Coordinated functioning of multiple brain regions is crucial for normal brain function, encompassing perception, cognition, emotions, and mood responses. A significant amount of evidence points to a dysfunctional local circuitry in schizophrenia in the prefrontal cortex (PFC) and its connections with other brain regions, particularly those associated with the limbic system (Lewis et al., 2005). In the past two decades, numerous studies involving neuroimaging techniques like fMRI have yielded compelling findings indicating abnormal activity within the local prefrontal network and disrupted integration of information processes in the PFC and other brain regions among individuals with schizophrenia (Anticevic et al., 2014, 2015; Hunt et al., 2017). Although the evidence supporting disconnectivity in schizophrenia is robust, understanding its causes is complex, and there is ongoing debate regarding its mechanisms and significance concerning clinical symptoms (Gao W. et al., 2022).

Investigations using fMRI have consistently revealed disrupted connectivity in individuals with schizophrenia, both during resting-state conditions and while engaged in specific cognitive tasks (Garrity et al., 2007; Whitfield-Gabrieli et al., 2009; Sheffield and Barch, 2016; Erdeniz et al., 2017; Godwin et al., 2017). More recently, a meta-analysis and an original article reported consistent changes in local functional connectivity in schizophrenia. It was found that patients showed significantly higher Regional Homogeneity (ReHo) in the bilateral medial superior frontal gyrus, while lower ReHo in the bilateral post-central gyrus, right pre-central gyrus, and right middle occipital gyrus (Cai et al., 2022); and differences in the functional connectivity between the salience network and certain brain regions, including the right inferior and middle temporal gyrus, left caudate, and right pre-central gyrus (Huang H. et al., 2022). These findings suggest that there are consistent aberrant local functional connectivity patterns in schizophrenia.

The assessment of functional connectivity in schizophrenia relies predominantly on fMRI scanning data acquired from adult individuals diagnosed with the disorder. However, due to the dependence of fMRI on hemodynamic fluctuations associated with neural activity, it is unsuitable for capturing rapid transitions in brain functional connectivity configurations with high temporal resolution (Jamadar et al., 2021). Moreover, the spatial resolution of this technique is limited to a millimeter scale. As a result, our comprehension of the cellular mechanisms underlying the aberrant brain functional connectivity observed in schizophrenia remains incomplete.

Multiple etiological hypotheses have been proposed to elucidate the abnormal brain connectivity seen in schizophrenia. The dopaminergic hypothesis posits that abnormal dopaminergic neurotransmission contributes to the development and manifestation of schizophrenia (Creese et al., 1976; Toda and Abi-Dargham, 2007). Several lines of evidence support the dopaminergic hypothesis such as alterations in dopamine receptor density and availability in affected individuals revealed by PET and SPECT imaging (Patel et al., 2010). Specifically, an increased number of dopamine D2 receptors has been observed in the limbic striatum. Hyperactivity of D2 receptors in the mesolimbic pathway is thought to contribute to the positive symptoms of schizophrenia (Howes et al., 2009).

The glutamatergic hypothesis was also proposed as an additional perspective on the pathophysiology of schizophrenia (McCutcheon et al., 2020). For instance, decreased glutamate levels have been found in the anterior cingulate cortex and prefrontal cortex, regions implicated in cognitive and emotional processing (Chen et al., 2017). Moreover, PET studies have shown abnormalities in the expression, binding, and availability of glutamate receptors in various brain regions of individuals with schizophrenia (Beck et al., 2021).

It is becoming increasingly evident that the dopaminergic and glutamatergic hypotheses alone fall short of providing a comprehensive explanation for the disorder highlighting the need to consider additional neurochemical systems involved in schizophrenia, such as the GABAergic system (Jahangir et al., 2021). fMRI and PET studies have also provided insights into the altered neural connectivity and network dynamics associated with GABAergic abnormalities in the disorder (Shukla et al., 2019; Marques et al., 2021).

Employing MRI, researchers investigated gray/white-matter contrast in sensory and motor regions of the cortex in schizophrenia revealing reduced myelin in three bilateral sensory and motor regions (Jørgensen et al., 2016). Furthermore, a study employing Diffusion Tensor Imaging (DTI-fMRI) observed significantly lower fractional anisotropy (FA) values in white matter tracts of patients with psychosis compared to healthy controls (Xu et al., 2022). Additionally, the study demonstrated a positive correlation between decreased white matter tract integrity and cognitive performance in patients with psychosis. Electron microscopy of brain tissue from individuals with schizophrenia revealed ultrastructural signs of apoptosis and necrosis in oligodendroglial cells within the cortex and the caudate nucleus with damage to myelin sheath lamellae, and a significant decrease in the nucleus area and volume density of mitochondria (Uranova et al., 2001).

Dendritic spines are the primary sites of excitatory synaptic connections (Papa et al., 1995). As such, alterations to their morphology directly impact the neuronal circuitry within and across multiple brain regions, potentially contributing to the pathogenesis of schizophrenia. Studies on schizophrenia subjects have revealed reductions in dendritic spine density, dendritic arborization and plasticity in several cortical and non-cortical areas (Glantz and Lewis, 2000; Konopaske et al., 2014; MacDonald et al., 2017). By employing confocal microscopy, researchers have investigated the formation, maturation, and pruning of synaptic connections, using in vitro models (Sellgren et al., 2019). Studies in human stem cell-derived neural models have revealed increased synapse elimination and significant developmental and connectivity issues, including the abnormal spread of proliferating neural progenitor cells from the ventricular zone to the intermediate and cortical zones (Stachowiak et al., 2017). Interestingly, maturing neurons were found to be abundantly developed in the deeper neural structure (analogous to subcortical regions) but were notably depleted in surface layers (analogous to the cortical region) of schizophrenia neural organoids.

In vivo mapping of the regional distribution of network abnormalities is a crucial way to define precisely the site of seizure onset. The identification of the site where seizures start and how they propagate is critical to understanding both the pathophysiology of epilepsies and developing therapeutic approaches. Macroscale neuroimaging techniques, such as high-resolution MRI and fMRI, are the entrance step in providing insights into the topological organization of brain networks and connectivity disruptions in epilepsy patients.

Extensive findings have emerged from quantitative structural MRI investigations employing volumetry, voxel-based morphometry, cortical thickness mapping, and structural covariance analysis. In DTI investigations, several parameters can be obtained to characterize white matter microstructure including tractography, tensor-derived metrics, and connectivity matrices (Bartolomei et al., 2005).

In structural MRI, volumetric analysis frequently reveals atrophy in limbic structures, such as the hippocampus, entorhinal cortex, and amygdala, which often correlates with histological evidence of neuronal loss in excised temporal lobe epilepsy (TLE) brain tissue (Bartolomei et al., 2005; Bernhardt et al., 2013). Likewise, post-processing methods in quantitative MRI, such as voxel-based morphometry and cortical thickness analysis, have also revealed that TLE is linked to widespread neocortical irregularities. Covariance analyses of these abnormalities extend beyond mesial temporal structures to comprise prefrontal, frontocentral, cingulate, occipitotemporal, and lateral temporal neocortex (Bernasconi et al., 2004; Bernhardt et al., 2012, 2013). ENIGMA-Epilepsy MRI scans showed gray and white matter changes in different epilepsy types, with more widespread and bilateral extra-hippocampal gray matter differences in left TLE (Whelan et al., 2018; Hatton et al., 2020; Sisodiya et al., 2020). Also, in individuals with TLE, the investigation of preoperative structural connectivity using DTI-fMRI and its association with post-operative seizure control outcomes revealed specific preoperative connectivity patterns that are associated with improved surgical outcomes (Bonilha et al., 2013).

While there has been substantial progress in understanding structural connectivity abnormalities at the macroscale, we face limitations due to our access being restricted to network topology without achieving a finer neuronal resolution and specificity. In this regard, the mesoscale provides a more precise comprehension by pinpointing the particular neural components contributing to local connectivity. Thus, the gold standard for noticing abnormal structural connectivity in a mesoscale is anterograde and retrograde viral neuronal tracing (Lanciego and Wouterlood, 2020). These tracers exhibit high accuracy and sensitivity, especially when mapping long-range connections, thus contributing to a comprehensive and detailed understanding of connectivity across various brain areas (Saleeba et al., 2019). Their invasive nature restricts the use to animal models. Du et al. (2017) employed a rat model induced by pilocarpine and utilized rabies tracing techniques to discern intricate morphological details of projections within the dynamic hippocampal circuit. This study revealed that newly formed dentate granule cells (DGCs) in adults, triggered by seizures, receive excitatory signals from pyramidal cells in the cornu Ammonis (CA3) and repeated excitatory inputs from other DGCs.

In fMRI, Englot et al. (2016) explored local and distant synchronization of resting-state fMRI signals in TLE and focal epilepsy patients. They observed altered connectivity within and between various brain regions, highlighting the impact of epilepsy on network organization. Likewise, analysis of resting state in focal cortical dysplasia (FCD) identified distinct patterns of functional connectivity with the hypo-connected patterns in cases with FCD type IIB, whereas the hyperconnected lesions were predominantly associated with type IIA (Hong et al., 2019).

Alterations in the size and shape of neuronal cell bodies have been detected across diverse brain regions, encompassing the hippocampus, neocortex, and other regions linked to abnormal neural connectivity (Stouffer et al., 2016). The connection between these morphological alterations and epileptogenesis has already been confirmed (Abdijadid et al., 2015). These deviations in neuronal cell body structure can influence the interconnection and communication between neurons, potentially influencing the onset and advancement of epilepsy (Hsieh et al., 2016; Wu et al., 2022). More precisely, these alterations in local and global connectivity can impact the manifestation of seizures, determining whether abnormal connectivity and hyperexcitability result in focal or generalized seizures (Sheybani et al., 2018; Represa, 2019).

In focal epilepsies, malformations of cortical development are associated as the primary substrate in which the presence of morphologically abnormal neurons significantly affects neural connectivity (Mainen and Sejnowski, 1996; Richards and Van Hooser, 2018). The existence of atypical neurons could influence the subsequent stages of development that regulate cortical synaptic connectivity (Subramanian et al., 2020). Avansini et al. (2022) observed an enhanced level of network connectivity (termed effective connectivity) along with increased neuronal excitability in human neural organoids derived from pluripotent stem cells of patients with FCD. The aberrant connectivity seen in FCD appears to be influenced by neuronal morphological abnormalities, particularly the presence of dysmorphic neurons. Using 3D confocal microscopy, the researchers detected enlarged cell bodies and increased dendritic complexity, potentially contributing to a more interconnected neural circuitry and the formation of an epileptogenic network in FCD.

Using high-resolution synchrotron x-ray microtomography and Golgi-Cox staining, Fonseca et al. noticed an altered distribution of neurons and a reduction of cell number in the hippocampus in a status epilepticus mouse model. These approaches allowed the assessment of the 3D cytoarchitecture, neuron density, and morphology (Fonseca et al., 2018).

The integration of dendrites and synapses into functional networks is heavily affected by how neocortical neurons are positioned during development (Martineau et al., 2018). Malpositioned neurons in the cortex cytoarchitecture are called heterotopic neurons (Ishii et al., 2015). These neurons alone may play a role but do not seem to be sufficient to trigger seizures (Aghakhani et al., 2005). The aberrant organization of cortical cytoarchitecture potentially leads to aberrant connections within these developing neuronal networks. Additionally, the recruitment of distinct microcircuits from different cortical locations could alter synchronicity, leading to abnormal neural oscillations (Dubeau et al., 1995; Abdijadid et al., 2015).

Neuronal disorganization and clusters of heterotopic neurons are primarily observed in human specimens from cortical migration malformations such as periventricular heterotopia (Ekşioğlu et al., 1996) and FCD type I (Coras et al., 2021) using light microscopy with immunohistochemical and DiI tracing techniques. Additionally, in animal epilepsy models (Mello et al., 1993), there have been observations of heterotopic granule cells in the dentate gyrus, resembling those found in human epilepsy. Heterotopic granule cells establish new connections and potentially impact synaptic reorganization (Babb, 1991).

The presence of dendritic spine pathologies and abnormal dendritic arborization have been suggested to be implicated in epilepsy worsening, increasing neuronal hyperexcitability in the circuits, and contributing to cognitive deficits, synaptic remodeling, and aberrant plasticity (Fiala et al., 2002).

Dendritic spines are mostly observed in excitatory synapses and neurons respond to epileptogenic changes in the circuitry by modifying the structure of their dendritic trees. Alterations in the distribution, quantity, and morphology of dendritic spines have been proposed to have a direct impact on seizures and epileptogenesis (Jiang et al., 1998; Jean et al., 2023). However, it remains unclear whether these changes are the cause or are a consequence of seizure recurrence (Wong and Guo, 2013). Dendritic pathology in epilepsy can be broadly categorized into two main fields, as described below:

Neuronal dendritic arborization: Morphological changes of dendrites can affect neuronal excitability. Abnormalities in dendritic length, shape, and branching patterns have been described in epilepsies associated with either hippocampal sclerosis, or tumors, or microdysgenesis (von Campe et al., 1997), and also associated with the presence of varicose swelling of the dendrites of granular dentate neurons of the hippocampus (Blümcke et al., 1999).

Dendritic spine pathology: The initial observation of dendritic spine loss occurred in hippocampal pyramidal neurons and dentate granule cells among individuals with TLE (Scheibel et al., 1974), providing a plausible mechanism to elucidate the learning and memory challenges experienced by these patients (Chen et al., 2010). In Lennox-Gastaut syndrome, a childhood epileptic disorder linked to intellectual disability, pyramidal neurons from brain biopsy were observed to possess a reduced number of spines using EM (Renier et al., 1988). In human cerebral cortices derived from FCD patients, a reduction of dendritic spines, and sporadic filopodia-like protrusions emerging from the soma in dysmorphic neurons were noticed using Golgi impregnation and confocal microscopy (Rossini et al., 2023).

Employing Golgi-Cox staining, optionally combined with immunohistochemistry, as well as DiI tracing, and utilizing both confocal microscopy and EM techniques, provides a comprehensive method for assessing the morphology and structure of dendritic arborization, as well as the density and morphology of neuronal dendritic spines in epilepsy.

Extensively imaging and tracing axons throughout the brain provides a mesoscale view of regional connectivity, offering insights into soma and neurite density as well as assessing total cell reductions and identifying cell death in specific brain regions (Bazinet et al., 2023). Although mesoscale imaging strategies can unveil a series of histological structures, they present their computational challenges.

The first challenge encountered in the mesoscale is usually the stitching of large quantities of high-resolution microscopy images. Image stitching refers to the process of aligning and overlaying two or more images of the same object taken from different, consecutive, and overlapping FOV. Through image registration processing (Sarvaiya et al., 2009), corresponding features or structures in these images are spatially aligned, making it possible to combine them into a single and panoramic image. In the case of brain microscopy, this involves merging multiple images, sometimes acquired at varying scales, into a comprehensive, high-resolution representation of the brain. Registering can mean either tile stitching multiple consecutive FOVs, without overlapping or positioning microscopic images into a larger dataset using fiducial markers, or a common reference frame to localize them into the brain. In both cases, it is expected that mesoscale imaging strategies generate the largest amount of raw data. BigStitcher is a method of stitching consecutive FOVs into a single high-resolution image (Hörl et al., 2019). To manage such large amounts of data, the BigSticher software computes shifts between overlapping image tiles by using a phase correlation method in downsampled images, which optimizes the computational time necessary for image stitching (Hörl et al., 2019). Moreover, DeepSlice is a CNN specifically trained on a substantial histological dataset to automatically align coronal mouse brain two-photon microscopy images with the Allen Common Coordinate Framework (CCF) (Carey et al., 2023).

The following step is to detect neuronal cell bodies in the entire image and quantify soma density across brain regions. The size of mesoscale data makes manual handling impractical and prone to biases; hence, automatic or semi-automatic tools are more suitable for its processing (Bjerke et al., 2023). Soma detection can be made either by traditional image enhancement filters followed by intensity thresholds, such as in ClearMap (Renier et al., 2016) and MIRACL (Goubran et al., 2019), or by advanced machine learning techniques for pixel classification as deep learning approaches (Tyson and Margrie, 2022). Intensity thresholding approaches also work well with DAPI-stained nuclei images (Kim et al., 2015), and modifications of thresholding can be done to address large-scale GFP images even with a large variability in contrast (Frasconi et al., 2014). These modifications consist of first using mean shift clustering to detect soma centers followed by image deconvolution and finally manifold learning for filtering false positives (Frasconi et al., 2014). However, intensity thresholding and morphological approaches fail especially with densely packed images and that is precisely where deep learning can be used. For example, (Hu et al., 2021) combined 3D U-shaped full CNN with multi-task learning to perform soma segmentation in Nissl stained images. This strategy is done in small patches and would take a long time to train in teravoxel mesoscale images. As a faster approach, (Wei et al., 2023) used a lightweight MCC-Net to reduce computational complexity in soma detection. Then, in the second stage, they employed SFS-Net for precise soma localization in mouse brain images, utilizing advanced segmentation techniques. Experimental results confirmed the excellent performance of the method and its capacity to provide valuable information for neuron reconstruction (Wei et al., 2023). The user-friendly software CellPose (Stringer et al., 2021) also has a DeepLearning module that has been used to count pyramidal neurons in histopathological images (Oltmer et al., 2023). An alternative option is to employ Suite2p software (Pachitariu et al., 2017), which offers AutoROI cell segmentation designed for simultaneous analysis of functional and morphological two-photon calcium images. The compilation of the main soma quantification methods for mesoscale connectivity is presented in Table 2.

The next step to a mesoscale connectivity view is to detect, trace, and quantify neurites across the brain. In the study conducted by Allen Mouse Brain Connectivity Atlas (Oh et al., 2014), axonal detection at the mesoscale level starts with a very similar process to single neuron morphology mesh tracing. Signal detection algorithms, such as filament tracing, can be used as an initial segmentation. The authors then rescale image intensity and remove noise using filters and morphological techniques. Candidate signal objects were identified based on adaptive edge/line detection and morphological attributes such as length and area. Additionally, high-intensity pixels near these objects were considered. In a post-segmentation step, objects considered artifacts were removed. It is important to note that passing fibers and terminals were not differentiated. The outcome is a high-resolution mask classifying each 0.35 μm × 0.35 μm pixel as a signal or background (Oh et al., 2014).

Also in the mesoscale, TRAILMAP uses a modification of a 3D UNet to extract axonal projections from uncleared brain tissue in light-sheet microscopy (Friedmann et al., 2020). This method focuses on segmenting axons from the background in a generalized way that can be applied to all brain regions. Unlike filament tracing methods, it does not address neurite branching numbers or spatial positions. The purpose is mainly to address axonal fiber density and compare it across brain regions (Tyson and Margrie, 2022).

Another possibility is to use the MIRACL pipeline and toolbox. MIRACL is based on a multimodal approach that integrates CLARITY data at the microscopic level with macroscopic in vivo and ex vivo imaging data, including structural, diffusion, and quantitative MRI, all aligned to the Allen atlas reference frame “ARA.” This integration facilitates various analyses, including the examination of histological features across network graphs and fiber tracts, as well as connectivity analyses based on projection terminals. Additionally, MIRACL supports group-level statistics, multimodal correlations, and comparisons of connectivity maps across different scales (Goubran et al., 2019).

The morphology of a neuron can have a big impact on its connectivity with other local neuronal circuits. Neurons with a complex dendritic branching pattern tend to have a larger surface area and a denser synaptic field, allowing them to have more candidate synapses (van Pelt and van Ooyen, 2013). According to Peter’s rule, the colocalization of dendritic and axonal arbors are reasonable predictors of connectivity among neuron types (Rees et al., 2017). Nevertheless, a greater number of potential synapses does not always mean a greater number of functional synapses (Rees et al., 2017). Axo-dendritic overlapping is a necessary but not sufficient condition to ensure a synaptic connection. Light microscopy is the ideal imaging technique to study both neuronal morphology (using cytoplasmic markers) and synaptic connectivity (using puncta colocalization) (Wang et al., 2020). In contrast, neuron morphology has recently been described to predict non-random connectivity in local networks and circuits (Udvary et al., 2022). The authors state that the specificity in neural wiring is influenced by morphological factors such as similarities in neurite projections, packing density, and the diversity of cell types in the neuropil (Udvary et al., 2022). High values in these factors lead to recurring patterns in the network, while lower values result in a more feedforward network structure (Udvary et al., 2022).

To address single neuron morphology using imaging techniques we must first extract from the image which pixels belong to each neuron. This process is called instance segmentation or neuron instance individualization (You et al., 2019). The main techniques used to measure multiple neurons are light microscopy (including confocal, two-photon, STED, and light-sheet) and serial-section EM.

The first and most important parameter to quantify single neuron morphology is the radial profile of neuron dendrite spanning tree, also known as the Sholl Intersection Profile (SIP) (Bird and Cuntz, 2019). The complete Sholl analysis includes measuring the total length of the dendrite, the axon domain maximum and minimum from the soma, and the angular distribution of dendritic segments that deviate from a direct path to the soma. According to the authors (Bird and Cuntz, 2019), a larger dendrite extension length implies a larger region where synapses can occur, peaks in the SIPs are related to regions where synapses have a higher probability to occur and valleys in the SIPs are regions to where synapses have a lower probability to occur. The angular distribution is related to a neuron’s centripetal bias and implies a neuron that minimizes wiring to ensure an efficient propagation of electrical impulses.

Alternative ways to measure neuronal shape include parameters such as the total length of neurites, the minimal occupied volume, the distribution of branch lengths as represented in histograms, and the frequency of distances between successive bifurcations along the neural trajectory. These measurements are obtained through the open-source software L-measure, as outlined in the work of Scorcioni et al. (2008).

A further challenge is to classify neuron types using only their morphological assets without any molecular markers (Polavaram et al., 2014) used L-measure to extract morphological features of neurons in the NeuroMorpho database. They subsequently applied principal component analysis (PCA) as a statistical tool to identify key morphological parameters capable of effectively classifying dendritic structures across diverse metadata categories. Their findings highlight the importance of specific measures like branching density, size, tortuosity, bifurcation angles, arbor flatness, and topological asymmetry in capturing meaningful features of dendritic trees. Similarly, Khalil et al. (2021) extracted L-measure metrics and modified Sholl descriptors from the NeuroMorpho database and used PCA and KNN clustering to classify neuronal types.

Deep learning revolutionized feature extraction and image classification and has been used to classify neurons. For example, GraphDINO used a Transformer-based Graph Neural Network to create 3D spatial embedding representations of neuronal graphs and later classified them into neuronal types (Weis et al., 2021). The authors adapted positional encoding and introduced a novel attention mechanism called AC-Attention to fit neuronal graphs and achieved results comparable to expert-manual classification without prior knowledge about neuronal structural features and outperforms previous methods in predicting expert labels on quantitative benchmarks (Weis et al., 2021). Similarly, MorphoGNN is a novel approach for embedding single neuron morphologies using graph neural networks (GNN) and learns spatial relationships between nodes in reconstructed neuron fibers by considering their nearest neighbors on each layer. This process generates a reduced-dimensional representation of individual neurons using an end-to-end model that incorporates densely connected Densely Connected Convolutional layers and a dual pooling operator (Zhu et al., 2023).