2927

Algorithm for automated lesion segmentation in patients with familial Cerebral Cavernous Malformations(CCM)
Sivakami Avadiappan1, Marc Mabray2, Blaine L Hart2, Melanie A Morrison1, Angela Jakary1, Leslie Morrison3, Atif Zafar3, Helen Kim4,5, and Janine M Lupo1

1Department of Radiology and Biomedical Imaging, University of California, San Francisco, San Francisco, CA, United States, 2Department of Radiology, University of New Mexico, Albuquerque, NM, United States, 3Department of Neurology, University of New Mexico, Albuquerque, NM, United States, 4Department of Anesthesia and Perioperative Care, University of California, San Francisco, San Francisco, CA, United States, 5Department of Epidemiology and Biostatistics, University of California, San Francisco, San Francisco, CA, United States

Synopsis

Cerebral cavernous malformations (CCMs) are collections of small blood vessels in the brain that are enlarged and irregular in structure and have clinical manifestations that include seizures and hemorrhage. In this work, we developed and evaluated an automated algorithm for counting and quantifying different sized CCM lesions on SWI images. The total lesion burden increased with overall symptom score in baseline scans from 50 patients. Large lesion burden increased at follow-up in 16/17 cases. Our automated algorithm is a consistent method for counting microbleeds and accurate volume estimation and can thus facilitate lesion burden calculation and tracking in CCM patients.

Purpose

Cerebral cavernous malformations (CCM) are collections of small blood vessels in the brain that are enlarged, irregular in structure, and prone to repetitive microhemorrhages. Familial CCM is a rare, autosomal dominant disease with highly variable severity among individuals with the same genetic mutation or age1. Common symptoms include seizures, headaches, and focal neurological deficits2. Susceptibility-weighted imaging (SWI) has been used to identify the presence of these lesions, with the majority of studies focusing on the location, number, and size of the malformations. Although the number of CCM lesions has been used as a surrogate biomarker for disease severity 3-5, manual counting of lesions is tedious and subject to high variability. Accurate automated routines for CCM lesion detection are lacking, with a prior study reporting a sensitivity of 50% for detecting smaller lesions known as cerebral microbleeds (CMBs) in this population6. Although large CCM lesions are easily identified visually, accurate volume calculation remains a challenge and would benefit from automated methods. The goal of this work was to 1)develop a robust automated pipeline for segmenting (and quantifying) different sized CCM lesions on SWI images and 2)evaluate the resulting quantitative metrics as potential markers of clinical disease severity.

Methods

Subjects: Fifty familial CCM type-1 patients with the Common Hispanic Mutation (Q455X) were scanned with SWI (resolution=1x1x1.5mm, FOV=25.6x19.2cm, TE=20ms, TR=28ms) on a 3T Siemens MR scanner. Seventeen of these patients had a follow-up SWI scan ~3 years later. An experienced neuroradiologist manually annotated the lesions for 8 patients.

Algorithm Development: To identify CCM microbleeds, an algorithm that employed a Fast Radial Symmetry Transform (FRST; Figure 1A) to detect radiation-induced microbleeds on SWI images7 was modified by adding a second pass of region growing to further remove false positive mimics (FPs) using MATLAB. Separate pipelines were developed for detecting medium and large lesions (Figure 1BC) that involved using a Frangi-based thresholding method followed by imposing certain area and circularity criteria. For large lesions, circularity criteria were removed to capture lesions that were irregularly shaped, and the lesion had to exist on at least 2 slices. Probable candidates were fed to a GUI8 for user-guided FP removal, automatic lesion segmentation, and subsequent algorithm validation.

Evaluation: Manual and automated counts were compared and the sensitivity and number of FPs were calculated both before and after manual adjustment by the GUI. A patient-specific correction factor was then determined to remove the majority of remaining FPs. In order to verify the accuracy of the segmentation of medium and large lesions, the manually segmented lesion volumes were compared with the automatic segmentations for 10 lesions. Differences in lesion burden were assessed among clinical symptom scores (range 1-5) with a Kruskal Wallis test and between serial scans using a Wilcoxon signed-rank test.

Results

Figure 2 shows clear demarcation of an irregular-shaped large lesion on consecutive slices, with only a few missed smaller CMBs. Figure 3A,B depicts challenging low contrast CMBs detected by our algorithm and FPs that were caused by: ending, cross-section, or turning points of vessels; artifacts from SWI processing; or ventricles. Missed CMB examples are shown in Figure 3C.

CMB lesions: The algorithm correctly identified 112 of 138 CMBs (sensitivity=82%, ICC=0.95, p<0.0005) and the mean number of FPs before correction was 40/patient. After applying the user-guided GUI, the number of FPs was reduced to 2/patient. However, since the number of FPs increased linearly as a function of true microbleed count (Figure 4B), we were able to correct for the number of FPs detected individually for each patient to create an adjusted CMB count. Figure 4A shows the improvement in sensitivity and reduction of FPs and false negatives(FNs) with the adjusted counts. Automated and manual microbleed counts were in agreement at both baseline and follow-up timepoints.

Larger lesions: The automated algorithm identified all medium and large lesions (sensitivity=1) and the remaining FPs (~7/patient) were successfully removed using the GUI (Figure 4C). The automatically generated lesion volume was slightly less than the manual segmented with higher variability for larger lesions (Figure 4D).

Clinical evaluation: Total lesion burden was found to increase with overall symptom score in the 50 baseline patients (p<0.05; Figure 5A). The large lesion burden was found to significantly increase at follow-up compared to baseline scans in 16 out of 17 cases (p<0.05; Figure 5B).

Conclusion

Our automated algorithm provided accurate quantification of various size CCM lesions which will greatly facilitate lesion burden estimation and tracking for larger correlative studies of disease severity. The pipeline developed also has the potential to assess other types of cerebral vascular lesions.

Acknowledgements

The authors would like to acknowledge the support received from the NINDS grant U54-NS065705 and NICHD grant R01-HD079568

References

1. Riant F, Bergametti F, Ayrignac X, Boulday G, Tournier-Lasserve E. Recent insights into cerebral cavernous malformations: the molecular genetics of CCM. FEBS J. 2010;277:1070–1075

2. Al-Shahi Salman R, Hall JM, Horne MA, Moultrie F, Josephson CB, Bhattacharya JJ, Counsell CE,Murray GD, Papanastassiou V, Ritchie V, Roberts RC, Sellar RJ, Warlow CP (2012) Untreated clinical course of cerebral cavernous malformations: a prospective, population-based cohort study. Lancet Neurol 11(3):217–224

3. Choquet H, Pawlikowska L, Lawton MT, Kim H (2015) Genetics of cerebral cavernousmalformations: current status and future prospects. J Neurosurg Sci 59(3):211–220 2

4. Choquet H, Nelson J, Pawlikowska L, McCulloch CE, Akers A, Baca B, Khan Y, Hart B,Morrison L, Kim H(2014)Association of cardiovascular risk factors with disease severity in cerebral cavernous malformations type 1 subjects with the common Hispanic mutation. Cerebrovasc Dis 37(1):57–63. doi:10.1159/000356839 5. Shenkar R, Shi C, Rebeiz T, Stockton RA, McDonald DA, Mikati AG, Zhang L, Austin C, Akers AL, Gallione CJ, Rorrer A, Gunel M, MinW, Marcondes de Souza J, Lee C, Marchuk DA, Awad IA (2015) Exceptional aggressiveness of cerebral cavernous malformation disease associated with PDCD10 mutations. Genet Med 17(3):188–196. doi:10.1038/gim.2014.97

6. Zou X, Hart BL, Mabray M, et al. Automated algorithm for counting microbleeds in patients with familial cerebral cavernous malformations. Neuroradiology. 2017;59(7):685-690

7. Bian W, Hess CP, Chang SM, Nelson SJ, Lupo JM. Computer-aided detection of radiation-induced cerebral microbleeds on susceptibility-weighted MR images. Neuroimage Clin. 2013;2:282-90. Published 2013 Feb 9. doi:10.1016/j.nicl.2013.01.012

8. Morrison MA, Payabvash S, Chen Y, et al. A user-guided tool for semi-automated cerebral microbleed detection and volume segmentation: Evaluating vascular injury and data labelling for machine learning. Neuroimage Clin. 2018;20:498-505. Published 2018 Aug 4. doi:10.1016/j.nicl.2018.08.002

Figures

Figure1. Pipeline for detecting different sized CCM lesions. For small lesions, the thresholds have been modified to increase sensitivity and reduce FPs. The number of detected CMBs in each stage of the pipeline is shown for a patient with 8 true CMBs. For medium lesions, a Frangi filter was modified to detect blobs rather than cylinders, followed by thresholding. For large lesions, adaptive thresholding was used to first highlight the lesions, followed by additional conditions to eliminate the FPs. All the potential lesion candidates are fed to the GUI where the user labels the true and false positives.

Figure 2. Detected lesions on consecutive slices for a representative patient. Our algorithm worked well in detecting both small and large lesions. The large lesion that is irregular in shape was segmented accurately.

Figure 3.Representative examples of TP, FP, and FN lesions. A.Our microbleed detection algorithm performs well at detecting low contrast CMBs(blue arrows) B.Examples of FPs in the choroid plexus within ventricles, midline veins, and in areas of SWI artifacts. These get detected because of their hypo-intense signal and circular shape. C.Examples of CMBs missed by our algorithm. The more challenging CMBs(red arrows) that are missed are barely visible while more obvious CMBs that the algorithm missed are present in inferior slices(yellow arrows). Some non-circular CMBs that are attached to veins(green arrows) are often missed because they do not meet circularity criteria.

Figure 4. A.For CMB detection, average sensitivity was 0.82 in the first part of the algorithm. FPs were reduced to ~2/patient using GUI. B.The relation between total detected count and the manual count was fitted with a line. The equation of the line is used to calculate the adjusted CMB count based on just the total count of detected lesions(TP+FP). By comparing the adjusted CMB and manual count, the sensitivity is 0.95 and the adjusted FP and FN/person is 1.2 and 1, respectively. C.For medium and large lesions, sensitivity was 1. D.Volume difference between auto and manual segmented lesions.

Figure 5. Comparison of lesion burden with clinical outcome. A. With increasing symptom score, there seems to be an increase in total lesion burden. The overall symptom score has been derived from the sum total of symptoms experienced by the patient. A score of 1 indicates that the symptom is present and 0 absent. The symptoms include headache, hemorrhage, seizures, and focal neurological deficits. B. When comparing large lesion burden between serial scans, 16/17 patients showed an increase at follow-up compared to baseline. The mean time interval between baseline and follow-up scans was 39.4 ± 5.3 months.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)
2927