Introduction

So-called ‘black holes’ have been identified as a specific type of highly destructive lesions or white matter hyperintensities, which are T₂ hyperintense and become hypointense on T₁w images with time¹. These are pathologically characterized by severe axonal and neuronal loss associated with chronic and irreversible tissue damage^2-4. In multiple sclerosis (MS), black holes (BH) have been traditionally identified on spin-echo (SE) T₁w images where their presence has been linked to clinical severity^5,6. As 3D imaging has gained momentum, MPRAGE or other gradient-echo T₁w images are more commonly acquired to allow for both tissue and lesion segmentation⁷. However, it is known that the level of lesional hypointensity differs strongly between SE and gradient-echo scans and that a large proportion of lesions on gradient-echo images appears ‘black hole’-like^8,9. Here, we assessed whether a fixed intensity threshold may be identified across participants to match BH identification between gradient-echo (MPRAGE) and SE T₁ scans.

Methods

3T MRI data were collected in 25 relapsing-remitting MS patients enrolled in a non-inferiority trial of two MS treatments (NCT04688788), considering only baseline data prior to treatment initiation. 3D MPRAGE and 3D FLAIR images were collected prior to contrast administration, in addition to 2D SE T₁w 8.5 min post-contrast. All images were acquired with 1x1 mm² resolution in-plane, 3D scans 1 mm through-plane (sagittal), and 2D SE T₁w with 3 mm axial slices (MPRAGE: TE/TR/TI=3.6/2700/1090 ms – SE T₁: TE/TR=8.7/700 ms). First, the MPRAGE image was aligned to the MNI standard space, skull-stripped and bias field corrected. All other images were thereafter registered to the aligned MPRAGE image, skull-stripped and bias corrected (Ants, N4)¹⁰. A fully convolutional neuronal network learning architecture (FLEXCONN) was employed for automated segmentation of white matter lesions, using both FLAIR and. MPRAGE images¹¹. Tissue segmentations from FAST (FSL library)¹² were used to remove spurious lesional voxels outside of the white matter. All T₁w image intensities were normalized to the image’s maximum intensity value. Similar to the approach by Tam et al.¹³, the subset of BHs among all detected T₂ lesions was identified by determining an upper SE T₁ intensity threshold for each lesion considering the local white matter intensity and cerebrospinal fluid signal across the image. Thereafter, lesion maps were thresholded at a series of MPRAGE T₁ intensity values (first 10 intensity values between 0 and white matter, thereafter 50 values around the previously identified intensity corresponding to the maximum dice coefficient). The DICE similarity coefficient and volume difference relative to the reference spin-echo BH mask were computed to estimate the most appropriate T₁ intensity threshold. Using the group’s mean T₁ intensity threshold, the final MPRAGE BH masks were obtained. This procedure was followed using different intensity weights for the SE BH identification (varied between 0.3-0.8).

Results

Using a weight of 0.8 in Tam et al.’s method, corresponding to an optimal agreement between manual BH identification and SE intensity thresholding, we found that the relative intensity on pre-contrast MPRAGE images yielded best agreement to the obtained SE BH masks when using a threshold of 0.44 ± 0.07. Nevertheless, the agreement between SE and gradient-echo BH estimations was moderate, with a median DICE score across the 25 subjects of 0.43 (ranging between 0.195-0.753). Even when using this optimal intensity threshold, MPRAGE images detected on average a 45% greater BH volume than SE scans. When modifying the intensity weight of the SE images to restrict ’black holeness’ definition to lower T₁ intensities, thought to produce better correlation with clinical metrics, the DICE coefficient declined further (median 0.16). Results for all SE-weightings are summarized in Table 1 and Figure 1 compares segmentations with weights of 0.8 and 0.6 in one subject. When applying a common average intensity threshold to all subjects the median DICE across all subjects was reduced to 0.39 for the 0.8 weighting, as the average intensity threshold may be too low or too high for a given MPRAGE image. An example comparison of the effect of the individual and group thresholds is shown in Figure 2.

Discussion and Conclusion

MPRAGE images are known to overestimate BH counts and volume compared to SE estimates. We showed that even when applying intensity thresholds on the MPRAGE images for BH identification, agreement with the traditional BH segmentations from SE T₁w images remained moderate. Our data are in good agreement with a recent study at 1.5T, which also demonstrated over 40% volume difference between SE and gradient-echo scans⁸. Note that due to the cross-sectional nature of the current data, we could not establish whether the identified BHs were persistent BHs. More work is needed to understand how SE T₁ images may be replaced by 3D gradient-echo scans in order to identify BHs in a similar fashion. However, it should also be evaluated which additional information MPRAGE data may provide to potentially identify lesional degrees of damage, and how T₁ relaxation times may be used as supplementary information for the identification and monitoring of BHs^14,15. Importantly, our data showed that the choice of T1 scan is still a relevant consideration at 3T for the estimation of BH, even if scans are now acquired at higher spatial resolution (both in 2D and 3D).

References

[1] van Waesberghe et al. Patterns of lesion development in multiple sclerosis: Longitudinal observations with T1-weighted spin-echo and magnetization transfer MR. AJNR Am J Neuroradiolog 1998;19:675-683.
[2] Bitsch et al. A longitudinal MRI study of histopathologically defined hypointense multiple sclerosis lesions. Ann Neurol 2001;49:793-796.
[3] Brück et al. Inflammatory central nervous system demyelination: correlation of magnetic resonance imaging findings with lesion pathology. Ann Neurol 1997;42:783-793.
[4] van Waesberghe et al. Axonal loss in multiple sclerosis lesions: magnetic resonance imaging insights into substrates of disability. Ann Neurol 1999;46:747-754.
[5] Truyen et al. Accumulation of hypointense lesions (“black holes”) on T1 spin-echo MRI correlates with disease progression in multiple sclerosis. Neurology 1996;47(6):1469-1476.
[6] Paolillo et al. Brain atrophy in relapsing-remitting multiple sclerosis: relationship with ‘black holes’, disease duration and clinical disability. J Neurol Sci 2000;174(2):85-91.
[7] Valcarcel et al. A dual modeling approach to automatic segmentation of cerebral T2 hyperintensities and T1 black holes in multiple sclerosis. NeuroImage Clinical 2018;20:1211-1221.
[8] Dupuy et al. MRI detection of hypointense brain lesions in patients with multiple sclerosis: T1 spin-echo vs. gradient-echo. Eur J Radiol 2015;84:1564-1568.
[9] Lapucci et al. Degree of microstructural changes within T1-SE versus T1-GE hypointense lesions in multiple sclerosis: relevance for the definition of “black holes”. Eur Radiol 2020;30:3843-3851.
[10] Tustison et al. N4itk: improved n3 bias correction. IEEE Trans Med Imag 2010;29(3):1310-1320.
[11] Roy et al. Multiple sclerosis lesion segmentation from brain MRI via fully convoluted neural networks. 2018, Preprint arXiv:1803.09172.
[12] Zhang et al. Segmentation of brain MR images through a hidden Markov random field model and the expectation-maximation algorithm. IEEE Trans Med Imag 2001;20(1):45-47.
[13] Tam et al. The impact of intensity variations in T1-hypointense lesions on clinical correlations in multiple sclerosis. Mult Scler J 2011;17(8):949-957.
[14] Thaler et al. T1-thresholds in black holes increase clinical-radiological correlation in multiple sclerosis. PLoS One 2015;10(12):e0144693.
[15] Thaler et al. T1 recovery is predominantely found in black holes and is associated with clinical improvement in patients with multiple sclerosis. AJNR Am J Neuroradiol 2017;38(2):264-269.