Automatic quality measurement of aortic contrast-enhanced CT angiographies for patient-specific dose optimization

The first step of the automatic assessment is the detection of a suitable axial slice of interest (SOI) for each ROI. The aorta and the artery do not have a generalized shape for patients undergoing a CTA examination. Therefore, the algorithm is searching for the following significant anatomical structures (Fig. 3) which are located on the corresponding slices of the ROIs:

For detecting the aforementioned structures in ROI 1 and 3 we applied a template matching approach [5]. The templates were extracted as 2D image patches from the CTA volumes. Before matching the spacing of the templates was adjusted to the CTA data of the assessed patient. The search space was chosen as a subset of all slices in the considered CTA data to optimize the computational demand. To determine SOI 1 the search space was limited to the upper two thirds of axial slices. For SOI 3 the lower two thirds were chosen and at least two templates were used containing the left and right femoral head as a set extracted from the same reference data. The template matching algorithm was applied to each slice of the search space and the value of the similarity function was recorded.

As the HU values in CTA volumes can show a high degree of variability in the artery a normalized cross-correlation measure was implemented as similarity function between the template and the image patches of the considered axial slice extracted in a sliding-window manner. The slice that resulted in the highest value of the similarity function was determined as the SOI regardless of the number of used templates or in case of SOI 3 the number of sets of templates. To measure the similarity for the determination of SOI 3 a set of templates was defined and their mean of their individually obtained cross-correlation values was computed (Fig. 4).

The slice determination of SOI 2 forms an exception. Because of the high variability of the anatomical structures in the abdomen, the appropriate slice for SOI 2 in the standard case is chosen as the mean of the determined slices of SOI 1 and 3. In a few exceptional cases of scans, the CTA volumes do not contain the axial slices suited for SOI 1 or 3. This causes a problem for the aforementioned determination of SOI 2. It must therefore first be established whether the required slices for SOI 1 and 3 exist. Based on the similarity value of the determined slice taken from the template matching a threshold was empirically defined. If the similarity value is below 0.6 the algorithm declares a SOI as non-existing. The algorithm then uses an offset applied to the existing SOI (SOI 1 or SOI 3) to determine SOI 2. The offset was chosen as the halved average distance between ROI 1 and 3 of the expert-annotated axial slices. For our data the offset was therefore 46 slices or 23 cm based on the constant z-spacing of 5 mm.

After all available SOIs are determined the algorithm proceeded with the determination of ROI 1, 2 and 3. Because of the generally round shape of the aorta in the axial plane the circle Hough transform was applied to determine ROI 1 and 2. The smaller size of the arteria femoralis communis leads to the decision to use the vesselness filter [6] for the determination of ROI 3.

By applying the circle Hough transform the CTA image slice is transformed into a Hough space represented as an accumulator matrix parameterized by the radius and the coordinates of the circle center. As the aorta is not the only circular structure (e.g., the vertebral body) a dynamic programming approach [4] computing a cost function for the selection of the most probable aorta was implemented. The six highest accumulations in the Hough space are taken as candidates for the aorta and were used as starting point for the calculation of the cost function. The costs for final circle candidate \(x_k\) with \(k = 1,\ldots , 6 \) is \( C(x_k ) = \sum _{i}^{i+4} C(x_{ki})\) , (i being the according SOI) cumulates the costs of possible circles over adjacent axial slices. The individual cost for each considered circle is defined as followed:where \(C_\mathrm{v}(x_{ki})\) describes the variance within the circle \(x_{ki}\). \(C_\mathrm{r}(x_{k,i-1},x_{ki})\) describes the radius difference of two circles between slice \(i - 1\) and i and \(C_\mathrm{d}(x_{k,i-1}, x_{ki} )\) describes the distance of two circles centers of two adjacent slices. The starting circle of the combination of circles is selected that appears the most constant over the five considered slices resulting in the lowest overall cost. To avoid calcification on the vascular walls the circle radius is reduced by 4 pixels. If no circle achieves a score under 200 no circle will be selected which implies that the aorta will not be detected and has to be determined by the medical practitioner.

Because of the lower diameter of the arteria femoralis communis after applying the vesselness filter to the SOI 3 a thresholding is applied to exclude bone material and calcification. The algorithm returns ROI 1, 2 and 3 as input for the concluding assessment.

Our data consist of 73 contrast-enhanced CTA volumes that can contain pathological findings. These images were generated at the Department of Radiology and Nuclear Medicine located at the UKSH Lübeck.The Siemens SOMATOM Definition AS+ and the ulrich medical CT motion contrast media injector with RIS/PACS interface were used. The administered iodine containing CA dose was 100 ml. As CA Imeron was used at a concentration of 300 mg/ml, it was administered with an injection rate of 5 ml/s.

The volumes have been generated using an aorta scan protocol. The z-spacing is 5 mm for all volumes with varying spacing for the remaining dimensions. All axial slices have a resolution of \(512\times 512\) voxel. The image data and the corresponding metadata were stored in the DICOM format. For each CTA volumes radiological experts annotated the SOIs, the ROIs and the image contrast quality class (Table 2). These annotations provide the reference standard for the evaluation of our provided automatic image contrast measurement.

To increase the generalization of the template matching we created a pool of possible templates taken from our CTA volumes. We tested each combination of templates for the number of templates \(t = 1, \ldots ,5\). As a mean to evaluate the templates and the slice detection we computed the absolute difference in slices between the predicted SOI and the slice chosen by the radiological expert. Our best results shown in Table 3 were achieved by a combination of \(t = 3\) templates for each SOI.

For the validation the CTA volumes from which the templates were taken, were omitted for calculating the results. Figure 5 visualizes the corresponding results. The algorithm detected 56 out of 68 slices for SOI 1 with a only an detection error of max. 2 slices implying that 82.35% were detected within max. 10 mm deviation of the expert’s chosen slice. For SOI 3 the corresponding rate was 82.76%. For SOI 2 40.90% of the detected slices were deviating up to two slices, but regarding a slightly higher max. deviation up to 5 slices or 25 mm still 83.33% fell into this category.

In addition the algorithms ability to confirm or deny the existence of a slice that would be considered as the SOI in the current CTA volume is evaluated. Table 4 shows the number of missing and available slices of SOI 1 and 3 in comparison with the experts choice. The algorithm achieved a sensitivity and specificity of 100% for SOI 1 and a sensitivity of 100% and a specificity of 83.33% for SOI 3. This shows that the algorithm is capable to a rather high degree.

For the evaluation of the ROI determination three possible outcomes of the segmentation were considered. “Correct” was recorded when the ROI was placed within the defined vessel. “Not found” corresponds to the circle conditions (“Automatic ROI determination” section) not being met leading to no set ROI. If a wrong circular anatomical structure (e.g., the left ventricle) was determined as the ROI we assigned the label “incorrect.” For ROI 1 our algorithm reached 78.87% accuracy locating the aorta. An accuracy of 60.67% was achieved for ROI 2 reflecting the higher occurrence of pathological findings like aneurysms and dissections in the abdomen area deforming the overall shape of the aorta. The low accuracy 4.9% for ROI 3 shows that the vesselness filter was the wrong choice and will be replaced in further research. The detailed results can be found in Table 5.

At last, the extent to which the classification of the image contrast results corresponds with the classes the experts chose for our aortic CTA volumes has to be evaluated. In Table 6 are the results of the algorithms results presented in comparison to the experts ones. In 95.89% of the cases our quality measurement chose the same image contrast class. This corresponds to 70 out of 73 CTA volumes. An additional positive note is that if a CTA volume is misclassified the deviation to the correct class is only 1 class.

Considering our focus on potentially giving recommendations for a CA dose reduction the system should be particularly skilled at identifying class 3 cases. We analyzed this scenario by merging class 1 and 2 together to calculate the positive predictive value of class 3 on the resulting two-class problem. This resulted in a positive predictive value of 100% meaning that all cases classified as class 3 by the system actually belong to class 3.