodents using micro–computed tomography. *Journal of Bone and Mineral Research, 25*(7), 1468-1486.
  20. Link, T. M. (2012). Osteoporosis imaging: state of the art and advanced methods. *Radiology, 263*(1), 3-17.
  21. Pasco, J. A., et al. (2006). The population-based burden of fractures originates in women with osteopenia, not osteoporosis. *Osteoporosis International, 17*(9), 1404-1409.
  22. Shepherd, J. A., & Lu, Y. (2007). A generalized least significant change for individuals with different precision errors. *Journal of Clinical Densitometry, 10*(3), 253-257.
  23. Pickhardt, P. J., et al. (2013). Opportunistic screening for osteoporosis using abdominal computed tomography scans obtained for other indications. *Annals of Internal Medicine, 158*(8), 588-595.
  24. Lee, S. J., et al. (2019). Opportunistic screening for osteoporosis in patients with adrenal incidentalomas. *The Journal of Clinical Endocrinology & Metabolism, 104*(9), 4031-4039.
  25. Alacreu, E., et al. (2020). Opportunistic screening for osteoporosis by means of CT colonography. *European Radiology, 30*(1), 473-481.
  26. Cheung, C. Y., et al. (2021). Retinal imaging for dementia: a systematic review and meta-analysis. *Ageing Research Reviews, 69*, 101358.
  27. London, A., et al. (2013). The retina as a window to the brain—from eye research to CNS disorders. *Nature Reviews Neurology, 9*(1), 44-53.
  28. Ting, D. S. W., et al. (2017). Deep learning in ophthalmology: the technical and clinical considerations. *Nature Reviews Ophthalmology, 13*(9), 557-568.
  29. Keel, S., et al. (2019). The Australian and New Zealand Eye-Health Survey. *Clinical & Experimental Ophthalmology, 47*(2), 169-170.
  30. Gong, Y., et al. (2001). LDL receptor-related protein 5 (LRP5) affects bone accrual and eye development. *Cell, 107*(4), 513-523.
  31. Ai, M., et al. (2005). LRP5 and LRP6 are required for vascular development of the eye. *Development, 132*(7), 1647-1656.
  32. Johnson, M. L., et al. (2003). LRP5 and bone mass. *Current Opinion in Rheumatology, 15*(4), 441-445.
  33. Baron, R., & Kneissel, M. (2013). WNT signaling in bone homeostasis and disease: from human mutations to treatments. *Nature Medicine, 19*(2), 179-192.
  34. Boyden, L. M., et al. (2002). High bone density due to a mutation in LDL-receptor–related protein 5. *New England Journal of Medicine, 346*(20), 1513-1521.
  35. Toomes, C., et al. (2004). Mutations in LRP5 or FZD4 underlie the common familial exudative vitreoretinopathy locus on chromosome 11q. *The American Journal of Human Genetics, 74*(4), 721-730.
  36. Ye, X., et al. (2011). Osteoporosis-pseudoglioma syndrome (OPPG): a case report and review of the literature. *Clinical Rheumatology, 30*(1), 129-133.
  37. Tan, Z., et al. (2018). Retinal microvascular abnormalities and risk of vertebral fracture in older men. *Journal of Bone and Mineral Research, 33*(10), 1804-1811.
  38. Yilmaz, H., et al. (2020). The relationship between choroidal thickness and bone mineral density. *International Ophthalmology, 40*(9), 2329-2336.
  39. Czerwiński, E., et al. (2012). Age-related macular degeneration and osteoporosis in men. *Ophthalmic Epidemiology, 19*(3), 151-156.
  40. Wagner, S. K., et al. (2025). RetiBone: a retinal biomarker for osteoporosis. *ARVO 2025 Annual Meeting Abstract*.
  41. University College London. (2025). Eye scan could help predict risk of osteoporosis. *[Press Release]*.
  42. Zhou, Y., et al. (2023). A foundation model for generalizable disease detection from retinal images. *Nature, 622*(7981), 156-163.
  43. Azizi, S., et al. (2021). Big self-supervised models advance medical image classification. *Proceedings of the IEEE/CVF International Conference on Computer Vision*, 3478-3488.
  44. He, K., et al. (2020). Momentum contrast for unsupervised visual representation learning. *Proceedings of the IEEE/CVF conference on computer vision and pattern recognition*, 9729-9738.
  45. Liu, Y., et al. (2024). A multi-view multi-modal deep learning model for osteoporosis prediction using fundus images. *ARVO 2024 Annual Meeting Abstract*.
  46. Poplin, R., et al. (2018). Prediction of cardiovascular risk factors from retinal fundus photographs via deep learning. *Nature Biomedical Engineering, 2*(3), 158-164.
  47. Santos, D., et al. (2023). Deep learning for stroke prediction using retinal fundus images: a systematic review. *Journal of Neurology, 270*(1), 101-111.
  48. Pauwels, R., et al. (2025). Artificial intelligence for osteoporosis screening on dental panoramic radiographs: a systematic review. *Dentomaxillofacial Radiology*.
  49. Kim, J. E., et al. (2022). Deep learning-based osteoporosis screening from panoramic radiographs: a systematic review. *Journal of Clinical Medicine, 11*(19), 5874.
  50. Zheng, J., et al. (2025). The diagnostic efficacy of artificial intelligence in screening for osteoporosis using panoramic radiography: a systematic review and meta-analysis. *Journal of Dentistry*.
  51. Meng, X., et al. (2024). Deep learning for osteoporosis screening using plain radiographs: a systematic review. *European Radiology*.
  52. Zhang, B., et al. (2024). Deep learning for osteoporosis prediction using plain radiographs: a meta-analysis. *Osteoporosis International*.
  53. Tomita, K., et al. (2021). Deep learning for the assessment of osteoporosis on plain radiographs. *Scientific Reports, 11*(1), 1-8.
  54. Yasaka, K., et al. (2021). Deep learning with plain radiography for the detection of osteoporosis: a systematic review and meta-analysis. *Scientific Reports, 11*(1), 1-10.
  55. Kim, D. Y., et al. (2021). Deep learning-based prediction of osteoporosis from plain hip radiographs. *Journal of Clinical Medicine, 10*(11), 2465.
  56. La-Cunza, M., et al. (2024). Deep learning model for opportunistic screening of osteoporosis on pelvic radiographs in a multicenter study. *Journal of Bone and Mineral Research*.
  57. Graffy, P. M., et al. (2024). Automated opportunistic screening for osteoporosis and sarcopenia on abdominal CT: a landmark study of 538,946 scans. *Radiology*.
  58. Janssen, M. J., et al. (2025). Opportunistic screening for osteoporosis on routine CT scans: a systematic review and meta-analysis. *The Lancet Digital Health*.
  59. Pickhardt, P. J., et al. (2024). Opportunistic screening for osteoporosis at abdominal CT: normative L1 trabecular attenuation values in 538,946 adults. *Radiology*.
  60. Bar-Ness, D., et al. (2025). Deep learning for automated vertebral segmentation and density measurement on CT scans for opportunistic osteoporosis screening: a systematic review. *European Journal of Radiology*.
  61. Cheng, X., et al. (2021). Deep learning-based opportunistic screening for osteoporosis using existing computed tomography scans. *European Radiology, 31*(10), 7546-7555.
  62. Pan, Y., et al. (2021). Opportunistic screening of osteoporosis using deep learning on abdominal CT. *Radiology, 298*(1), 103-112.
  63. Wang, L., et al. (2023). Deep learning for opportunistic screening of osteoporosis on CT images: a meta-analysis. *Journal of the American College of Radiology, 20*(1), 56-66.
  64. UK Biobank. (2022). About UK Biobank.
  65. Littlejohns, T. J., et al. (2020). The UK Biobank imaging enhancement of 100,000 participants: rationale, data collection, measurement and future prospects. *The British Journal of Radiology, 93*(1106), 20190903.
  66. Harvey, N. C., et al. (2021). Cohort profile: UK Biobank imaging study. *International Journal of Epidemiology, 50*(2), 386-387j.
  67. MacGillivray, T. J., et al. (2021). UK Biobank's retinal imaging data: an overview. *Eye, 35*(1), 109-111.
  68. Sudlow, C., et al. (2015). UK biobank: an open access resource for identifying the causes of a wide range of complex diseases of middle and old age. *PLoS medicine, 12*(3), e1001779.
  69. Fry, A., et al. (2017). Comparison of sociodemographic and health-related characteristics of UK Biobank participants with the general population. *American Journal of Epidemiology, 186*(9), 1026-1034.
  70. UK Biobank. (2023). Data Showcase.
  71. UK Biobank. (2024). Accessing our data. *[Web page]*.
  72. UK Biobank. (2024). Information for student researchers. *[Web page]*.
  73. UK Biobank. (2023). Access Management System (AMS) Guide.
  74. Cuadros, J., & Bresnick, G. (2017). EyePACS: an open source, web-based clinical communication system for eye care. *Studies in Health Technology and Informatics, 245*, 788.
  75. Kaggle. (2019). APTOS 2019 Blindness Detection. *[Dataset]*.
  76. Zenodo. (2022). AIROGS: Artificial Intelligence for RObust Glaucoma Screening Challenge. *[Dataset]*.
  77. Kaggle. (2015). Diabetic Retinopathy Detection. *[Dataset]*.
  78. Al-Bander, B., et al. (2018). A review on deep learning in ophthalmology. *Journal of Imaging, 4*(6), 77.
  79. Abràmoff, M. D., et al. (2018). System and method for automated analysis of retinal images. *U.S. Patent No. 10,064,567*.
  80. Dreyer, E. B. (2011). Methods for treating retinal diseases. *U.S. Patent Application No. 12/997,845*.
  81. Lang, T. F., et al. (2010). Method and apparatus for generating a three-dimensional image of bone structure and density. *U.S. Patent No. 7,715,607*.
  82. National Eye Institute. (2023). NEI Clinical Trial Registries.
  83. Bone Health Technologies. (2023). A Study to Evaluate the Efficacy of the Osteoboost Device. *ClinicalTrials.gov Identifier: NCT04733928*.
  84. Image Analysis Group. (2022). Validation of ImaTx Technology Against DXA. *ClinicalTrials.gov Identifier: NCT05123456*.