Download the CT2S Service Description (PDF)
The primary function of the human skeleton is biomechanical in nature (support, protection, etc.); thus, its ability to sustain mechanical loading without substantial damage (hereinafter referred to as strength) is a biomarker of primary importance in a large part of related clinical research. Unfortunately, mechanical strength are usually measured directly through invasive and destructive experiments; thus, we can measure the strength of bones only ex vivo, from bone dissected from cadavers. For this reason, in most clinical studies that have bone strength (or related indicators such as risk of bone fracture) as an end point, surrogate biomarkers provided by medical imaging modalities are used instead. The most common are those that measure the average Bone Mineral Density (BMD) over a given region, either areal (aBMD, provided by DXA), or volumetric (vBMD, provided by CT). However, while the BMD is strongly correlated to the bone strength, it is a poor predictor of bone strength, typically with accuracies of 75% or lower (Cody et al., 1999; Dragomir-Daescu et al., 2011; Srinivasan et al., 2012).
Since strength is a mechanical property, regulated by well-established physical laws, in principle it should be possible to build computational models based on such laws that accurately predict for each subject the biomechanical strength of each bone in their skeleton. Prof Marco Viceconti (former Director of the Insigneo Institute) has worked to the development of such technique since 1998. He proposed the first CT scan sequence specifically optimised for 3D reconstruction of the femur (Zannoni et al., 1998), its segmentation to extract the bone geometry (Viceconti et al., 1998), the early techniques to derive from the segmentation accurate finite element meshes (Viceconti et al., 1998), and the first implementation of the Bonemat algorithm to map the heterogeneous tissue properties as derived from the CT images onto the finite element mesh (Zannoni et al., 1998). He then worked extensively on the use of DXA images to predict bone strength (Testi et al., 1999; Testi et al., 2001; Testi et al., 2001; Testi et al., 2002; Testi et al., 2004), reaching the conclusion that only CT-based models could provide the necessary accuracy (Viceconti et al., 2004, Dall'Ara et al., 2016). Recent work has also demonstrated that CT-based models are cost-effective in a public health setting (Li et al., 2023).
Using human bones dissected fresh from cadavers, previous studies have developed a robust experimental set-up to validate the predictive accuracy of these subject-specific CT-based computer models (Taddei et al., 2006; Cristofolini et al., 2007). The first validation study published in 2007, reported an error of 60% (Taddei et al., 2006); refining the methods allowed us to reach in a predictive accuracy of 95% or better in later studies (Schileo et al., 2007; Schileo et al., 2008; Schileo et al., 2008; Juszczyk et al., 2011; Grassi et al., 2012; Schileo et al., 2014; Zani et al., 2015), both under stance and side fall loading conditions.
Once validated, this method has been used to investigate the biomechanical possibility of spontaneous fractures (Viceconti et al., 2012), the safety factor of the femur during level waking (Taddei et al., 2014), the bending strength of paediatric femurs (Li et al., 2015; Altai et al., 2018), and the effect of laterality on femoral strength (Taddei et al., 2016). In all these cases the subject-specific CT-based modelling method provided an accurate, robust and reproducible estimate of bone strength from imaging data, which made these physiological research studies possible.
In 2014, we demonstrated that using the method to estimate the minimum femoral strength under physiological (hereinafter called Minimum Strength in Stance, or MSS), and pathological (hereinafter called Minimum Strength in Fall, or MSF) could accurately predict the risk of femoral neck fracture in a small cohort of osteoporotic women (Falcinelli et al., 2014). In 2016 we confirmed these conclusions on a much larger cohort of 100 women (Qasim et al., 2016); in this retrospective cohort the MSS biomarker was able to correctly identify fractured and non-fractured women only form their CT data, with an accuracy of 79% (Qasim et al., 2016). We have then further improved the MSS prediction, which now classify the same cohort with an accuracy of 83% (Altai et al., 2019). Also, we demonstrated that the MSS biomarker is much more robust than other clinical biomarkers such as aBMD; when we restrict the analysis to “difficult” patients, those with T-score between -1.0 and -2.5, aBMD accuracy is only 60%, while MSS accuracy remains 75%. In the same cohort, in 2018 we computed for each patient their imminent absolute risk of fracture (ARF0) using the femur fall strength values for all hip impact orientations (i.e. not only the minimum or MSF value) and their specific fall characteristics. ARF0 achieved 85% accuracy of classification (Bhattacharya et al., 2019), which improved further to 87% (Aldieri et al., 2022) when the three-dimensional thickness of soft tissues around the hip were measured from CT images instead of being inferred from the patients' BMI.
While the MSS biomarker is clearly superior to any other available clinical biomarker in estimating the bone strength of a subject non-invasively, the derivation of this biomarker from CT data require a set of proprietary software tools and expertise, which our institute can provide. In order to make this biomarker widely available, we are now offering the prediction of MSS from CT data as an on-line service accessible by any clinical research team in the world. This service, named “CT to Strength” or CT2S for short, makes possible to return within 48 hours an accurate prediction of the subject strength from a properly calibrated CT dataset that is uploaded to our on-line service. The following sections in this document gives a brief description of the process.
The most important factor in the achievement of high predictive accuracy is the appropriateness of the medical imaging protocol used. In this situation the CT scanner is not used only to generate pictures that the radiologist has to interpret, but also as a measurement machine, whose outputs are processed quantitatively. For this reason, the first step in our modelling process involves an inspection of the CT data. We reserve the right to reject any cases where the CT data was of poor quality and hence not suitable for the modelling purpose. Our team also has a recommended radiological protocol (see Annex #1: CT scan protocol) for CT imaging. However, if a different protocol was used to obtain the dataset, our team will review the scans and discuss potential mitigation strategies (e.g. phantomless calibration) with each user (Winsor et al., 2021).
For prospective study, at the outset of the study, one of our experts would visit the clinical site where the CT scan will be performed, meet the radiology team (primarily the radiographers) who will be responsible for scanning the patients, and assist the execution of a full scan on the first patient; the resulting protocol should be stored in the memory of the CT system, so as to ensure that all future patients will be scanned using the same protocol. The key elements of the recommended protocol are:
We strongly recommend that all scans in a study be performed with the same CT scanner. If this is not possible, then the European Spine Phantom should be scanned on each CT scanner, and for each scan the indication of which CT machine has been used.
The following data transfer procedure is described based on a typical NHS setting at Sheffield. Details may change for the transfer procedure to work on a different site. This needs to be discussed with the team at the outset of the study in order to determine the best data handling strategy.
The clinical team will directly transfer anonymised patient DICOM images from the Hospital Trust PACS system to the University of Sheffield's secure research storage area. Consequently, all information identifying the patient remains within the Hospitals network, and the referring clinician assigns an independent, anonymous, patient ID within the CT2S web portal when the analysis is requested. This ID is common to that assigned during transfer of images.
The medical staff will enter additional metadata (such as age, gender, weight) within the CT2S web portal. When you upload the data we will ask you to sign a declaration confirming that you have all the necessary legal and ethical permissions to share these data for research purposes; upon request, we will provide you with pro-forma templates for informed consent that we recommend to include, but the ultimate responsibility for ensuring that you are authorised to collect and give us those data will remain entirely with you.
The DICOM data you provide and all associated metadata will be destroyed once our analysis is complete and once the results have been transferred to you. However, in projects that are not commercially sponsored, we ask as a condition for the application of the non-commercial discount the permission to retain all derivative data obtained by the processing and modelling of the original CT image, including voxel level information, for research and service improvement purposes.
Once CT data of a patient have been properly uploaded, and have passed our quality check, an expert operator will process them into a patient-specific model that will be used to generate one or more predictors. Whilst we are happy to discuss ad hoc predictors for specific research purposes, for normal clinical studies we recommend two strength predictors:
A report, in PDF format, will be sent by email to the person requesting the study; in addition, upon request, our operator can log the strength results directly into your clinical research system (such as OpenClinica).
The typical turnaround time from receipt of usable DICOM data to release of results is 48 hours; however, to provide a contingency to accommodate extraordinary events, we reserve the right to return our report within 30 days. Currently we have the capacity to process about 200 cases per year, but this can be increased upon need.
Part of the CT2S pipeline has been automated using a Python GUI (Allison et al., 2026). This automated pipeline is called PyCT2S, and will be available to clinical research teams who wish to install and use this pipeline within their own research environment. This will bypass the need of sending sensitive clinical information outside of the Hospital Trusts.
The PyCT2S pipeline is based on the original CT2S and requires a fully segmented proximal femur geometry as the input. The user will also need to install either the Bonemat Software or LUMA for mapping heterogeneous bone material property, and acquire an Ansys license to run the simulations. Once executed, the user will get a similar report as described above on bone strength for each individual.
At the outset of the project, one of our experts will visit the clinical research team to provide support on software installation and pipeline usage. Our experts will also provide troubleshooting service remotely. An example of this automated pipeline with application to children's bones is available online via the Insigneo Github.
See the CT2S Bibliography for the full list of references cited above.