Automatic Classification and 3D Visualisation of Abdominal Aortic Aneurysms to Predict Aneurysm Expansion and Rupture
Yolanda Georgia Koutraki1,2, Rachael O. Forsythe2, Chengjia Wang1,3, Olivia Mcbride2, Jennifer Robson2, Tom J. MacGillivray1, Calum D. Gray1, David E. Newby1,2, and Scott I. Semple1,2

1Clinical Research Imaging Centre, University of Edinburgh, Edinburgh, United Kingdom, 2Centre for Cardiovascular Science, University of Edinburgh, Edinburgh, United Kingdom, 3Toshiba Medical Visualization System-Europe, Edinburgh, United Kingdom

Synopsis

The measurement of the diameter of abdominal aortic aneurysms (AAA) as a criterion for repair has been proved to be imperfect, thus new methods are required. Uptake of Ultrasmall Superparamagnetic Particles of Iron Oxide (USPIO) in AAA has been shown to correlate with aneurysm growth-rate. We previously suggested the use of an automatic AAA classification technique in order to replace manual processing. We have now improved our algorithm to include 3D data analysis and visualisation, multivariate analysis of metrics, batch processing and a Graphical User Interface. We are improving growth prediction with full reproducibility, 40 times faster than before.

Introduction

Abdominal aortic aneurysms (AAA) are responsible for 1-3% of deaths in men aged 65 to 85 in the western world1. Currently decisions for AAA repairs are based on ultrasound measures of the aneurysm diameter (>5.5cm), which is an imperfect criterion since 60% of AAA>5.5 cm never rupture, while 10-20% of AAA< 5.5 cm do rupture2,3. Ruptured AAA cause 80%-90% mortality, so there is an imperative need for better methods to accurately predict AAA expansion and rupture. Richards et al4 demonstrated that uptake of Ultrasmall Superparamagnetic Particles of Iron Oxide (USPIO) in MRI identifies cellular inflammation, while differentiation in patterns of inflammation correlates with aneurysm growth-rate: AAA with distinct mural uptake of USPIO (“inflammatory hotspots”) were found to expand significantly faster. This processing of the data on a 2D slice-by-slice basis however is time-consuming and it uses an empirically-defined threshold which may exclude important information, while inter- and intra-observer variability are introduced by subsequent manual classification. We previously suggested the use of a classification technique6 which automatically detects hotspots of inflammation and classifies AAA. We have now developed our algorithm to include 3D processing of the data. The inflammation throughout the whole volume of the AAA can be quantified and visualised for the first time; this enables us to begin sub-classification of the current groups and higher accuracy of growth prediction in our existing classification. We are also incorporating anatomical measurements to further assist our classification with multivariate analysis. Our algorithm is now included in a Graphical User Interface (GUI) and we have enabled batch processing to greatly reduce classification time.

Materials and Methods

350 patients were imaged using a 3-T MRI Verio (Siemens GmbH, Erlangen) before and 24+ hours after administration of USPIO (Rienso); sub-groups were randomly selected for our algorithm to be tested. A multi-echo, gradient-echo T2*W sequence was used to produce T2* maps to detect the accumulation of USPIO within the AAA. The percentage change in T2* (%ΔT2*) was calculated and displayed as a colour scale. The datasets were registered automatically using a previously described custom algorithm5. Our program was built in MATLAB-R2015a (Mathworks) and uses non-thresholded data. The periluminal area of the AAA is automatically masked. In order to detect ‘hotspots’ of USPIO uptake, an adapted k-means clustering (k=7) algorithm and 2D and 3D-connectivity are applied to the %ΔT2* data. Metrics (eg. lumen size and shape) are calculated using MATLAB and the 3D visualisations are created in MATLAB and Paraview (Kitware).

Results

In the subpopulation of 16 patients initially processed, classification of 12 out of 16 patients was in agreement between the automatic classification and the clinicians’ manual classification (92% of hotspots agreed). However when we checked the outcome of the percentage US growth of the AAA at one year, the automatic classification was more predictive of growth than manual classification (Figure 1). This might be the result of using non-thresholded data in the automatic processing, so that the automatically detected 2D hotspots appear larger and therefore less potential hotspots are discarded (Figures 2, 3). We are now in the process of using the 3D-connectivity between hotspots of different slices and the metrics to subclassify the AAA according to hotspot size and shape. The total processing time with our program for each patient ranges between 70 to 95 seconds. The corresponding processing time by trained observers ranges between 45 to 65 minutes per patient per observer.

Discussion

Our automatic classification program appears to have a high success rate in reproducing the clinicians’ manual classification, while introducing improvements to the process that increase aneurysm growth-rate prediction accuracy. This software may provide clinicians with more automated, robust and fast data processing and can effectively assist in the assessment of future AAA patients. By using non-thresholded data both in 2D and 3D, we obtain more reliable measurements of USPIO uptake, including areas missed in manual processing. The clustering technique used in our algorithm adapts to every individual patient, while the 71% threshold used in the manual processing is population-based. The processing time of the program is approximately 40 times faster than the manual processing, without taking into consideration the extra time needed for observer training. The results are fully reproducible removing inter- and intra-observer variability. With the incorporation of anatomical metrics and 3D connectivity information we have the opportunity to investigate further sub-classifications within the AAA patients. Furthermore, these techniques can be adapted in the future to assist with the imaging of inflammation throughout the body in different clinical application, for example USPIO uptake targeting inflammation post myocardial infarction.

Acknowledgements

This work is funded by the Medical Research Council, British Heart Foundation and the Scottish Universities Physics Alliance INSPIRE award.

References

1. Sakalihasan, N., R. Limet, and O.D. Defawe, Lancet, 2005.

2. Hardman, D., et al., International Journal for Numerical Methods in Biomedical Engineering, 2013.

3. Scott, R.A.P., et al., European Journal of Vascular and Endovascular Surgery, 2001.

4. Richards, J.M.J., et al., 2011.

5. Wang, et al., ISMRM annual meeting, 2014.

6. Y.G. Koutraki, et al., ISMRM annual meeting, 2015.

Figures

Classification of 15 patients into 3 groups based on inflammation patterns; comparison between automatic and manual classification. (growth measured with ultrasound, error bars: standard deviation)

Comparison of manual (by trained observer) against automatic detection of inflammatory hotspots. The hotspots chosen by our automated process appear bigger on each slice and additional hotspots are detected, due to the absence of thresholding.

Automated 3D Hotspot Identification and 3D-connectivity algorithms have been applied to the same difference map, with no threshold. The Hotspots identified by the clinician are now identified as 2 separate 3D Hotspots. The Hotspots chosen by our automated process appear bigger on each slice, due to the absence of thresholding.



Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
2630