Data-driven model for evaluation of cerebrovascular-reserve measurement with hypercapnia BOLD
Lenka Vondráčková1,2, Pawel Krukowski3, Johannes Gerber3, Jennifer Linn3, Jan Kybic2, and Jan Petr1

1Institute of Radiopharmaceutical Cancer Research, Helmholtz-Zentrum Dresden-Rossendorf, Dresden, Germany, 2Center for Machine Perception, Faculty of Electrical Engineering, Czech Technical University in Prague, Prague, Czech Republic, 3Department, University Hospital Carl Gustav Carus, Dresden, Germany

Synopsis

Hypercapnia BOLD with the breath-holding task is a technically easier and more clinically available alternative to cerebrovascular reserve (CVR) mapping than administration of CO2 enriched air using an air-tight mask. The disadvantage is complicated data evaluation in case the subject does not adhere to the breathing protocol completely. Here, a data-driven approach for evaluation is presented that is more robust to protocol deviations and produces a reasonable CVR map in most cases where the standard model-driven approach fails. This is demonstrated on randomized evaluation of CVR maps of a group of 86 subjects with stenosis or vessel occlusion.

Purpose

Current methods for cerebrovascular reserve (CVR) mapping based on BOLD-fMRI use hypercapnia stimulation induced by breathing of CO2 enriched air or by breath-holding. The advantage of breath-holding is an easy implementation and wide clinical availability without the need for a dedicated hardware1. Performance of this method is, however, poor when the subjects do not adhere completely to the breathing protocol and the PETCO2 measurement is not available2. To overcome this drawback, we propose a data-driven method3 for analysis of hypercapnia-BOLD measurements that increases the robustness of the analysis in such cases. This is demonstrated in 86 patients with vessel stenosis/occlusion. A reasonable CVR maps are obtained in most of the cases where the standard analysis failed.

Methods

For a standard analysis with a model-driven approach (MDA), a block function indicating periods of normal breathing and breath-holding convolved with hemodynamic response function (HRF) was used as a regressor. The general delay of subject’s response was estimated using cross-correlation of the subject’s mean response with the regressor. The shifted regressor was then least-square fitted to the response of each voxel to generate a CVR map. Absolute CVR quantification was not possible due to missing PETCO2 information. The direction of the BOLD change rather than the magnitude was thus used to identify pathological regions4. In the data-driven approach (DDA), we have assumed that normal CVR response across the brain is largely similar5,6, and that each patient has at least a single large region that exhibits normal CVR. The mean response in the left and right regions corresponding to the anterior cerebral artery, middle cerebral artery, posterior cerebral artery, and cerebellum was calculated, see Figure 1. The region that had a mean response with the highest Pearson’s correlation coefficient with the MDA regressor was considered as normal. This mean response was used as a regressor in the DDA analysis. The MDA and DDA methods were compared in a group of 86 subjects (acquired in 154 sessions) with stenosis/occlusion of an extra/intracranial vessel. FLAIR and EPI-BOLD sequences were acquired with a 1.5T Siemens Sonata with an 8-channel head-coil. EPI-sequence parameters were: TR/TE 3330/54ms, matrix 384x384, voxel size 3.28x3.28x4mm3, 26 slices, 6 phases of 23.3s breath-holding followed by 46.6s of normal breathing. SPM8 toolbox was used for pre-processing (motion correction, spatial normalization, and smoothing with a 6mm FWHM Gaussian). The CVR maps from both methods were randomized and evaluated by a neuroradiologist with 5-years experience. The reader was completely blinded from the session and patient number, method, clinical findings, and other sequences. The overall quality was graded (0 unreadable, 1 suboptimal, 2 optimal quality). Main criteria were cortical predominance, no motion artifacts, no positive response within the ventricles in absence of plexus choroideus. Pathological regions were identified as having negative response, or response clearly decreased compared to contra-lateral side or the overall response amplitude. The regions with reduced CVR were assessed separately in the right and left hemispheres according to the ASPECTS program.

Results

The quality of the DDA was substantially higher than for the MDA. MDA scored optimal, suboptimal and unreadable on 46, 77, and 31 sessions, respectively. DDA was evaluated optimal on 88, suboptimal on 58 sessions, and unreadable only on 8 sessions, see Figure 2. In 93 sessions, the findings of MDA and DDA were either identical or no pathological regions were detected in either of the sessions (in 13 sessions, DDA scored 1 or 2 and MDA was not readable), see Figure 3. In 30 sessions, the MDA-identified pathological regions were larger than on DDA or more regions were detected, and the extra findings corresponded to the clinical symptoms, to stenosis evaluated from angiography, or to CVR findings from other sessions of the same subject. In 17 sessions, DDA performed better, see Table 1. For the remaining sessions, this could not be decided from the available clinical data.

Discussion and Conclusions

We have demonstrated the feasibility of the data-driven approach. DDA had in general slightly lower sensitivity than MDA. In most cases of different findings, the pathological regions detected by DDA were similar to that of MDA but of smaller size, see Figure 4. Thus the performance and sensitivity of DDA still needs to be verified in a larger clinical population. However, DDA improved the quality of the results substantially especially in the cases scored as unreadable in the standard analysis. This makes it a valuable tool that enables the analysis of problematic subjects from breath-holding hypercapnia measurement where the standard method fails to produce any meaningful CVR map.

Acknowledgements

No acknowledgement found.

References

1. Spano, V. R., D. M. Mandell, and J. Poublanc. CO2 Blood Oxygen Level – dependent MR Mapping of Cerebrovascular Reserve in a Clinical Population: Safety, Tolerability, and Technical Feasibility. Radiology, 266(2):592-8, 2013.

2. Tancredi, Felipe B, and Richard D Hoge. Comparison of cerebral vascular re- activity measures obtained using breath-holding and CO2 inhalation. Journal of cerebral blood flow and metabolism, 33(7):1066-74, 2013.

3. Vondrácková Lenka. Functional MRI of Hypercapnia Data. Master thesis. Czech technical university in Prague, Faculty of electrical engineering, Department of cybernetics, 2015. URL: http://hdl.handle.net/10467/61539

4. Pillai, Jay J., and David J. Mikulis. Cerebrovascular Reactivity Mapping: An Evolving Standard for Clinical Functional Imaging. American Journal of Neuroradiology, 36(1):7-13, 2015.

5. Goode, S.D., S. Krishan, C. Alexakis, R. Mahajan, and D.P. Auer. Precision of Cerebrovascular Reactivity Assessment with Use of Different Quantification Methods for Hypercapnia Functional MR Imaging. American Journal of Neuroradiology. 30(5):972-7, 2009.

6. Chang, Ting Yu, Wan Chun Kuan, Kuo Lun Huang, Chien Hung Chang, Yeu Jhy Chang, Ho Fai Wong, Tsong Hai Lee, and Ho Ling Liu. Heterogeneous Cerebral Vasoreactivity Dynamics in Patients with Carotid Stenosis. PLoS ONE, 8(9):e76072, 2013.

Figures

Figure 1: Seven regions used for the DDA method are on the left. The MDA response and the mean regional responses are on the right. Pathological response is visible in the left middle cerebral artery (MCA) region. The right MCA region has the highest correlation with the MDA response.

Figure 2: Two slices of the CVR map overlaid over the FLAIR image are shown for MDA (left) and DDA (right) methods. Clear improvement from unreadable to optimal quality is visible in one subject (upper row) and from sub-optimal to optimal quality (lower row) in another subject.

Figure 3: Identical findings for both DDA and MDA are shown for two subjects with impaired CVR (middle and bottom row) and one subject with no CVR impairment detected (upper row). Two slices are shown for each method and subject.

Table 1: Total count of sessions by the quality scored on MDA and DDA evaluation (left). Only sessions where the CVR pathological findings were better with DDA (middle) or with MDA (right) are shown. 17 sessions unreadable with MDA were readable with DDA (vice-versa for 4 sessions only).

Figure 4: The region with impaired CVR appears larger in the MDA CVR map. The CVR improves after treatment in the later sessions. The sensitivity of the MDA method is slightly better than for DDA on the second session.



Proc. Intl. Soc. Mag. Reson. Med. 24 (2016)
3801