RPS 2305 - Healthy aging, body composition and prevention: the true potential of AI?

Author Block: J. Y. Anifowose, Z. Li, G. Agarwal, E. Aboagye, B. Ariff, S. Copley, M. Chen; London/UK
Purpose: To develop an automated software for assessing coronary calcification in non-small cell lung cancer (NSCLC) patients using an optimised deep learning nnUNet framework for disease prognostication.
Methods or Background: Cardiovascular risk is higher in NSCLC patients than in the general population, but is often underdiagnosed in clinical practice. Attenuation correction CTs from routinely acquired PET-CT staging scans are ungated unenhanced studies which offer an opportunity to assess this risk without incurring additional radiation exposure or radiology workload. nnUNet is a state-of-the-art deep learning architecture demonstrating superior performance in medical image segmentation applications. We trained nnUNet models for coronary calcification on ungated unenhanced chest CTs (n = 100) from a public domain dataset (Stanford AIMI) and tested them on independent data: attenuation correction CTs of PET-CT scans of NSCLC patients acquired at our multi-centre institution between 2012 and 2018 (n = 287, age: 66.8 ± 10.1, male: female 174:113). The reference truth segmentations were drawn and verified by two radiologists of 8 and 2 years of experience. Models with varying batch sizes and convolutional filters were developed and benchmarked ; with the best performing one selected to develop a composite prognostic predictor, based on model-derived coronary calcification score and significant NSCLC features.
Results or Findings: The best performing nnUNet has a 3D_fullres configuration with batch size of 4 and patch size 28x224x224. All cases of coronary calcifications were successfully detected. Multivariable Cox showed statistical significance of disease histology and stage on patient survival. The composite predictor achieved statistically significant prognostic risk stratification (p-value < 0.05).
Conclusion: An optimised nnUNet framework can facilitate automated coronary calcification assessment on ungated unenhanced CT to support a composite prognostic predictor in NSCLC patients.
Limitations: Retrospective study. Single external validation cohort.
Funding for this study: Academy of Medical Sciences award SGL026∖1024.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Retrospective observational study
IRAS: 243592
REC: 18HH4616

Author Block: M. Jung¹, M. Reisert², H. Rieder², S. Rospleszcz², M. T. Lu¹, F. Bamberg², V. Raghu¹, J. Weiß²; ¹Massachusetts General Hospital, Boston/US, ²Freiburg/DE
Purpose: Chronological age is one of the cornerstones of medical decision-making, but it's an imperfect measure of health. We propose a deep learning framework (MRI-Age) for estimating biological age from MRI and investigated its value in predicting cardiometabolic outcomes in the general population beyond chronological age.
Methods or Background: We used 30,389 individuals from the German National Cohort (NAKO) to develop MRI-Age, which takes MRI-derived volumetric body composition, including subcutaneous (SAT), visceral (VAT), intramuscular adipose tissue (IMAT), and skeletal muscle (SM) from the 1st to 5th lumbar vertebra as input and outputs an age estimate in years. For downstream analyses, we calculated MRI-Age acceleration, defined as an age-specific z-score of the age estimate. We then validated this framework in an external testing set of 36,317 individuals from the UK Biobank (UKBB). Incident outcomes were diabetes, MACE, and all-cause mortality. Multivariable Cox regression assessed the association between MRI-Age acceleration categories “negative” (MRI-Age acceleration <-1), “reference” (MRI-Age acceleration -1 to 1), and “positive” (MRI-Age acceleration >1) and outcomes adjusted for traditional cardiometabolic risk factors in the UKBB.
Results or Findings: In 36,317 UKBB participants (65.1±7.8 years, 51.7% female; median follow-up 4.8 years), we found a higher incidence of diabetes, MACE, and death in individuals with positive MRI-acceleration. In multivariable-adjusted Cox regression, we observed a significant positive association between positive MRI-Age acceleration and diabetes (aHR: 1.87, 95% CI [1.56-2.25], p<0.001), MACE (aHR: 1.26, 95% CI [1.01-1.57], p=0.038), and all-cause mortality (aHR: 1.37, 95% CI [1.09-1.72], p=0.007).
Conclusion: Deep learning-derived biological age from MRI predicts cardiometabolic outcomes in the general population beyond chronological age and cardiometabolic risk factors. Individuals at high MRI-Age could benefit from personalized prevention strategies, lifestyle interventions, and treatment planning.
Limitations: Limited age-range. Predominantly white population.
Funding for this study: This project was conducted with data from the German National Cohort (NAKO) (www.nako.de). The NAKO is funded by the Federal Ministry of Education and Research (BMBF) [project funding reference numbers: 01ER1301A/B/C, 01ER1511D, and 01ER1801A/B/C/D], federal states of Germany, and the Helmholtz Association, the participating universities and the institutes of the Leibniz Association. This research has been conducted using the UK Biobank Resource under Application Number 80337. We thank all participants who took part in the NAKO and UKBB study and the staff of these research initiatives. MJ was funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) - 518480401. VKR was funded by Norn Group Longevity Impetus Grant, NHLBI K01HL168231, and AHA Career Development Award 935176.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Informed consent was obtained from all participants in the UK Biobank and the German National Cohort study. In addition, we received local IRB approval (IRB of the University of Freiburg: 23-1316-S1-retro and 24-1099-S1-retro).

Author Block: M. Jung¹, M. Reisert², H. Rieder², S. Rospleszcz², M. Russe², M. T. Lu¹, F. Bamberg², V. Raghu¹, J. Weiß²; ¹Massachusetts General Hospital, Boston/US, ²Freiburg/DE
Purpose: Body composition (BC) plays an important role in risk estimation in patients with cardiometabolic disease and cancer, but reference curves are missing to place individual measurements in context. We developed a deep learning framework to quantify BC from MRI to calculate reference curves and investigated its value for predicting cardiometabolic outcomes.
Methods or Background: BC extracted from MRI data of the UK Biobank (UKBB) and German National Cohort included 1) subcutaneous (SAT), 2) visceral (VAT), 3) intramuscular adipose tissue (IMAT), 4) skeletal muscle (SM), and 5) SM fat fraction (SMFF). Reference curves were generated using generalized additive models for each BC metric to calculate age, sex, and height-specific z-scores. Multivariable Cox regression assessed the association between z-score categories (low: z<-1; middle: z=-1-1; high: z>1) and outcomes (incident diabetes; major adverse cardiovascular events [MACE]; and all-cause mortality) adjusted for traditional cardiometabolic risk factors in the UKBB.
Results or Findings: Among 66,608 individuals (57.7±12.9 years; BMI: 26.2±4.5 kg/m2, 48.3% female), we observed sex differences in BC volumes and distributions with SAT, VAT, SMFF, and IMAT positively and SM negatively associated with age. We found graded associations between BC z-score categories and health outcomes in the UKBB. In multivariable adjusted Cox regression, z-score risk categories had hazard ratios of up to 2.69 for incident diabetes (high VAT), 1.41 for incident MACE (high IMAT), and 1.49 for all-cause mortality (low SM) compared to middle categories.
Conclusion: BC measures normalized for age, sex, and height are associated with cardiometabolic outcomes beyond traditional risk factors in the general population. We will provide open-source BC reference curves, which may accelerate the clinical translation of BC-based risk assessment for cardiometabolic disease and support future BC research.
Limitations: Study population is predominantly white Western European adults.
Funding for this study: This project was conducted with data from the German National Cohort (NAKO) (www.nako.de). The NAKO is funded by the Federal Ministry of Education and Research (BMBF) [project funding reference numbers: 01ER1301A/B/C, 01ER1511D, and 01ER1801A/B/C/D], federal states of Germany, and the Helmholtz Association, the participating universities and the institutes of the Leibniz Association. This research has been conducted using the UK Biobank Resource under Application Number 80337. We thank all participants who took part in the NAKO and UKBB study and the staff of these research initiatives. MJ was funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) - 518480401. VKR was funded by Norn Group Longevity Impetus Grant, NHLBI K01HL168231, and AHA Career Development Award 935176.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Informed consent was obtained from all participants in the UK Biobank and the German National Cohort study. In addition, we received local IRB approval (IRB of the University of Freiburg: 23-1316-S1-retro and 24-1099-S1-retro).

Author Block: S. Kim¹, D. W. Kim¹, C. Han¹, D. Kim², D. Yoon¹; ¹Seoul/KR, ²Daegu/KR
Purpose: To evaluate AI-derived imaging biomarkers from whole-body MRI in detecting and predicting diabetes mellitus (DM).
Methods or Background: An open-source multi-label segmentation model was applied to Dixon whole-body MRIs from the UK Biobank to segment organs and body compositions. Volume indices (volume/m³) and fat fractions of each structure were calculated automatically.
For DM detection at the time of MRI, logistic regression was performed. Excluding baseline DM, random survival forest analysis was performed for predicting future DM. Area under curve (AUC) and Harrell’s C-index was used. Performance of imaging biomarkers was compared to the Leicester Diabetes Risk Score.
Results or Findings: Among the 2,924 participants, 149 had DM at baseline. Of the 2,775 participants without baseline DM, 28 developed DM and were included in the survival analysis (median follow-up 4.1 years, up to 8.9 years).
For DM detection, adrenal gland volume index, kidney volume index, and pancreatic fat fraction (AUC 0.748, 0.716, and 0.710 respectively) were the top classifiers. The multivariate model, using 10 selected imaging features, achieved AUC of 0.802.
In survival analysis, pancreatic fat fraction, adrenal gland volume index, and torso fat volume index (C-index 0.713, 0.685, and 0.678 respectively) were the top predictors. The multivariate model with six selected imaging features achieved C-index 0.780, outperforming the Leicester Diabetes Risk Score (C-index 0.651). When imaging features were combined with clinical features, performance further improved (C-index 0.794).
Conclusion: AI-derived MRI biomarkers demonstrated strong performance in detecting current DM and predicting future onset, highlighting their potential utility in opportunistic screening.
Limitations: Further validation of the open-source segmentation model is necessary to assess its quantitative and qualitative performance.
Funding for this study: MD-PhD/Medical Scientist Training Program through the Korea Health Industry Development Institute, funded by the Ministry of Health & Welfare, Republic of Korea
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Our institution has received IRB approval for UK Biobank-related research, and any additional ethical considerations are adhered to in accordance with this approval.

Author Block: R. F. Cabini, A. Cozzi, S. Leu, B. Thelen, R. Krause, F. Del Grande, S. Rizzo, D. U. Pizzagalli; Lugano/CH
Purpose: This study aims to develop and validate CompositIA, an automated, open-source pipeline for quantifying body composition scores from thoraco-abdominal CT scans.
Methods or Background: CompositIA consists of three main steps: automatic identification of the L1 and L3 vertebrae, segmentation of image slices at these spinal levels, and quantification of body composition indices. Two Deep Learning models were used: MultiResUNet for detecting CT slices intersecting the L1 and L3 vertebrae, and UNet for segmenting the corresponding axial slices. The pipeline was trained on 205 contrast-enhanced thoraco-abdominal CT scans and tested on an independent dataset of 54 scans. Manual segmentation was performed by two radiology residents, who identified the centers of the L1 and L3 vertebrae and segmented the corresponding axial slices. Performance was evaluated via mean absolute error (MAE) for L1/L3 detection, volumetric Dice similarity coefficient (vDSC) for segmentation, and mean percentage relative error (PRE), regression analysis, and Bland–Altman plots for body composition indices estimation.
Results or Findings: On the independent dataset CompositIA achieved a MAE of about 5 mm in detecting slices intersecting the L1 and L3 vertebrae, with a MAE < 10 mm in at least 85% of cases, and a vDSC greater than 0.85 in segmenting axial slices. Regression and Bland–Altman analyses demonstrated a strong linear relationship and good agreement between automated and manual scores (p values < 0.001 for all indices), with mean PREs ranging from 5.13% to 15.18%.
Conclusion: CompositIA facilitated automated quantification of body composition scores, achieving high precision in independent testing.
Limitations: The main limitation of the study is the small size of the training set.
Funding for this study: Raffaella F. Cabini, Benedikt Thelen, Rolf Krause and Diego U. Pizzagalli were supported financially by the grants ExaTrain (to Rolf Krause), and FIR (to Diego U. Pizzagalli).
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study was approved by the local Ethics Committee (Comitato Etico Cantonale, Repubblica e Cantone Ticino, Switzerland; protocol code 2021-00943). All patients whose CT scans were included in the training set provided informed consent for the participation in the study.

Author Block: J. Garrett¹, M. Lee¹, A. Pyrros², R. Summers³, M. Kattan⁴, P. J. Pickhardt¹; ¹Madison, WI/US, ²Downers Grove, IL/US, ³Bethesda, MD/US, ⁴Cleveland, OH/US
Purpose: To derive and test a CT-biological age (CTBA) model using explainable fully automated abdominal CT-based tissue biomarkers predictive of survival in a large adult population.
Methods or Background: In this retrospective cohort study, an automated suite of explainable CT-based AI algorithms quantifying skeletal muscle (L3 level), fat (L3 level), aortic calcification, bone density, and solid organs (liver/spleen/kidney volume) was applied to a large adult cohort undergoing abdominal CT between January 2001-January 2021. Multivariable Cox proportional hazards regression survival analysis was performed to determine final CT biomarker selection based on index of prediction accuracy (IPA). Using all-cause mortality as a primary outcome, the CTBA model informed only by CT biomarkers and blinded to demographics was compared to a model based on demographic data (chronological age/sex/race). The model was also applied to an external validation cohort of 40,718 adults.
Results or Findings: 123,281 adults (mean age, 53.6 years [SD 17.4]; 47% women) underwent abdominal CT during the study interval. Median post-CT follow-up was 5.3 years (IQR,1.9-10.4 years). CT biomarkers of greatest importance to the model were (in descending order): muscle attenuation, aortic calcification, visceral fat attenuation, and bone density. The CTBA model significantly outperformed demographic data for predicting longevity (IPA=29.2 vs. 21.7; 10-year AUC=0.880 vs. 0.779; p<0.001). Age- and sex-corrected survival HR for highest-vs-lowest risk CTBA quartile was 8.73 (95% CI,8.14-9.36); HR for highest-risk vs remaining quartiles was 3.13 (95% CI,3.04-3.23). CTBA model performed well in the external validation cohort (IPA=28.6; AUC=0.888).
Conclusion: A novel CTBA model informed only by objective fully automated "opportunistically” derived abdominal CT biomarkers outperformed a demographics (CA/sex/race) based model and improves survival prediction.
Limitations: Data from a single large academic medical center were used for model training.
Funding for this study: None
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: IRB Waiver of consent; retrospective analysis.

Author Block: Y. Wen, J. H. Eil, K. A. Borys, J. Kohnke, K. Arzideh, J. Haubold, F. Nensa, O. Pelka, R. Hosch; Essen/DE
Purpose: This study aims to demonstrate the integration of AI-generated body composition and organ measurements from CT images with Fast Healthcare Interoperability Resources (FHIR) to standardise and enhance CT-derived biomarker interoperability.
Methods or Background: FHIR is a widely used interoperability standard that enables health information exchange across different healthcare systems. With the development of AI models, modern AI applications cannot only analyse the data but also generate relevant data for patient monitoring and assessment, such as models for body composition analysis. The missing step in advancing personalised medicine is combining AI-generated healthcare results with an interoperable and standardised format. Therefore, this study integrated the results of the Body and Organ Analysis (BOA) model into FHIR profiles, including measurements of 11 semantic body regions, seven tissues, and 104 landmarks, to streamline and provide interoperability of CT-derived biomarkers in radiology.
Results or Findings: Two FHIR profiles, Body Composition Analysis Observation and Body Structure Volume Observation profiles, have been developed to capture body composition measurements and record the volume of body structure generated from the BOA model, incorporating terminology coding and references to related FHIR resources.
Conclusion: The presented FHIR profiles provide an interoperable format for AI-generated body composition data, standardising the storage and exchange of AI-generated biomarkers derived from CT images. The contributed profiles can also be extended in future work to support other radiological modalities (e.g. MRI) or other image-based AI model biomarkers (e.g. CT-based bone mineral density).
Limitations: The created profiles focus on tissue and organ volumetrics and should be enhanced to include other available image-based markers and imaging modalities.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: This study does not require ethics committee approval, since no identifiable or sensitive patient data was used.

Author Block: C. Bojahr¹, J. Haubold¹, O. Pollok¹, C. S. Schmidt¹, K. A. Borys¹, M. Mancino², L. Umutlu¹, F. Nensa¹, R. Hosch¹; ¹Essen/DE, ²Rome/IT
Purpose: This study aims to adapt CT-based deep learning (DL) models using CycleGAN-based style transfer, enabling accurate body composition analysis (BCA) without extensive manual annotation on T2-weighted MRI sequences.
Methods or Background: This study analyzed data from 120 patients (96 train, 24 test) who underwent whole-body CT and MRI within 48 hours. A CycleGAN was trained to convert CT images to T2-weighted MRIs, producing synthetic MRIs that preserve CT structures with MRI styling. BCA was assessed on the corresponding CT scans using the Body and Organ Analysis (BOA) framework, and 10 body-region class segmentations were transferred to synthetic MRIs to train an initial nnU-Netv2 3D segmentation model. This model was used to generate proposals for all 120 MRIs, which two trainees under guidance of a radiologists (with 8 years of experience) refined. A second model was then trained on the refined segmentations, and evaluated by comparing both models to expert annotations using the Sørensen-DICE score.
Results or Findings: The comparison between the two models (style transfer vs. expert refined) revealed the following DICE-scores: subcutaneous tissue (0.835 vs. 0.978), muscle (0.845 vs. 0.965), abdominal cavity (0.943 vs. 0.988), thoracic cavity (0.895 vs. 0.977), bone (0.774 vs. 0.919), glands (0.576 vs. 0. 899), pericardium (0.697 vs. 0.945), mediastinum (0.731 vs. 0.914), brain (0.894 vs. 0.965), spinal canal (0.886 vs. 0.970) and the average of all classes (0.808 vs. 0.952).
Conclusion: The presented approach shows rapid adaptation of CT BCA models to MRI without manual annotation, achieving notable segmentation performance. When refined by experts, these metrics are further enhanced, enabling precise body composition analysis with reduced annotation effort.
Limitations: Validation of different MRI scanners is necessary to ensure the generalizability and robustness of the proposed method.
Funding for this study: No funding was provided for this study.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Informed consent was waived by the ethics committee due to the retrospective setting.

Author Block: J. Kohnke, K. Schmidt, F. Espeter, K. Pattberg, J. Haubold, F. Nensa, R. Hosch; Essen/DE
Purpose: Acute Respiratory Distress Syndrome (ARDS) is a severe condition with high morbidity and mortality. Early risk assessment is crucial for improving outcomes and guiding treatment. While body composition parameters have recently emerged as prognostic factors, they are not commonly considered. However, image-based Body Composition Analysis (BCA) can help extract relevant information about patients. By leveraging deep learning, these features can be effectively used for enhanced risk stratification using detailed body information.
Methods or Background: Thoracic CT scans from 960 ARDS patients (37.4 % female; median age = 54.7; interquartile range 43.0 - 64.6), were analyzed. The scans were obtained within two days before or after ICU admission. Extracted BCA features include lung volume and sarcopenia marker (muscle volume / bone volume). Based on the features, tertiles were determined separately for both genders (lower tertile vs. others). Kaplan-Meier, Log-Rank, and Cox-regression analyses compared 30-days survival between the tertiles.
Results or Findings: Kaplan-Meier analysis revealed significant differences in survival based on lung volume markers (p = 0.02 for male; p = 0.09 female) and sarcopenia (p = 0.01 male; p = 0.52 female). Cox regression indicated that Lung volume (p = 0.02) and gender (p = 0.01) had significant effects on survival, while sarcopenia (p = 0.07) was slightly not statistically significant for survival.
Conclusion: The results suggest that image based BCA from routine CT imaging could improve risk predictions in ARDS patients by using additional information of the patient's body.
Limitations: Although BCA parameters show promise, generalizability is limited as all data were from a single center, highlighting the need for validation in broader clinical settings. Furthermore, the differences between the results depending on gender require further investigation.
Funding for this study: None
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: This study adhered to all guidelines defined by the approving institutional review board of the investigating hospital. The Institutional Review Board waived written informed consent due to the study's retrospective nature. Complete anonymization of all data was performed before inclusion in the study.

Author Block: J. Garrett¹, M. Golden¹, M. Lee¹, S. Berry², N. Appel³, H. Edgar³, P. J. Pickhardt¹; ¹Madison, WI/US, ²Kalamazoo, MI/US, ³Albuquerque, NM/US
Purpose: To correlate fully automated post-mortem CT (PMCT)-based measures of aortic calcification, skeletal muscle, and intra-abdominal fat of decedents with causes of death and comorbidities.
Methods or Background: Retrospective study of the New Mexico Decedent Image Database (NMDID) with non-contrast PMCT scans between 2010-2017. An automated pipeline of AI-driven algorithms for quantifying skeletal muscle, subcutaneous fat, visceral fat, and aortic calcification (Agatston score) from the abdominal component of PMCT scans was used. Scans with more than minimal decomposition were excluded. Cause of death was categorized as acute or chronic. CT-based biological age was derived using a predetermined model.
Results or Findings: The final cohort included 6638 decedents (mean age 50 ± 18 years; 74% male). Deaths were classified as 80% acute, 10% chronic, and 10% uncertain. Muscle density and area at the L3 level were higher in the acute group compared to the chronic group (26 HU vs. 18 HU, p<0.001; 192 cm² vs. 183 cm², p<0.001) and higher in those without cancer (25 HU vs. 16 HU, p<0.001; 190 cm² vs. 169 cm², p<0.01). Aortic Agatston scores were higher in heart disease deaths (5120 vs. 2098, p<0.001). Diabetic patients had higher L3 visceral fat area (227 cm² vs. 175 cm², p<0.001) and lower muscle density (17 HU vs. 25 HU, p<0.001). The chronic group had a larger biological-chronological age gap than the acute group (median age gap, 19 years vs. 10 years; p<0.001).
Conclusion: Fully automated quantitative CT-based tissue biomarkers from PMCT scans match expectations based on previous studies on living patients and correlate with acuity of death and chronic co-morbidities.
Limitations: The process imperfect of categorizing decedents into “acute” or “chronic” causes of death based on death certificates is imperfect without accounting for all potential medical confounders.
Funding for this study: None
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: IRB Exempt study; non-human subjects per HIPAA

Author Block: U. Zidan¹, N. Bi¹, A. Chandrashekar¹, M. Bown², E. Joviliano³, V. Grau¹, E. R. Ranschaert⁴, R. Lee¹; ¹Oxford/UK, ²leicester/UK, ³São Paulo/BR, ⁴Ghent/BE
Purpose: To develop high-performance ML/DL pipelines for the detection and characterization of cardiomegaly in a diverse international repository of CT scans.
Methods or Background: The AICT consortium (www.aict.ai) consists of 10 clinical sites across 3 continents, contributing CT scans in an agnostic fashion to a common research repository. The ultimate goal is to collect 1 million CT studies, enabling ML/DL training at an unprecedented scale. . This pilot analysis includes the first 5487 unique individuals encompassing 1978 chest CT scans performed from March 2017 to September 2024. Two published models (Superem Total Segmentor) were used to detect cardiomegaly.
Results or Findings: Here we report the findings on cardiomegaly (defined as a cardiothoracic ratio [CTR] > 0.50). Of the 1978 individuals, 1577 did not exhibit cardiomegaly (784 males and 793 females), and 401 had cardiomegaly (178 males and 223 females). The overall prevalence of cardiomegaly is 20%. The prevalence is higher among females (22%) compared to males (19%) (p<0.05). The average age of those with cardiomegaly is on average 70 years old (range: 21-96) [m:65,(21-96); f:75,(36-95); p<0.05]. The mean CTR in those with cardiomegaly is 0.55 (±0.05) [m: 0.55 (±0.05); f: 0.56 (±0.06); p=ns].
Conclusion: The AICT Consortium repository, combined with high-throughput ML/DL analytic pipelines, provides novel insights into the prevalence and demographic distribution of cardiomegaly in a contemporary international cohort. This data enhances our understanding of cardiomegaly epidemiology and supports the development of advanced detection methods.
Limitations: [To be added based on study outcomes]
Funding for this study: Horizon Europe and UK Research Innovation
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: The study was approved by HRA (22/HRA/2302)

Author Block: V. Mathew, D. Pearce, N. Kate Rose, S. Saini, E. Bogoch; Toronto, ON/CA
Purpose: The detection of undiagnosed vertebral compression fractures (VCFs) is critical due to their association with increased risk of future fragility fractures. Primary objective is to evaluate the performance of Nanox.AI HealthOST in identifying incidental VCFs on outpatient chest and abdomen CT scans by assessing sensitivity, specificity, PPV, and NPV. Secondary objective is to quantify missed VCFs on by initial reporting radiologist.
Methods or Background: HealthOST is an AI solution from Nanox.AI, providing automatic image analysis of the spine from CT images to support clinicians in the evaluation and assessment of indicators of osteoporosis. Retrospective analysis on 590 outpatient CT cases from St. Michael’s Hospital at Unity Health Toronto. Two radiologists, including a senior musculoskeletal radiologist established a consensus “gold standard” for comparison with AI results. Two AI thresholds for vertebral height reduction were examined: mild (>20%) and moderate (>25%). Original radiologist reports were reviewed to quantify missed VCFs on these scans.
Results or Findings: At the 20% threshold, AI showed a sensitivity of 91.1%, specificity of 52.7%, PPV of 17.1%, and NPV of 98.2%. At the 25% threshold, sensitivity decreased to 79.9%, while specificity improved to 94.2%, with a PPV of 50.7% and NPV of 98.4%. AI increased fracture detection by 88% compared to initial radiologist findings at the 20% threshold and 92% at the 25% threshold.
Conclusion: Nanox.AI HealthOST shows potential as an effective tool for VCF screening, with high sensitivity at the 20% threshold and improved specificity at 25%. Given the variable specificity and substantial rate of false positives, a secondary review by radiologists is recommended for accuracy. Increased detection rate by the AI in comparison to the initial radiologist report highlights the AI's capability to assist in fracture detection and enhancing diagnostic accuracy.
Limitations: None
Funding for this study: AMGEN Inc.
Has your study been approved by an ethics committee? Yes
Ethics committee - additional information: Ethics committee approval REB# 21-183