skip to primary navigationskip to content

Nikolaos Tsaousis

Cortical thickness estimation of the proximal femur from multi-view Dual X-ray Absorptiometry

Figure 1: The femoral cortex


My PhD research targets a very significant burden of our society: hip fractures. They are the leading cause of acute orthopaedic hospital admission amongst the elderly, with one out of five women over fifty years old enduring one. One year mortality rates are as high as 33%, with survivors suffering pain, reduced quality of life and disability. Although various preventative therapies are available, patient selection is difficult, and the current state-of-the-art risk assessment tool ignores focal structural defects, such as cortical bone thinning, a critical component in characterizing hip fragility. The cortex is a layer of mineral-rich tissue, just under the surface of the bone (see Figure 1), which is thought to contribute up to 90% to the strength. Moreover, various studies provide strong empirical evidence that fractures may initiate at locations where the cortical layer is eggshell thin and maximum stresses are observed.

Figure 2
Figure 2: Cortical thickness estimation from Computed Tomography (CT)


Currently, cortical thickness can be measured very accurately using Multi-Detector Computed Tomography (MDCT), which provides an information-rich 3-dimensional dataset of the bony structures inside the body (see Figure 2). However, any clinically practical screening protocol cannot rely on MDCT because of its high radiation dose and its high infrastructure and operational costs. Instead, the objective of this research is to use a medical imaging technique called Dual X-ray Absorptiometry (DXA) to create detailed spatial maps of femoral cortical thickness. DXA is currently the preferred imaging modality for assessing hip fracture risk and is used routinely in clinical practice, as its effective radiation is typically two orders of magnitude less than MDCT, and is both cheaper and faster. Unfortunately, DXA scans are 2-dimensional radiographic projections of the tissue (similar to X-rays) lacking all depth information (see Figure 3), leaving us with much sparser data to work with, and consequently making the accurate measurement of the cortex very hard.
Figure 3: Multi-view Dual X-ray Absorptiometry

 




We first developed a tool capable of producing detailed cortical thickness estimates using information from multi-view DXA scans (see Figure 3) and a deformable template model of a femur. Figures 4 and 5 briefly explain the workflow of our approach: we first estimate the cortical thickness at many cross-sections of the bone, and then we present the results as colour-maps for ease of visualisation.

We then examined an alternative approach: we extended our model by incorporating more prior information about the shape and density distribution of the femur using a large training cohort of more than 700 femurs. The statistical models we built capture mathematically common anatomical variations and are used to compensate for the sparseness of information in the DXA datasets. Figures 6-9 show the two most important shape and thickness modes of variation we obtained from our statistical study. Subsequently, these models are fitted to patient's DXA scans using an iterative optimisation algorithm. Similarly to our previous approach, the final result is a colour-mapped femoral surface, where the colour represents the cortical thickness at each location – the results can be seen in Figure 10.


Our results for the clinically relevant experiments are encouraging and it remains to be seen whether the current performance meets clinical requirements; this will depend on the location and size of fracture-predicting regions of interest that are only now beginning to emerge from cohort studies. We hope that this technique will allow early detection of fragile bones and assist in treatment monitoring.

figure4.png
Figure 4: Estimation of cross-sectional cortical thickness


Figure 5: Cortical thickness estimation from Dual X-ray Absorptiometry (DXA), using a deformable template model
Figure 6: Statistical model: first 2 principal shape mode variations
Figure 7: Statistical model: first 2 principal shape mode variations

 

 

 

 

 

 

 

 

 

 

Figure 8: Statistical model: First 2 principal thickness mode variations
 
Figure 9: Statistical model: First 2 principal thickness mode variations
Figure 9: Statistical model: First 2 principal thickness mode variations
Figure 10: Cortical thickness maps, and error maps