Improving the Hip Fracture Risk Prediction with a Statistical Shape-and-Intensity Model of the Proximal Femur

Weakly Supervised Bayesian Shape Modeling from Unsegmented Medical Images

Anatomical shape analysis plays a pivotal role in clinical research and hypothesis testing, where the relationship between form and function is paramount. Correspondence-based statistical shape modeling (SSM) facilitates population-level morphometrics but requires a cumbersome, potentially bias-inducing construction pipeline. Traditional construction pipelines require manual and computationally expensive steps, hindering their widespread use. Furthermore, such methods utilize templates or assumptions (e.g., linearity) that can bias or limit the expressivity of the variation captured by the constructed SSM. Recent advancements in deep learning have streamlined this process in inference by providing SSM prediction directly from unsegmented medical images. However, the proposed approaches are fully supervised and require utilizing a traditional SSM construction pipeline to create training data, thus inheriting the associated burdens and limitations. To address these challenges, we introduce a weakly supervised deep learning approach to predict SSM from images using point cloud supervision. Specifically, we propose reducing the supervision associated with the state-of-the-art fully Bayesian variational information bottleneck DeepSSM (BVIB-DeepSSM) model. BVIB-DeepSSM is an effective, principled framework for predicting probabilistic anatomical shapes from images with quantification of both aleatoric and epistemic uncertainties. Whereas the original BVIB-DeepSSM method requires strong supervision in the form of ground truth correspondence points, the proposed approach utilizes weak supervision via point cloud surface representations, which are more readily obtainable. Furthermore, the proposed approach learns correspondence in a completely data-driven manner without prior assumptions about the expected variability in shape cohort. Our experiments demonstrate that this approach yields similar accuracy and uncertainty estimation to the fully supervised scenario while substantially enhancing the feasibility of model training for SSM construction.

Nonunion scaphoid bone shape prediction using iterative kernel principal polynomial shape analysis

BACKGROUND

The scaphoid is an important mechanical stabilizer for both the proximal and distal carpal columns. The precise estimation of the complete scaphoid bone based on partial bone geometric information is a crucial factor in the effective management of scaphoid nonunion. Statistical shape model (SSM) could be utilized to predict the complete scaphoid shape based on the defective scaphoid. However, traditional principal component analysis (PCA) based SSM is limited by its linearity and the inability to adjust the number of modes used for prediction.

PURPOSE

This study proposes an iterative kernel principal polynomial shape analysis (iKPPSA)-based SSM to predict the pre-morbid shape of the scaphoid, aiming at enhancing the accuracy as well as the robustness of the model.

METHODS

Sixty-five sets of scaphoid images were used to train SSM and nine sets of scaphoid images were used for validation. For each validation image set, three defect types (tubercle, proximal pole, and avascular necrosis) were virtually created. The predicted shapes of the scaphoid by PCA, PPSA, KPCA, and iKPPSA-based SSM were evaluated against the original shape in terms of mean error, Hausdorff distance error, and Dice coefficient.

RESULTS

The proposed iKPPSA-based scaphoid SSM demonstrates significant robustness, with a generality of 0.264 mm and a specificity of 0.260 mm. It accounts for 99% of variability with the first seven principal modes of variation. Compared to the traditional PCA-based model, the iKPPSA-based scaphoid model prediction demonstrated superior performance for the proximal pole type fracture, with significant reductions of 25.2%, 24.7%, and 24.6% in mean error, Hausdorff distance, and root mean square error (RMSE), respectively, and a 0.35% improvement in Dice coefficient.

CONCLUSION

This study showed that the iKPPSA-based SSM exploits the nonlinearity of data features and delivers high reconstruction accuracy. It can be effectively integrated into preoperative planning for scaphoid fracture management or morphology-based biomechanical modeling of the scaphoid.

Phantomless calibration of CT scans for hip fracture risk prediction in silico: Comparison with phantom-based calibration

Finite element models built from quantitative computed tomography images rely on element-wise mapping of material properties starting from Hounsfield Units (HU), which can be converted into mineral densities upon calibration. While calibration is preferably carried out by scanning a phantom with known-density components, conducting phantom-based calibration may not always be possible. In such cases, a phantomless procedure, where the scanned subject's tissues are used as a phantom, is an interesting alternative. The aim of this study was to compare a phantom-based and a phantomless calibration method on 41 postmenopausal women. The proposed phantomless calibration utilized air, adipose, and muscle tissues, with reference equivalent mineral density values of -797, -95, and 38 mg/cm3, extracted from a previously performed phantom-based calibration. A 9-slice volume of interest (VOI) centred between the femoral head and knee rotation centres was chosen. Reference HU values for air, adipose, and muscle tissues were extracted by identifying HU distribution peaks within the VOI, and patient-specific calibration was performed using linear regression. Comparison of FE models calibrated with the two methods showed average relative differences of 1.99% for Young's modulus1.30% for tensile and 1.34% for compressive principal strains. Excellent correlations (R2 > 0.99) were identified for superficial maximum tensile and minimum compressive strains. Maximum normalised root mean square relative error (RMSRE) values settled at 4.02% for Young's modulus, 2.99% for tensile, and 3.22% for compressive principal strains, respectively. The good agreement found between the two methods supports the adoption of the proposed methodology when phantomless calibration is needed.

DXA-based statistical models of shape and intensity outperform aBMD hip fracture prediction: A retrospective study

Areal bone mineral density (aBMD) currently represents the clinical gold standard for hip fracture risk assessment. Nevertheless, it is characterised by a limited prediction accuracy, as about half of the people experiencing a fracture are not classified as at being at risk by aBMD. In the context of a progressively ageing population, the identification of accurate predictive tools would be pivotal to implement preventive actions. In this study, DXA-based statistical models of the proximal femur shape, intensity (i.e., density) and their combination were developed and employed to predict hip fracture on a retrospective cohort of post-menopausal women. Proximal femur shape and pixel-by-pixel aBMD values were extracted from DXA images and partial least square (PLS) algorithm adopted to extract corresponding modes and components. Subsequently, logistic regression models were built employing the first three shape, intensity and shape-intensity PLS components, and their ability to predict hip fracture tested according to a 10-fold cross-validation procedure. The area under the ROC curves (AUC) for the shape, intensity, and shape-intensity-based predictive models were 0.59 (95%CI 0.47-0.69), 0.80 (95%CI 0.70-0.90) and 0.83 (95%CI 0.73-0.90), with the first being significantly lower than the latter two. aBMD yielded an AUC of 0.72 (95%CI 0.59-0.82), found to be significantly lower than the shape-intensity-based predictive model. In conclusion, a methodology to assess hip fracture risk uniquely based on the clinically available imaging technique, DXA, is proposed. Our study results show that hip fracture risk prediction could be enhanced by taking advantage of the full set of information DXA contains.

A statistical shape analysis for the assessment of the main geometrical features of the distal femoral medullary canal

Statistical Shape Models (SSMs) are widely used in orthopedics to extract the main shape features from bone regions (e.g., femur). This study aims to develop an SSM of the femoral medullary canal, investigate its anatomical variability, and assess variations depending on canal length. The canals were isolated from 72 CT femur scans, through a threshold-based segmentation. A region of interest (ROI) was selected; sixteen segments were extracted from the ROI, ranging from 25% of the full length down to the most distal segment. An SSM was developed to identify the main modes of variation for each segment. The number of Principal Components (PCs) needed to explain at least 90% of the shape variance were three/four based on the length of the canal segment. The study examined the relationship between the identified PCs and geometric parameters like length, radius of curvature, ellipticity, mean diameter, and conicity, reporting range and percentage variation of these parameters for each segment. The SSMs provide insights into the anatomical variability of the femoral canal, emphasizing the importance of considering different segments to capture shape variations at various canal length. These findings can contribute for the design of personalized orthopedic implants involving the distal femur.

Effect of Artificial Intelligence or Machine Learning on Prediction of Hip Fracture Risk: Systematic Review

BACKGROUND

Dual energy X-ray absorptiometry (DXA) is a preferred modality for screening or diagnosis of osteoporosis and can predict the risk of hip fracture. However, the DXA test is difficult to implement easily in some developing countries, and fractures have been observed before patients underwent DXA. The purpose of this systematic review is to search for studies that predict the risk of hip fracture using artificial intelligence (AI) or machine learning, organize the results of each study, and analyze the usefulness of this technology.

METHODS

The PubMed, OVID Medline, Cochrane Collaboration Library, Web of Science, EMBASE, and AHRQ databases were searched including "hip fractures" AND "artificial intelligence".

RESULTS

A total of 7 studies are included in this study. The total number of subjects included in the 7 studies was 330,099. There were 3 studies that included only women, and 4 studies included both men and women. One study conducted AI training after 1:1 matching between fractured and non-fractured patients. The area under the curve of AI prediction model for hip fracture risk was 0.39 to 0.96. The accuracy of AI prediction model for hip fracture risk was 70.26% to 90%.

CONCLUSIONS

We believe that predicting the risk of hip fracture by the AI model will help select patients with high fracture risk among osteoporosis patients. However, to apply the AI model to the prediction of hip fracture risk in clinical situations, it is necessary to identify the characteristics of the dataset and AI model and use it after performing appropriate validation.

Site-Specific Differences in Bone Mineral Density of Proximal Femur Correlate with the Type of Hip Fracture

The aim of this study was to investigate whether site-specific differences in bone mineral density (BMD) of proximal femur correlate with the type of hip fracture using quantitative computed tomography. Femoral neck (FN) fractures were classified as nondisplaced or displaced subtypes. Intertrochanteric (IT) fractures were classified as A1, A2, or A3. The severe hip fractures were identified as displaced FN fractures or unstable IT fractures (A2 and A3). In total, 404 FN fractures (89 nondisplaced and 317 displaced) and 189 IT fractures (76 A1, 90 A2, and 23 A3) were enrolled. Areal BMD (aBMD) and volumetric BMD (vBMD) were measured in the regions of total hip (TH), trochanter (TR), FN, and IT of the contralateral unfractured femur. IT fractures exhibited lower BMD than FN fractures (all p ≤ 0.01). However, unstable IT fractures had higher BMD compared with stable ones (p < 0.01). After adjusting for covariates, higher BMD in TH and IT were associated with IT A2 (A1 vs. A2: odds ratios (ORs) from 1.47 to 1.69, all p < 0.01). Low bone measurements were risk factors for stable IT fractures (IT A1 vs. FN fracture subtypes: ORs from 0.40 to 0.65, all p < 0.01). There are substantial site-specific differences in BMD between IT fractures A1 and displaced FN fractures. Higher bone density was associated with unstable IT fracture when compared with stable ones. The understanding of biomechanics of various fracture types could help to improve the clinical management of these patients.