Cerebral small vessel disease (cSVD) is a group of cerebrovascular conditions that account for up to 30% of strokes and are a major cause of vascular contributions to cognitive impairment and dementia (VCID)—an irreversible and progressive form of dementia that has become a significant public health issue (Petty et al., 2000; Bamford et al., 1987; Pantoni, 2010; Pantoni and Gorelick, 2014). Monogenic and idiopathic forms of cSVDs (Rutten-Jacobs et al., 2015; Yamamoto et al., 2011; Choi, 2015; Joutel et al., 2016; Joutel and Faraci, 2014) share clinical manifestations, suggesting common pathogenic mechanisms. Monogenic forms of complex disorders tend to be more severe and occur at younger ages. Thus, monogenic forms of cSVD may represent experimentally tractable settings to provide insight into pathogenic processes underlying common age-related disease subtypes. The genes encoding type IV collagen alpha 1 (COL4A1) and alpha 2 (COL4A2) share a common genetic locus that is reproducibly associated with cSVD hallmarks in large-scale genetic studies including white matter hyperintensities (Malik et al., 2018; Persyn, 2020; Sargurupremraj et al., 2020; Traylor et al., 2021; Mishra et al., 2022; Duperron et al., 2023), enlarged perivascular spaces (Duperron et al., 2023), ischemic stroke (Mishra et al., 2022), and small vessel stroke (Mishra et al., 2022), even after accounting for hypertension (Sargurupremraj et al., 2020), making COL4A1/A2 mutations a promising study subject for revealing the disease-causing mechanisms that eventually lead to cSVD.

COL4A1 and COL4A2 assemble into heterotrimers [α1α1α2(IV)] that are conserved throughout the animal kingdom and are fundamental constituents of specialized extracellular matrix structures called basement membranes. Rare, dominant, coding mutations of COL4A1 and COL4A2 cause monogenic cerebrovascular disease as part of a highly variable multisystem disorder named Gould syndrome. CSVD is one of the most notable features of Gould syndrome and encompasses a constellation of clinical manifestations, including porencephaly, early-onset stroke, leukoencephalopathy, intracranial aneurysms, and recurrent intracerebral hemorrhage (ICH) (Breedveld et al., 2006; Gould et al., 2005; Gould et al., 2006; Sibon et al., 2007; Guey and Hervé, 2022). Missense mutations of highly conserved glycine residues in the COL4A1 and COL4A2 triple helical domain represent the most prevalent class of mutations (Jeanne et al., 2015; Whittaker et al., 2022). Murine models of Col4a1 and Col4a2 mutations faithfully replicate Gould syndrome and allelic heterogeneity was shown to modulate the severity of ICH in an allelic series of Col4a1 and Col4a2 mutant mice (Jeanne and Gould, 2017; Labelle-Dumais, 2019; Kuo et al., 2014).

While the qualitative diagnosis of Gould syndrome relies on molecular genetic analysis (Sondergaard et al., 2017; Guey et al., 2021), neuroradiology, in particular magnetic resonance imaging (MRI), provides valuable early detection and monitorization of the progression of cSVD, serving as an indispensable tool for timely clinical intervention. The typical neuroimaging findings of cSVD include white matter lesions, lacunar infarcts, and hemorrhagic lesions, which can be readily detected by different MRI techniques, such as T2-weighted imaging (T2WI), Fluid Attenuated Inversion Recovery (FLAIR), and susceptibility-weighted imaging (SWI). Allelic heterogeneity in Col4a1 mutant mice offers an opportunity to develop and evaluate multimodal imaging for studying the breadth of phenotypes in this model of monogenic cSVD. This may provide insight into correlations between genotypes, radiological features, and histopathology findings, which can improve the understanding of cSVD generally and Gould syndrome more specifically.

In this study, we used five different Col4a1 mutant mouse strains that recapitulate the clinical spectrum of cerebrovascular manifestations associated with Gould syndrome (Figure 1A). Using this unique allelic series, we characterized the neuroradiology features of cerebrovascular lesions using multimodal 14.1 Tesla MRI. We established a machine learning-based imaging analysis pipeline to quantify radiological changes and used ex vivo studies to elucidate the pathological basis of the respective MR lesions (Figure 1B). Overall, our study demonstrates that multimodal MRI combined with machine learning successfully identify typical cSVD lesions and heterogeneous expressivity in a Col4a1 murine allelic series in terms of anatomical changes, and lesion prevalence, number and volume.

The primary goal of this study was to optimize multimodal MRI pipeline to detect lesions and disease burden in mouse models of a monogenic form of cSVD. To maximize the potential to detect and characterize cSVD lesions, we chose to investigate old animals (52.1 ± 1.26 weeks) as the severity of cerebrovascular manifestations is exacerbated with age in Col4a1 mutant mice (Jeanne et al., 2015).

We used 5 Col4a1 mutant strains that well mimic the disease spectrum observed in individuals with Gould syndrome. Each strain carries a distinct mutation, including 4 different glycine missense mutations in the triple helical domain (Col4a1^+/G658D, Col4a1^+/G912V, Col4a1^+/G1038S, and Col4a1^+/G1180D) and one mutation in the NC1 domain (Col4a1^+/S1582P). While there was no difference in age across all animal groups (Supplementary Figure S1A), 3 out 5 strains have reduced body weight compared to WT (Col4a1^+/+; Supplementary Figure S1B), which is consistent with previous report of reduced body size in Col4a1 mutant mice and might contribute to potential change in overall brain size.

Gould syndrome is a highly variable multisystem disorder caused by COL4A1 and COL4A2 mutations, whose clinical presentation can vary widely between and within families ranging from asymptomatic COL4A1/COL4A2 mutation carriers to severely disabled individuals (Guey and Hervé, 2022). The key manifestation of Gould syndrome in adults is a wide range of cSVD conditions, including porencephaly, ICH of variable severity, and white matter lesions (Guey and Hervé, 2022; Vahedi, 2007; de Vries, 2009). Previously, we successfully demonstrated that Col4a1 and Col4a2 mutant mice can recapitulate pathophysiological hallmarks of Gould syndrome, including ICH, whose severity was modulated by allelic heterogeneity (Jeanne et al., 2015; Jeanne and Gould, 2017). In this study, we show that our multimodal MRI approach can detect radiological features of cSVD in Col4a1 mutant mice that are similar to those observed in individuals with Gould syndrome, and extend our previous findings of allelic heterogeneity contributing to the variability of cerebrovascular manifestations in Col4a1 mutant mice (Jeanne et al., 2015).

COL4A1 mutations were initially discovered in humans as a genetic cause of porencephaly (Gould et al., 2005) and cSVD (Gould et al., 2006). Subsequent studies identified similar consequences for COL4A2 mutations (Yoneda et al., 2012; Jeanne et al., 2012; Verbeek et al., 2012; Gunda et al., 2014). Large scale studies validated COL4A1 mutations as frequent causes of porencephaly and schizencephaly and COL4A1 and COL4A2 as frequent causes of prenatal hemorrhagic or ischemic cerebral lesions (Maurice et al., 2021; Coste et al., 2022). Individuals with pathogenic variants in COL4A1 and COL4A2 had clinical stroke for which hemorrhagic events were the most common causes. Radiological findings include white matter hyperintensities, cerebral hemorrhage, ischemic lesions, and cerebral microbleeds. Less frequently reported findings included brain atrophy, enlarged perivascular spaces, calcification, and cerebral aneurysms (Whittaker et al., 2022). Missense mutations of highly conserved glycine residues in COL4A1 or COL4A2 triple helical domains are the most common class of mutation. Mice with Col4a1 or Col4a2 mutations model the human disease and demonstrate that genetic background and allelic heterogeneity both contribute to the variable expressivity (Jeanne et al., 2015; Labelle-Dumais, 2019; Kuo et al., 2014; Gould et al., 2007; Mao, 2015; Mao, 2021). Here we used different strains from a Col4a1 allelic series in an attempt to model the breadth of clinical manifestations reported in human patients.

The typical neuroimaging lesions detected by our multimodal MRI approach included both enlarged ventricle size and ICH. Previous studies indicate that the enlarged ventricle might be secondary to the recurrence of ICH⁴⁷. In line with this notion, dilated ventricles were mainly found among Col4a1^+/G1038S mice, which also exhibited the highest ICH burden. This is consistent with previously reported histological findings using a larger murine allelic series of Col4a1 and Col4a2 mutations (Jeanne et al., 2015; Kuo et al., 2014). Meanwhile, the majority of ICH lesions across all mutant strains were located in periventricular subcortical regions, similar to what had been reported in individuals with Gould syndrome (Guey et al., 2021). This study also demonstrated that the Col4a1 murine allelic series can model a range of ICH lesion characteristics, including prevalence, number, volume, and bleeding activity. In particular, by the age of 12-month old, Col4a1^+/G1038S mice presented the highest lesion burden, the largest solitary lesion size, and the most extensive affected region among all Col4a1 mutations examined in this study; however, only 1.3% of Col4a1^+/G1038S showed Gd(+) which radiology feature indicates active ICH.

This multimodal imaging modality also detected porencephaly and schizencephaly, two radiological findings associated with Gould syndrome (Meuwissen, 2015), in Col4a1 mutant mice. Porencephaly was detected in one mouse in the Col4a1^+/G1038S background but not in other genotypes (Supplementary Figure S3). To better characterize the lesion, the infiltration of CSF into the corpus callosum was confirmed through combined analysis of TW2I and FLAIR within the same brain slice. Considering porencephaly can cause severe consequences leading to early lethality, its higher prevalence may thus have been precluded in this study, which suggests future MRI study should be conducted at an earlier time point (<12 month) for the pathogenesis of porencephaly. Additionally, schizencephaly, although rarely reported in individuals with Gould syndrome (Meuwissen, 2015), was also detected with low incidence in Col4a1 mutant mice (one Col4a1^+/G1038S mouse, Figure 3). Those two cases suggest that Col4a1^+/G1038S mice may serve as an ideal pre-clinical platform to study the evolution of various radiological phenotypes of Gould syndrome, although a larger sample size is needed for a more comprehensive phenotypic description of other COL4A1 and COLA2 mutations. Notably, periventricular leukoencephalopathy, a frequent MRI finding in individuals with COL4A1 and COL4A2 mutations characterized by FLAIR hyperintensity, was not identified in this study. Since proton tends to have longer T 1 and T 2 values at ultra-high field and the signal-to-noise ratio (SNR) of inversion recovery sequence follows the equation SNR ∝ ( 1 − 2 e − T 1 / T 1 ) . ( 1 − 2 e − TR / T 1 ) .2 e − TE / T 2 , the absence of white matter hyperintensity in this study could be attributed to the current FLAIR sequence lacking enough SNR to afford adequate tissue contrast at 14.1 T. Alternatively, leukoencephalopathy, like porencephaly or schizencephaly, may have a low prevalence in the Col4a1 mutant mice studied here, and further large-scale studies with a wider age range and a more white matter-specific pathological validation (like Luxol Fast Blue staining) are needed to investigate this possibility.

This study was designed as a methodological exploration to establish an imaging platform to characterize phenotypic heterogeneity in a monogenic murine disease model. Now that the neuroimaging and analysis pipelines are established, future studies can characterize disease progression longitudinally for each genotype. In particular, longitudinal MRI can track the formation and evolution of ICH, making it highly efficient to study acute/subacute and chronic-stages of ICH by leveraging SWI and Gd-enhanced T1WI techniques, and to explore potential difference between genotypes.

In this study, two convolutional networks with U-net architecture were implemented to perform the brain structure segmentation on the acquired MRI data, enabling quantitative morphometric analysis. To prevent potential overfitting, we trained and validated both networks on a previously published mouse brain MRI dataset before applying the optimized model to the animals in this study. The first segmentation task was whole brain extraction, commonly referred to as skull-stripping, for which numerous automatic algorithms have been proposed for human brain imaging (Kalavathi and Prasath, 2016) but remained underdeveloped for preclinical research. Given that U-Net-Based segmentation has demonstrated robustness against inter-subject variability and efficiency in handling rodent brain MRI data (Hsu, 2020), it was chosen as an optimal architecture for MRI contrast-indifferent skull-stripping while limited dataset was available in this study. The loss function used in our convolutional networks is Local Weighting-Modified Dice coefficient, which was reported to offer greater robustness than the conventional Dice coefficient as for class-imbalanced segmentation tasks (Sugino, 2021). The same neural network architecture and loss function were used for ventricle segmentation, where both T2WI and FLAIR were fed as input data, which proved sufficient to differentiate CSF from brain tissue. This multi-modality input strategy was also applied in another U-net (not reported here) aimed at quantifying leukoencephalopathy lesions, however, the scarcity of such pathological changes in this study precluded sufficient training.

Although the multimodal MR imaging in this study was performed at non-clinical field strength, our findings suggest that multimodal MRI combined with machine learning holds significant promise for aiding individuals with Gould syndrome in a clinical setting. As demonstrated here, multimodal MRI enabled the identification of subtle brain changes, that potentially enable early detection of the disorder, and personalized treatment strategies tailored to each patient’s unique neurological profile. Moreover, machine learning algorithms can analyze vast amounts of multimodal MRI data to uncover complex patterns and biomarkers indicative of disease progression, facilitating timely interventions and potentially improving patient outcomes. By integrating multimodal MRI and machine learning, clinicians could enhance their understanding of Gould syndrome and develop more effective therapeutic interventions to alleviate its symptoms and improve the quality of life for affected individuals.