Brain atrophy, assessed through magnetic resonance imaging (MRI), is a crucial in vivo measure of neurodegeneration and a predictor of disability progression in conditions such as Alzheimer’s disease (AD) (Boublay et al., 2020; Traini et al., 2020; Zhu et al., 2022) and multiple sclerosis (MS) (Fujimori et al., 2020; Koch et al., 2022; Rocca et al., 2017). MRI-based brain volumetry has been widely used in clinical trials to evaluate the efficacy of disease-modifying therapies at a group level (De Stefano et al., 2017; Giorgio and De Stefano, 2013; Kappos et al., 2016; Kolind et al., 2023; Swanson et al., 2021).

Quantifying the volume of the entire brain or specific brain regions using MRI is also valuable for analyzing neurodegeneration in murine disease models. Given the small size of the mouse brain, achieving sufficient spatial resolution, contrast, and signal-to-noise ratio (SNR) for reliable segmentation of brain subregions is challenging and often requires long acquisition times, which can be prohibitive from an animal welfare perspective. An alternative approach involves performing volumetric analysis ex vivo, which benefits from greater spatial resolution and sensitivity due to the lack of constraints on imaging time, the use of radiofrequency (RF) coils tightly fitting the organ, and the absence of movement artifacts (Lerch et al., 2012). High SNR three-dimensional images of isolated brains with isotropic voxel volumes of approximately (25)³ µm³ have been obtained in several hours (Crater et al., 2022; Holmes et al., 2017; Ma et al., 2019), providing data suitable for sophisticated analyses. However, ex vivo tissue preparation, such as formalin fixation, is necessary to preserve the tissue during long scanning times, which may lead to distortions and shrinkage due to the removal of water and other fluids (Ma et al., 2019).

In vivo imaging avoids these drawbacks and allows for longitudinal analyses of structural changes. Repeated non-invasive assessments enhance the statistical power of experiments and enable pharmacological studies to evaluate therapy effects at multiple time points in the same mouse (Obrecht et al., 2023). However, in-life examinations require compromises, as animals must be anesthetized during acquisitions. Published mouse brain volumetry studies using 7 T or 9.4 T scanners with conventional RF coils report measurement times between 12 and 90 min per animal for voxel sizes ranging from (230)³ to (100)³ μm³ (Borg and Chereul, 2008; Holmes et al., 2017; Hussain et al., 2017; MacKenzie-Graham et al., 2012; Smith et al., 2018; Yoo et al., 2023). With helium-cooled cryoprobes that improve SNR (Baltes et al., 2009), acquisition times of 38 min and 6 min have been reported at 9.4 T for voxel volumes of 38x38x250 μm³ (Hamilton et al., 2019) and 68x68x300 μm³ (Pallast et al., 2019), respectively.

To perform volumetric analyses, semantic segmentation algorithms are employed to classify voxels in the acquired images according to anatomical regions. Registration based segmentation algorithms map the acquired images to a space of reference brain volumes (ATLAS) for which manual annotations are available (De Feo and Giove, 2019). Deep learning-based methods involve training neural networks to classify individual voxels. In animal brain models, these methods have been used for semantic segmentation to detect lesions (Valverde et al., 2020), perform skull-stripping (Porter et al., 2024), and conduct volumetric analyses of anatomical regions in mouse brains (De Feo et al., 2021). Image registration methods have also been utilized alongside deep learning models to generate reference datasets (Kohler et al., 2024). Additionally, deep learning based image reconstruction methods have decreased the examination time in clinical MRI, by reducing the required number of averages in the scans (Cochran et al., 2025; Ursprung et al., 2023) or by reconstructing from undersampled k-space data (Calivá et al., 2020; Sriram et al., 2020). Their success in the clinical arena points to a promising application to mouse brain imaging, where anesthesia requirements may demand compromises in acquisition time and voxel size. While previous publications have analyzed the performance of deep learning approaches on standalone test datasets, to the best of our knowledge, the adoption of these techniques for tracking longitudinal changes in animal models of brain pathology has received little attention in the literature. This is a crucial step toward evaluating their potential for application in the preclinical context.

In this work, we demonstrate how deep learning can be used to reliably quantify the volumes of brain regions such as the hippocampus, caudate putamen, and cerebellum in mice from T₂-weighted images with a pixel volume of 78x78x250 μm³ acquired in 4.3 min at 7 Tesla using a conventional radiofrequency (RF) coil. Following validation in healthy C57BL/6 J animals, we evaluate its performance to quantify brain volumetric changes in longitudinal studies corresponding to models of cuprizone-induced demyelination in C57BL/6 J mice, transactive response DNA binding protein 43 (TDP43) mice modeling amyotrophic lateral sclerosis, and in C57BL/6 J mice within an EAE model.

Quantitative and qualitative results for the deep learning model applied to mouse brain volumetry analyses are now presented. We begin by discussing the training and validation processes of the deep learning model for semantic segmentation. Additionally, we provide quality control and reproducibility assessments to evaluate the performance of the trained neural network. Following this, the trained network is applied to volumetric analysis of three animal models: TDP43 transgenic mice, EAE mice, and the cuprizone model.

MRI is inherently slower than other in vivo imaging modalities like ultrasound or computerized tomography due to the acquisition of multi-dimensional k-space data through one-dimensional signals. Over the past decades, efforts to accelerate MRI data acquisition have led to the development of compressed sensing, which enables accurate reconstruction from sparsely sampled k-space data (Feng et al., 2017; Ye, 2019). The FDA’s approval of compressed sensing protocols for clinical imaging attests to the maturity of this approach. Although applications to animals have emerged (Braig et al., 2020; Farias et al., 2018; Wang et al., 2020), routine use in small rodents remains challenging due to the technical optimization required for imaging small structures.

In this study, we pursued an alternative strategy to reduce the acquisition time for mouse brain images at 7 Tesla by using a standard MRI sequence and emphasizing deep-learning-assisted data evaluation. Compared to our previous work (Smith et al., 2018), we optimized the acquisition protocol by reducing the in-plane pixel size to 0.078 × 0.078 mm². Further optimization of the RARE sequence, and/or the use of Fast Imaging with Steady-State Free Precession (FISP) as demonstrated by Gao et al. (2014), combined with deep learning-based image analysis, could lead to additional reductions in acquisition time. Despite being developed independently, our procedure resembles efforts recently reported by Man et al. (2023), demonstrating that deep learning enabled fast three-dimensional MRI of the human brain at low fields. Before applying the algorithm to brain volumetry in animal models, reproducibility assessments were performed in healthy mice. Based on these evaluations, we concluded that the approach demonstrated potential to detect minor volumetric changes expected under pathological conditions.

Following validation and reproducibility assessments, the deep learning-based segmentation approach was used for volumetric analysis of different animal models. The significantly reduced brain volume in TDP43 transgenic mice compared to age-matched wildtype mice (Figure 6) is consistent with the accumulation of TDP43 aggregates in the central nervous system, a common feature of many neurodegenerative diseases such as amyotrophic lateral sclerosis (ALS), frontotemporal dementia (FTD), and Alzheimer’s disease (AD) (Jo et al., 2020; Shen et al., 2023). Our results align with reports of atrophy and inflammatory events (astroglial and microglial reactivities) in both cortical (medial prefrontal cortex) and subcortical (hippocampus) structures in TDP43-related FTD mouse models (Santos-García et al., 2023; Tsai et al., 2010). Interestingly, brain atrophy was present early in our animal model, at 8 weeks of age. Recent studies have shown that transient protein folding targets aggregation in the early phase of TDP43-mediated neurodegeneration (San Gil et al., 2024).

Similar to multiple sclerosis (MS), the EAE model is characterized by inflammation and neurodegeneration in both the spinal cord and brain. Previous work revealed inflammation, demyelination, and neurodegeneration in several brain regions of EAE animals, including the cortex (Clark et al., 2016; MacKenzie-Graham et al., 2012), caudate putamen (Gentile et al., 2015), cerebellum (MacKenzie-Graham et al., 2009), hippocampus (Ziehn et al., 2010), and corpus callosum (Mangiardi et al., 2011). The significantly reduced total brain and cerebellum volumes in EAE mice compared to control animals (Figure 6) are consistent with earlier MRI volumetric assessments (Hamilton et al., 2019; MacKenzie-Graham et al., 2009, 2012; Smith et al., 2018). The significant increase in total brain volume during the study in both experimental groups align with reports of brain weight increases in C57BL/6 male mice up to 12 months of age (Lessard-Beaudoin et al., 2015).

In MS patients MRI has detected atrophy in the same brain areas as those mentioned before for the EAE model (see Matthews et al., 2023 for a recent review). MRI measures of atrophy have been proposed as a complementary approach to lesion assessment, facilitating the prediction of clinical outcomes and assessing treatment responses in MS (Sastre-Garriga et al., 2020). The question arises whether clinical learnings can be reproduced in the EAE animal model. Indeed, MRI volumetry has shown that a sphingosine-1-phosphate (S1P) receptor modulator had neuroprotective effects in MS patients (Kappos et al., 2015; Zivadinov et al., 2018) and in murine EAE (Smith et al., 2018), strengthening the translational/back-translational character of the approach. Performing MRI volumetry in the EAE model is thus crucial for preclinical testing of therapies aimed at reducing neurodegeneration. Deep learning has the potential to impact such studies by significantly reducing acquisition time, as illustrated in this work. This will be even more evident when adopting preclinical chronic models where animals are followed for longer periods to effectively screen novel compounds targeting the progressive phase of MS.

Copper dyshomeostasis has been linked to neurodegenerative diseases (Benetti et al., 2010; Gromadzka et al., 2020). This may explain the brain neurodegeneration observed in young (3-month-old) and adult (6-month-old) female mice receiving cuprizone for 5 weeks (Figures 7A,B), accompanied by significant reductions in MTR in the corpus callosum, indicative of demyelination induced by the copper chelator (Beckmann et al., 2018). Reduced total brain volume assessed post-mortem by the fluid displacement technique and a caliper in formalin-fixed samples has also been reported for the 6-week rat cuprizone model (Abbasi et al., 2024). Design-based stereology on formalin-fixed brain revealed significant volume reduction in the corpus callosum and various subcortical areas, particularly the internal capsule and thalamus, in male mice subjected to cuprizone intoxication for 12 weeks, while brain volumes were not altered after 5 weeks of cuprizone (Hochstrasser et al., 2019). Differences in gender and volumetric assessment conditions (ex vivo vs. in vivo) may contribute to discrepancies between the literature and our data. In particular, formalin fixation used to preserve brain tissue may lead to distortions and shrinkage (Ma et al., 2019). Further investigation is needed to understand the reduced caudate putamen volumes observed in young and adult cuprizone-challenged animals in two independent experiments (Figures 7A,B). The lack of response to cuprizone in 18-month-old mice (Figure 7C) is consistent with higher doses and prolonged feeding durations being required for robust demyelination in aged mice (Gingele et al., 2020; Wang et al., 2013). A higher resistance of mature oligodendrocytes in aged animals against cuprizone-induced apoptosis and reduced phagocytic capacity of aged microglia, resulting in delayed removal of myelin debris (Natrajan et al., 2015; Ruckh et al., 2012), may have contributed to the reduced effect of cuprizone in old mice.

While our work investigates the use of supervised deep learning models for mouse brain segmentation, other approaches have also been reported in the literature. For instance, Ma et al. (2014) developed a multi-atlas framework for rat brain segmentation. Recently, Kohler et al. (2024) demonstrated the potential of brain atlas-driven deep learning models on a large cohort of historical data, comprising more than 11,000 MRI datasets acquired from 9,660 adult rats over 10 years. In our case, we initially evaluated an atlas-based segmentation approach, based on the framework of Ma et al. (2014), to generate segmentation annotations for the deep learning model. However, this approach resulted in inadequate segmentation performance after training, with higher prediction errors near edges. Matching our data including the sub-areas of interest turned out to be challenging and especially for the ventricle and the caudate putamen we were not able to resolve substantial discrepancies and variations from animal to animal. Consequently, we resorted to manual annotation to generate the training dataset. Our iterative approach involved initial training, algorithm application, and the addition to the training set of manually curated annotations, where substantial differences between predictions and observed brain morphology occurred. This method allowed us to use a modest amount of training data while achieving good performance in detecting whole brain and subarea volumes. De Feo et al. (2021) trained an MU-Net convolutional neural network on manually segmented multislice RARE images, thereby achieving higher segmentation accuracy than state-of-the-art multi-atlas segmentation methods. Results were reported for mice of different ages and for various murine Huntington models. Compared to our work, MRI acquisitions performed at 11.7 T took 10.7 min per data set and the slice thickness of the images was of 0.6–0.7 mm. We evaluated the pre-trained MU-Net of De Feo et al. (2021) but it did not provide the expected segmentation results probably because of domain shifts, e.g., different slice thicknesses. There is broad consensus that voxel geometry plays a critical role in determining the volumetric properties of regions of interest. Since each voxel represents an average of the signals within its volume, increased slice thickness leads to the inclusion of signals from multiple tissue types, thereby introducing partial volume effects. This hampers the ability to distinguish between brain subregions and accurately delineate brain boundaries, ultimately reducing segmentation accuracy. While higher in-plane resolution (≤0.1 mm) in mouse brain MRI improves segmentation precision, particularly for small or thin anatomical structures, lower resolutions (≥0.15–0.2 mm) may be sufficient for larger regions (Ma et al., 2008). Notably, the use of more isotropic voxels, where in-plane resolution and slice thickness are more closely matched, has been shown to significantly enhance both segmentation accuracy and anatomical fidelity (Mulder et al., 2019). Therefore, the impact of more isotropic voxel configurations - such as thinner slices combined with lower in-plane resolution - achievable within a comparable scan time of 4 min, relative to the current voxel dimensions of 0.078 × 0.078 × 0.25 mm³, warrants further investigation.

It is important to note that in the present work the convolutional neural network used for segmentation was trained on images without apparent brain lesions. The selection of such datasets and the choice of brain areas for segmentation were dictated by our specific applications and were sufficient for the models of interest. However, for brains displaying tumors or edema, additional training would be required. Similarly, further training would be needed if additional brain regions need to be segmented and/or if there are changes in the geometric parameters of the images. While we anticipate that a moderate number of images would suffice for further training based on our experience with supervised learning, these are interesting directions for future work.

In summary, deep learning-based analyses enabled us to reduce MRI data acquisition time for volumetric assessment from 25 min to 4 min at 7 Tesla using a conventional radiofrequency coil and an established standard protocol, without compromising quality and robustness in detecting mouse brain tissue morphology. The low acquisition time, involving minimal exposure to anesthetics, represents a significant refinement of the experimental procedure, contributing to the 3Rs (refine, reduce, replace) concept of animal welfare (Beckmann and Ledermann, 2017; Obrecht et al., 2023; Russell and Burch, 1959). With this dramatic reduction in acquisition time, brain volumetry can now be easily added to further analyses in the same imaging session without significantly extending its duration. The fully automated image analysis approach is highly beneficial for pharmacological studies on neurodegeneration (Obrecht et al., 2023; Smith et al., 2018), allowing for easy randomization of animals into groups before treatment begins, based on MRI volumetry metrics.