Deep learning in detection of osteoporosis: insights from recent evidence, and future directions

Sep 29, 2026

Osteoporosis is framed as a serious public health problem often diagnosed only after fragility fractures, despite current modalities being limited by factors such as radiation exposure and access constraints. Recent deep learning (DL) research is discussed for opportunistic screening using routine imaging (e.g., chest X-ray, CT, spine/hip radiographs, and dental films), with multiple models reporting strong diagnostic performance. Persistent implementation challenges are highlighted, including external validation, calibration, cost-effectiveness, regulatory and explainability needs, and equity concerns.

LanguageEnglish

CategoryHealthcare

FormatPDF

Pages2

Reading Time5 min

AccessFree to download

Letter to the Editor

Current diagnostic limitations and rationale for DL

Reported DL performance across imaging modalities

Evidence synthesis via meta-analyses

Challenges for clinical integration

Conclusion and future research priorities

Ethical approval and consent

Letter to the Editor Deep learning in detection of osteoporosis: insightsfrom recent evidence, and future directions Saad Arsalan Wasti, MBBSa , Areesha Wasti, MBBSb , Durjoy Acharjee, MBBSc,*, Hrithick Acharjee, MBBSd To the Editor, Osteoporosis is a silent but serious public health concern. Itisa common disease involving systemic impairment of bone massand microarchitecture, resulting in fragility fractures[1] . Thesefragility fractures drive morbidity, mortality and healthcare cost,yet the disease is commonly diagnosed only after fracture. Currentdiagnostic modalities suffer from limitations, including radiationexposure and limited access, and thus remain underutilized[2] .Recent work is exploring whether deep learning (DL) can utilizeexisting routine imagings [such as chest X-ray, computed tomo -graphy (CT), spine/hip radiographs, and panoramic dental films]to detect osteoporosis without having people face additionalradiation dosages or access-restraintsin diagnosing osteoporosis. A recent hybrid DL model has shown a high accuracy of 86%,with similarly high specificity and precision in detecting osteo -porosis using knee X-rays[3] . Another DL model for opportunis -tic osteoporosis screening from chest radiographs reportedpromising results across multiple settings, exceeding tests suchas calcaneal quantitative ultrasound in performance on externaldata[4] . DL convolutional neural network models to unenhancedchest CT also showed good results in predicting osteoporosis.The Densenet121 model showed area under the curves greaterthan 0.8 in external test sets[5] . In addition, cervical radiograph studies indicated the algorithm’sdiagnostic accuracy, sensitivity, and specificity, and DL’s rate ofcorrected answers were considerably higher than spine surgeons(80.0% compared to 60.6%). The diagnostic yield, too, was higherthan that of the spine surgeons[6] . Magnetic resonance imaging DLstudies also suggested clinical use, expanding DL’s anatomic scopebeyond the traditional hip and knee imagings[7] . Meta-analyses have also combined available evidence assessingthe diagnostic accuracy of DL model–based osteoporosis predic -tion. A recent meta-analysis synthesizing literature up to 2023 aDepartment of Medicine, King Edward Medical University, Lahore, Pakistan,bDepartment of Medicine, Continental Medical College, Lahore, Pakistan,cDepartment of Medicine, Dhaka Medical College and Hospital, Dhaka, Bangladeshand dDepartment of Medicine, Sylhet M.A.G. Osmani Medical College, Sylhet,Bangladesh Sponsorships or competing interests that maybe relevant to content are disclosedatthe end of this article. *Corresponding author. Address: Department of Medicine, Dhaka Medical Collegeand Hospital, Dhaka 1000, Bangladesh. E-mail: durjoyk75acharjee@gmail.com.Copyright © 2025 The Author(s). Published by Wolters Kluwer Health, Inc. This isanopen access article distributed under the Creative Commons Attribution License 4.0(CCBY), which permits unrestricted use, distribution, and reproduction in any medium,provided the original work is properly cited. Annals of Medicine & Surgery (2026) 88:963–964Received 26 October 2025; Accepted 31 October 2025Published online 14 November 2025http://dx.doi.org/10.1097/MS9.0000000000004316 included six studies and reported a sensitivity and specificity of81% and 87%, respectively, which were indicative of goodperformance[8] . It concluded that DL was effective in extractingbone density information from plain radiographs, and suggestedits possible clinical utility in opportunistic screening. The clinical integration of DL models, however, still posesnumerous challenges. Four interdependent aspects need extensiveaddressing. First, greater external validation and calibration isneeded. A model that performs on a single-center test set may failon different populations (age mix, scanner manufacturer, imagingprotocols) . Second, cost-effectiveness should be carefully consid -ered to see whether integr

FAQs

Why is osteoporosis often diagnosed only after fractures, and what limitations affect current diagnostics?

The letter states osteoporosis is frequently diagnosed after fragility fractures, and that existing diagnostic modalities have limitations including radiation exposure and limited access.

Which imaging modalities are mentioned for deep learning–based osteoporosis detection?

The document cites routine imaging such as chest X-ray, computed tomography, spine/hip radiographs, panoramic dental films, and knee X-rays and chest radiographs in specific studies.

What major challenges must be addressed before clinical deployment of deep learning models?

It highlights the need for greater external validation and calibration, careful cost-effectiveness evaluation, additional studies outside Asia, and attention to regulatory, explainability, and equity concerns, including transparency about uncertainty and bias.

Sep 29, 2026

Uploaded by eBook King