Clinically deployed imaging modalities, including MRI, PET, and single-photon emission CT (SPECT), have vastly facilitated the understanding of human brain function (1–3) and the development of disease biomarkers for brain disorders such as brain tumors, Alzheimer’s disease, ischemic stroke, and multiple sclerosis toward personalized medicine (4). MRI sequences for brain imaging, such as diffusion-weighted imaging (DWI), diffusion tensor imaging (DTI) for white matter integrity (5), structural T₁ and T₂ for regional brain atrophy, arterial spin labeling (ASL) for cerebral perfusion, and susceptibility-weighted imaging for microbleed assessment, are routinely performed both in animal models in the laboratory and in human patients in the clinical setting (6). The development of 7 T human MRI allows imaging of the living human brain at the mesoscopic level with a high spatial and temporal signal-to-noise ratio (SNR) (7, 8). There are currently seven Food and Drug Administration approved commercially available MR contrast agents with indications for central nervous system lesions (9). This gadolinium (III)-based contrast agents are widely used in the clinical setting. In addition, targeted agents, activatable agents, high-relaxivity agents, and gadolinium-free MR contrast agents are being developed (9). Gadolinium(III)-based contrast agents have been successful as they provide essential diagnostic information that often cannot be obtained with other non-invasive techniques. PET has been widely used as a highly quantitative non-invasive tool to detect neurotransmitter receptors and protein/enzyme levels, e.g., cerebral glucose metabolism ([¹⁸F]fluorodeoxyglucose, FDG), dopamine receptor ([¹¹C]raclopride), glial activation, and amyloid-β plaques, in the living human brain (10). It is also noted that optical imaging accounts for a large segment of clinical imaging and has been used in image-guided surgery, although far fewer contrast agents have been approved for optical imaging than PET (11, 12). Emerging optical imaging methods, including fluorescence and optoacoustic imaging (OAT) (13, 14), have shown increasing value for assisting diagnosis and surgical navigation. Observation of molecular, structural, and functional changes in the brains of small animal models that recapitulate human diseases is highly valuable for understanding physiological function (15) and the mechanisms underlying brain disorders (16). Small animal brain imaging has been indispensable for the development of novel therapeutic drugs and diagnostic imaging probes. A previous study showed that preclinical dosimetry studies and models facilitate the prediction of clinical doses of new PET tracers (17). Here, we summarize the gaps in the translation of preclinical neuroimaging and focus on the recent technical developments in improving its clinical relevance, especially regarding data acquisition (hybrid imaging and awake animal imaging), data analysis using deep learning (DL), and transcriptomics. We also outline the current outstanding challenges in closing the translational gap and propose outlooks for the future.

The most commonly used hybrid systems in small animal imaging are the integration of molecular imaging using PET, SPECT, fluorescence molecular tomography (FMT), and OAT with structural imaging using MRI and CT. Both PET/SPECT-CT and PET/SPECT-MRI provide research tools for probing molecular and structural information and have demonstrated significant value in brain research (18, 19). Molecular imaging modalities generally lack high-spatial resolution and soft tissue contrast to accurately allocate the distribution of specified molecular signals. Thus, it is essential to provide accurate anatomical information along with the molecular imaging modalities to better interpret the acquired molecular signals (20–22). There are two approaches to address this: sequential-mode using different standalone modalities and follow-up image processing/image registration (23) or hybridized multimodal imaging (24). Sequential-mode multimodal imaging allows convenient data acquisition and minimal interference from each modality, while hybridization of the multimodal method enables dynamic imaging data from multiple channel signal sources, enhanced image reconstruction with prior structural information, and improved quantitative information (25).

Translational gaps exist in developing imaging biomarkers, including those (1) between in vitro and in vivo animal studies, (2) between animal and patient translation as a robust medical research tool, and (3) between research tools and integrated clinical applications (97, 98). For clinical imaging, rapid and safe processes, reliable readouts and the added value to patients (in comparison to the existing option) are important. Imaging biomarker generates added value by facilitating drug development, monitoring of treatment response (99), and reducing cost per quality-adjusted life year gained by enabling early diagnosis (100). Closing the translational gaps requires enormous technical, biological, and clinical validation and also cost effectiveness assessment (101). The lack of a satisfactory animal model is a shared problem for research on brain diseases, including stroke, neurodegenerative diseases, psychiatric diseases, and multiple sclerosis (102). Many reviews have outlined the challenges in animal models and the need for models better mimicking human diseases to improve translational power (103, 104). Using the genome engineering technology CRISPR/Cas9, humanized knock-in animal models are under rapid development. However, species differences exist in the size of the brain, anatomical structure, cerebral cortical folding, parcelation, and connectivity neuron size in humans, non-human primates, and mice (105, 106). Microglia and astrocyte in humans and mice exhibit different vulnerabilities and responses to external stressors (107, 108). The differences in gene expression and protein level between small animals and humans increase the difficulty of translation of targeting ligands (109). Moreover, there are pharmacological and behavioral differences between mice and rats, which need to be considered when interpreting the results. In addition, strain- and substrain-dependent vulnerability to pathological interventions such as permanent focal cerebral ischemia in a mouse model has been documented (110, 111). Bailey et al. showed that the C57BL/6 strain background has an influence on tauopathy progression in the rTg4510 transgenic mouse model originally of the FVB/129 background (112). Moreover, the recent findings highlighted the unique vulnerability of humans to Alzheimer’s disease or primary tauopathy, and the amyloid-β and tau deposits formed in the brains from transgenic mouse models over 1–2 years are structurally different compared with those found in aged patients (113). A similar difference has been reported for α-synuclein aggregates in patients with Parkinson’s disease (114). These differences are reflected in the divergent binding properties observed in tau/amyloid-β/α-synuclein imaging probes binding to the brains of rodent models, non-human primates, and patients (115–117). In addition to the aforementioned developments, recent single-cell tracking PET, a cellular global positioning system, was able to detect a single cell over time in a living mouse and directly related the in vivo imaging pattern to a single cell with its molecular profiles (118). Moreover, Tournier et al. and Nutma et al. developed fluorescence-activated cell sorting to radioligand-treated tissues (FACS-RTT) to reveal the cell origin of the radioligand translocator protein binding (119, 120) (Figure 2F).

Anesthesia leads to alterations in the brain state, cardiovascular physiology, and hemodynamic properties. Functional readouts derived from rs-fMRI (BOLD signal) and ASL-MRI (cerebral blood flow) are especially affected by awake or anesthesia status (121). In addition, PET imaging in rodents, such as [¹⁸F]FDG for cerebral glucose metabolism (122) and [¹⁸F]MPPF for serotonin 1A receptor (123), is also largely influenced by anesthesia, restraint stress, and physiological stability of the animal. A key difference between rodent and human functional imaging is the use of anesthesia to reduce motion during scanning. To overcome this gap, imaging in awake behaving rodents has been increasingly reported in the two-photon optical imaging, functional ultrasound imaging, electrophysiology, PET, and MRI in the recent years (124–127). These imaging platforms enable the integration of molecular-, behavioral-level, and circuit-level understanding of the brain. However, the duration, setup, and imaging type suitable for use in awake rodents are still limited and require extensive habituation and further development.

Deep learning-based methods have recently gained great momentum in both image reconstruction (128) and postprocessing (129, 130). Here, we focus on the DL application in image postprocessing with emphasis on image segmentation in a mono-modality and registration between different modalities. For a standalone modality such as MRI, DL has successfully been used in assisting the diagnosis of brain disease and analyzing the whole brain vasculature (131). Several pipelines for registration and segmentation of high-resolution mouse brain data onto brain atlases have been developed, such as aMAP (132, 133) and AMaSiNe (134). In addition to conventional manual feature extraction, the emerging applications of DL entail an enormous advancement in discovering new characteristic features in data (135–137). Given sufficient training, it can learn complex non-linear functions from high-dimensional and unstructured data (138). The potential of artificial intelligence has already been revealed in medical applications such as computerized diagnosis and prognosis (139). Regardless of sequential-mode multimodal imaging or truly hybrid systems, developing software-based algorithms for accurate and automatic registering molecular information with structural information is of high interest. For small animal PET, SPECT, CT, and MRI data, several established pipelines and commercial software such as ANTs, statistical parametric mapping (SPM), AFNI, PMOD, etc., have been established and routinely used (140–142). For OAT data, challenges remain in reconstruction processing, including segmentation and registration. Different detection configurations generate OAT images with different features. As OAT images have limited soft-tissue contrast, segmentation, and registration are thus difficult and highly dependent on the user experience. Human inputs were involved in the previously reported OAT-MRI registration methods to a certain extent, such as piecewise linear mapping algorithms (82). Gehrung et al. developed an integrated protocol for OAT-MRI registration by using a customized animal holder and landmark-based registration software (86). Ren et al. reported semiautomated OAT-MRI brain imaging data registration software that used active contour segmentation as the first step and an adaptive mutual information-based registration algorithm (84). Hu et al. developed a fully automated registration method for OAT-MRI brain imaging data registration empowered by DL, which consists of (1) two U-net-like neural networks to segment OAT and MRI images (71) and (2) an adaptive neural network to transform these generated OAT/MRI masks (reference). The accuracy and robustness of such DL-based registration have been shown to be comparable with classic methods but at a much higher speed without manual efforts on either landmark selection or boundary drawing (143).

With the affordability of single-cell sequencing, metabolomics, and transcriptomics (144–146), deep phenotyping techniques allow us to elucidate the similarities and differences in the genetics (147) and protein expression in animal models and humans (109, 148–151). Single-cell transcriptomic profiling of aging and Alzheimer’s disease combined with data from amyloid and tau PET, [¹⁸F]FDG PET, and rs-Fmri, and structural MRI to unveil the gene and macroscopic factor interactions and the biological mechanisms underlying Alzheimer’s disease (148) (Figures 2A–E). Spatially resolved transcriptomics in particular, such as fluorescence in situ hybridization, showed promising application in convergent cellular, transcriptomic, and molecular neuroimaging data (149, 152, 153). Allen brain transcriptomics mouse and human datasets provide an excellent platform for exploring the link between mice and humans, integrated with connectivity and histology data (154). Other gene expression databases, such as ‘‘Brain RNA-Seq’’¹ (109) and machine-learning models to improve mouse-to-human inference ‘‘Found In Translation,’’² non-human primate platforms (155) are available to facilitate the interpretation of experimental observations. In addition, the uptake of several tracers, such as [¹⁸F]GE-180 (for glial activation) and [¹⁸F]FDG (for cerebral glucose metabolism), is influenced by sex (156–158). Further transcriptomic analysis may reveal sex-specific molecular differences associated with the uptake of different tracers.

There is a continued need for small animals as a powerful model system to advance neuroimaging and translational brain research (159). The biodistribution and pharmacokinetic information of pharmaceuticals in rodents and ex vivo target validation are prerequisites for phase 1 studies. Recent developments in organoids from induced pluripotent stem cells have introduced a paradigm shift for drug development (160). Cerebral organoids and blood–brain barrier organoids that mimic mouse or human physiology for investigating the permeability of compounds have been developed (161). With the goal of implementing the 3R principle (replacement, refinement, and reduction), recent efforts have been made to assess the behavior of imaging ligands using organ-on-chip systems employing organoids (162, 163). In addition, several recent studies reported systems for the characterization of cellular pharmacokinetics based on microfluidic systems, including a continuously infused microfluidic radioassay (164), microfluidics-coupled radioluminescence microscopy (165), and droplet-based single-cell radiometric assay (166). Further research on in vitro in vivo extrapolation will facilitate the translation and development of imaging ligands toward clinical application.

Imaging methods that non-invasively record rapidly changing functional and molecular imaging data have been crucial for our understanding of human and small animal brains. Recent developments in optogenetics, chemogenomics, two-photon microscopy, and all optical interrogation have improved the understanding of physiological and pathological processes in the rodent brains (167–175). However, these imaging methods are limited to preclinical applications. To further improve the translational power of PET, MRI, and CT and optical imaging, we suggest the following considerations:

In conclusion, developments in hybrid imaging, deep learning in data processing, awake animal imaging, and transcriptomics have greatly improved the translation of preclinical brain imaging. Further efforts on standardization, quality assurance/quality control, data sharing, and on-chip modeling for the translation of preclinical small animal brain imaging from bench to bedside need to be undertaken.

WR and RN wrote the draft. All authors contributed to the manuscript.

LC is employed in Shanghai Changes Tech, Ltd. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.