II. Relaciones intergubernamentales y cambio institucional: aspectos teóricos para el análisis de la dirección/gestión de la Educación Básica en México.
2. Relaciones Intergubernamentales, Políticas Públicas y Conducción/Gestión Intergubernamental (GIG) El contenido de las políticas.
2.4 Naturaleza de las RIG´s y análisis comparado
A series of binary image processing morphological operations has shown to provide a good compromise in terms of defining sufficiently accurate border detection (volume estimation) in a sufficiently short time, even when using a standard desktop PC. This approach was hence adopted in order to allow the tool to detect the vertebral body to minimise the length of calculations and to establish the most distant point in the vertebral cross- section. This approach was tested on different pathologies to assess its versatility and was validated experimentally against volume estimation. The following section details the particular background of such an approach.
A CT image can be represented as a three-dimensional array of attenuation coefficients expressed in grey-scale values. The spatial unit of the image-voxel (a three-dimensional pixel) comprises isotropic single value
information of a specific dimension depending on the scanning parameters and the resolution. Volumetric measurements are derived from the known voxel volume. Secondly, using the simple isometric voxel edge size, the distances and area measurements can be similarly derived.
To define the VB mask (i.e. to define a cross-sectional area slice-by- slice) and to automatically identify the most distant point of the vertebra, a semiautomatic subroutine was compiled in a computational environment using MATLAB [208]. The cross-section of a VB was established using two distinct sub-programs with additional subroutines. In the first program, the image in every slice was binarised based on a histogram cut and then filtered to remove disconnected elements with sequential flood fill morphological operations such as extrusion and dilation. All variables here listed are first predefined by the user according to the scanning resolution used and the quality of the bone, where for resolution of 70.8μm voxel edge size typical values may vary from 400-700mgHA/cm3 and 10-50 voxels for flood fill operations and increasing with increased resolution (due to lower partial volume effect). This step was repeated until the surrounding noise was removed providing clear and smooth boundaries of the vertebral body, as seen in the example in Figure 11 which shows the development and validation of the approach using porcine samples. Consecutively, this was repeated for every slice from the top to the bottom vertebrae with minimal intervention from the user.
However, as the tool is dependent on the difference between the degree of density inside the bone and the background, the automatic boundary estimation can be hampered for samples with a very low bone density or large lesions. The degree of rectitude is automatically assessed by comparing an area in the current slice to the area of the previous slice and if the criterion falls out of the tolerance area specified by the user, the border line is replaced by the one from the previous slice. This approach was shown to be sufficiently expedient and useful in non-uniform pathological samples and especially in samples with a high degree of lytic infiltration. In some cases the boundaries needed to be corrected manually. The manual correction (Figure 13) combined with the previously described self-correcting routine (Figure 12) have been proved to be suitable in all cases where the automatic tool failed to provide the boundaries independently. This subroutine significantly minimises pre-processing time with minimal effect on the final modulus map differences and needed to be deployed only on the more morphologically altered vertebrae.
Figure 11: Vertebral body boundary estimation: a microCT image is first taken and binarised according to a pre-set histogram threshold value, this is followed by a number of morphological operations such as filling the gaps and removing disjointed particles before the final mask is obtained (A). As shown in the GUI snapshot (B) this method can account even for more challenging low density human cadaver samples
B.)
A.)
When the fracture prediction tool is used, the defined boundaries are displayed for every slice indicating the proposed edge as a thin white line superimposed on the original image (such as depicted in Figure 11 and Figure 13 (C)). The user can then determine if the boundaries have been estimated correctly, apply the self-correction subroutine, change the parameters for both subroutines and either re-run the analysis or perform a manual correction. However, it was found that the manual correction was needed only for analysis of the historical data where samples had been scanned in air without removing the air-bubbles. Since then, the self- correction subroutine was shown to be a sufficient tool for capturing incorrectly proposed boundaries.
Figure 12 Manual correction of the VB boundaries combined with area- based filtering algorithm. The area-based filtering algorithm identifies slices where the pixel count of the masked area suddenly changes with respect to the previous slice. If such a step change occurs, the border is replaced with the previous one until the slice where the border area falls back into a user- specified difference between calculated cross-sectional areas is reached
Figure 13 Manual correction of boundary outlines. In case of a “morphologically challenging” sample (e.g. very low trabecular bone density or metastasis), manual correction is required. Such a problem can occur due to a low density gradient between the bone and the background (A). In such cases manual correction (B) allows the user to correct the boundaries to include the whole vertebral body (C)