Introduction

In the realm of clinical and biomedical research, the advent of sophisticated 3D imaging methodologies heralds a profound evolution, offering unprecedented insights into the complexity of human brain architecture and pathology. Standard histopathological analysis, mainly performed on Formalin-Fixed Paraffin-Embedded (FFPE) tissue, has long been the cornerstone of clinical research and diagnostics1,2. However, while providing valuable insights into tissue morphology and composition, histological techniques are inherently limited to two-dimensional (2D) representations, leading to a loss of spatial context and a failure to capture the intricate complexities of biological structures. Conversely, volumetric imaging analysis offers a comprehensive understanding of the spatial organization of human brain areas, enabling researchers to capture the full complexity of biological structures and correlate it to their function3. Histopathological analyses are routinely performed on ultra-thin slices (<10-μm-thick) of FFPE specimens. As a result, only a fraction of the FFPE tissue block is utilized, while the remaining is often archived and stored for extended periods. These archives of preserved clinical FFPE specimens represent a significant, yet underutilized resource for advanced volumetric analysis.

Recently, the combination of high-resolution 3D imaging with advanced 3D bioimage analysis has emerged as a powerful toolkit in neuroscience3,4,5,6. Thanks to the spread of automated image segmentation and classification methods, fluorescence imaging techniques, such as light-sheet fluorescence microscopy (LSFM) or two-photon fluorescence microscopy (TPFM), are evolving from purely observational to valuable quantitative analytical tools. Volumetric image reconstruction of human brain areas requires tissue optical transparency to reduce light scattering and absorption, thereby increasing the light penetration depth by refractive index matching7,8,9,10,11. Unlike for model organisms, clearing and labeling human brain tissues represents a challenging task. The major limitations are due to the variability in fixation and storage conditions of human samples, the high autofluorescence signal of blood vessels and lipofuscin, and the significant heterogeneity in cellular composition, structure, and optical properties across different regions and individuals12,13,14,15,16,17,18. To overcome some of these limitations, clearing methods such as CLARITY, SWITCH, SHIELD, iDISCO, and CUBIC, which were primarily developed for model organisms, have been adapted for human tissues18,19,20,21,22,23,24. Among the clearing methods specifically developed for large-scale human brain analysis, the SHORT tissue transformation method15,25,26 enables rapid and efficient clearing of thick brain slices. SHORT results in high tissue transparency, structure preservation, reduction of intrinsic autofluorescence, and uniform sample co-staining by means of several cellular markers.

To harness the full potential of clearing techniques on human FFPE samples, an essential step in the workflow involves the full removal of the paraffin matrix from tissue blocks. Harsh deparaffinization methods are routinely used for proteomics or genomics analysis; however, these protocols require a homogenization step either for proteins, DNA, or RNA extraction which does not preserve the architecture of the sample27,28,29,30. The development of volumetric reconstruction methods prompted researchers to optimize protocols of deparaffinization that maintain the structural organization of the sample, but the few alternatives proposed so far either for mouse or human tissues have been applied only to lungs, lymph nodes, and tumors31,32,33. The limited application of the deparaffinization step in combination with sample clearing and labeling methods highlights the difficulty in establishing a reliable protocol for paraffin removal from human brain tissues.

The ideal method for deparaffinization must strike a delicate balance, effectively removing the paraffin that could interfere with subsequent clearing steps, while preserving the integrity of tissue and protein structures. This is particularly difficult with human brain tissue since epitope preservation is essential for a homogeneous and specific labeling of the structures of interest, ensuring accurate visualization and analysis in downstream volumetric imaging.

Here, we present a mild deparaffinization method that preserves the molecular and structural tissue architecture and works on blocks of human brain tissue of different sizes. This method was applied to adult postmortem human brain areas (brainstem), to postsurgical brain specimens removed from pediatric patients affected by malformations of cortical development (MCD) and to adult patients affected by hippocampal sclerosis (HS) and cryptogenic epilepsy. Deparaffinized samples were processed with different clearing methods and reconstructed either with custom-made LSFM or TPFM setups.

Finally, to extract quantitative information on the spatial distribution and morphology of different classes of neurons from the 3D reconstructions of cleared samples, we implemented a machine-learning-based image segmentation workflow. This was employed to investigate the structural differences between two pathological samples of MCD, both presenting histologically focal cortical dysplasia type IIa (FCDIIa), but characterized by different genetic variations, and presenting in one case hemimegalencephaly (HME). The cytoarchitectural organization of the samples was characterized by analyzing the neuronal volume, ellipticity, local density, and clustering level of two different neuronal populations.

Overall, we present a comprehensive approach for performing 3D imaging and automated quantitative cellular analysis on FFPE human brain specimens. This workflow includes human brain tissue deparaffinization, optical clearing, multi-labeling, LSFM or TPFM imaging, automated neuronal segmentation, and quantitative cytoarchitectural analysis (Fig. 1).

Fig. 1: Overview of the entire pipeline for deparaffinization, clearing, labeling, imaging, automated neuronal segmentation and analysis.
Fig. 1: Overview of the entire pipeline for deparaffinization, clearing, labeling, imaging, automated neuronal segmentation and analysis.
Full size image

Schematic representation of the entire pipeline. a Deparaffinization of FFPE (Formalin-Fixed Paraffin-Embedded) adult and pediatric human brain tissues. b Clearing and labeling of deparaffinized slabs with the SHORT tissue transformation method, followed by volumetric imaging with LSFM (Light Sheet Fluorescence Microscopy) and TPFM (Two Photon Fluorescent Microscopy) custom-made setup. c Block diagram of the 3D image processing pipeline for quantitative cytoarchitectural analysis of TPFM images. Adjacent overlapping TPFM image stacks first undergo flat-field correction using spatial gain models estimated with the CIDRE retrospective method. ZetaStitcher is then used to align the corrected stacks in order to create high-resolution image reconstructions of the brain specimens. Supervised pixel and object random forest classifiers, trained using the ilastik interactive machine learning tool, assessed these reconstructions in sequence to automatically identify pRPS6+ and pRPS6- neuronal bodies. Finally, quantitative structural and morphological features are evaluated. The figure is partially created in BioRender. https://BioRender.com/7poqasl License agreement n. VE28MA1GGY.

This approach grants access to the wide range of paraffin-embedded clinical specimens, enabling 3D quantitative analysis with cellular resolution and, thus, providing unprecedented insights into human brain cytoarchitecture in both physiological and pathological tissue samples.

Results

Deparaffinization and clearing of FFPE human brain tissue blocks

FFPE tissues are the most abundant type of samples available in neuropathology laboratories. To perform tissue clearing and 3D volumetric imaging, we optimized a method to completely remove paraffin from human brain tissue blocks. We successfully applied our protocol to multiple samples derived from both healthy individuals and patients with focal MCDs, dysplastic megalencephaly (DMEG), cryptogenic epilepsy and hippocampal sclerosis (HS), both pediatric and adult. Moreover, the brain samples used differed in storage conditions, and donor age, thus highlighting the versatility of our pipeline (Table S1). In particular, we processed a post-mortem sample (brainstem) as well as samples surgically removed from an adult patient with HS and cryptogenic epilepsy, and pediatric patients with FCDIIa, FCDIIb (upon surgical treatment of drug-resistant epilepsy), and DMEG (post-mortem) (Figs. S1, S2a, S2c, S2e). Deparaffinization was achieved by melting the paraffin wax surrounding the tissue block in a water bath at 60 °C followed by several washes in xylene until the tissue block appeared completely translucent. These steps ensure the full removal of paraffin, both externally and deep inside the tissue. Following deparaffinization, xylene was removed with 100% ethanol, and the samples were rehydrated with a series of graded ethanol, until PBS was used for washing.

From the deparaffinized FFPE blocks, 500 μm-thick slices were cut and used to evaluate the compatibility of the deparaffinization protocol with two different clearing methods: the SHORT tissue transformation method15,26 and the iDISCO protocol22. To reconstruct the 3D structure of the cleared and stained volume, samples were imaged with a custom-made dual-view inverted LSFM setup25,34, with an isotropic voxel size of 3.64 μm following post-processing.

In order to demonstrate the compatibility of the methodology with the SHORT technique and with multiple staining (e.g., targeting various neuronal subpopulations), we tested several markers (Table S2). The acquired LSFM volumetric reconstructions showed uniform and specific co-staining throughout the tissue depth (Figs. 2, and S2). The slice from the brainstem was labeled with two markers to characterize different subpopulations of GABAergic interneurons: Somatostatin (SST) and Calretinin (CR) (Fig. 2a–c). Whereas the slabs from the two surgical hippocampi were labeled with the NeuN and the nuclear marker DAPI in the HS-affected hippocampus (Fig. 2d–f), and the tyrosine hydroxylase antibody along with the nuclear marker YOYO-1 in the cryptogenic hippocampus (Figs. S2e, S2f). The 3D reconstructions obtained with LSFM revealed a specific staining of neurons across various hippocampal regions (Figs. 2e, f, and S2f). As expected, in the volumetric reconstruction of the hippocampus derived from a HS-affected patient we observed segmental pyramidal cell loss in several Cornu Ammonis (CA) sectors (Fig. S3a, S3b), a typical histopathologic hallmark of HS35, as confirmed in FFPE tissue sections from the same block. Immunohistochemistry and histological staining on adjacent paraffin blocks confirmed this finding (Fig. S3c, S3d, S3e, S3f). Finally, the FCD and DMEG brain slabs were marked with the YOYO-1 dye, the pan-neuronal marker NeuN and the phospho-ribosomal protein S6 (pRPS6Ser235/236) (Figs. 2g–i, and S2b, S2d). The latter was used as a readout of the phosphatidylinositol 3-kinase (PI3K)/AKT/mammalian target of the rapamycin (mTOR) pathway dysregulation, which characterizes FCDII36.

Fig. 2: LSFM 3D reconstructions with deparaffinized human brain slabs processed with the SHORT method.
Fig. 2: LSFM 3D reconstructions with deparaffinized human brain slabs processed with the SHORT method.
Full size image

a Image of a postmortem adult human brainstem, before deparaffinization (FFPE) and after SHORT (TDE 68%). b Maximum intensity projection image (3.64 µm isotropic resolution) and volumetric rendering showing a mesoscopic reconstruction of a labeled for Somatostatin (SST, magenta) and Calretinin (CR, cyan). Scale bar: 1 mm. c Inset images of b showing the single markers used, SST (magenta) and CR (cyan). Scale bar: 100 µm. d Image of a postsurgical human hippocampus from a patient with hippocampal sclerosis (HS), before deparaffinization (FFPE) and after SHORT (TDE 68%). e Maximum intensity projection image and volumetric rendering showing a mesoscopic reconstruction of d labeled for NeuN (magenta) and DAPI (blue). Scale bar: 2 mm. f Inset images of e showing the single markers used, NeuN (magenta) and DAPI (blue). Scale bar: 100 µm. g Image of a postsurgical brain specimen from a pediatric patient with focal cortical dysplasia type IIa (FCDIIa), before deparaffinization (FFPE) and after SHORT (TDE 68%). h Maximum intensity projection image and volumetric rendering showing a mesoscopic reconstruction of g labeled for NeuN (magenta) and pRPS6 (cyan). Scale bar: 1 mm. i Inset images of h showing the single markers used, NeuN (magenta) and pRPS6 (cyan). Scale bar: 100 µm. The inset images in (c), (f), and (i) refer to the regions in white boxes in (b), (e) and (h), respectively.

To assess the compatibility of the deparaffinization protocol with different clearing methods, we also employed it in combination with the organic solvent-based technique iDISCO22. Using this technique, we achieved high levels of tissue transparency and multi-staining with selected markers (Table S2) in different human brain specimens: a postmortem adult brainstem (Fig. S4a, S4b), a postsurgical hippocampus (HS-affected patient) (Fig. S4c, S4d), and a pediatric FCD-affected cortical brain sample (Fig. S4e, S4f). However, high levels of natural autofluorescence from lipofuscin-type pigments were observed in adult samples when imaging with excitation wavelengths at 488 and 561 nm. We therefore demonstrated the compatibility of our deparaffinization approach with two distinct clearing methods: the SHORT tissue transformation technique and the organic solvent-based iDISCO. This entire process was applied to adult and pediatric specimens from various brain regions, enabling 3D reconstruction of the samples acquired with LSFM at micrometer resolution.

Quantitative cytoarchitectural analysis on deparaffinized pathological human brain samples

The successful deparaffinization, clearing, and co-labeling of pathological human brain samples prompted us to evaluate and compare quantitative cytoarchitectural features, namely local neuronal density and clustering—a structural parameter shown to have a marked functional connotation37,38—along with morphological cell properties, such as soma volume and ellipticity. To this aim, mesoscopic image reconstructions of 500 μm-thick sections with a resolution of 1.21 µm × 1.21 µm × 2 µm were acquired using a TPFM setup (Fig. 3). Two brain specimens surgically removed from pediatric patients were examined: two samples from patients with FCDIIa (S1 and S2), which in one case was hemispheric and associated with HME (S1) (Fig. 3a, b). These two patients underwent genetic testing. In patient S1, massive parallel sequencing of a panel containing 48 genes of the mTOR pathway revealed the heterozygous c.392C > T (p.Thr131Ile) variant of uncertain significance (VUS) in the PTEN gene (NM_000314.4), which was inherited from the healthy mother. Patient S2 corresponds to Patient 1 described by Guerrini and collaborators39 and carried the c.6644C > T (p.Ser2215Phe) MTOR (NM_004958.3) pathogenic variant, which was present at 5.5% alternative allele fraction (GS Junior sequencing validation: 2.46%) in the dysplastic brain tissue and absent in DNA extracted from saliva and blood.

Fig. 3: Volumetric imaging with TPFM and analysis on surgically removed pediatric brain specimens.
Fig. 3: Volumetric imaging with TPFM and analysis on surgically removed pediatric brain specimens.
Full size image

Representative middle plane of mesoscopic reconstructions obtained with TPFM (1.2 μm ×  1.2 μm × 2 μm resolution) are shown for surgically removed brain pieces from the two patients with FCD IIa carrying the VUS in PTEN (S1, a) and the pathogenic variant in MTOR (S2, b). Tissues were labeled for NeuN (magenta) and pRPS6 (cyan) while blood vessels were detected through autofluorescence (yellow). Scale bar: 200 µm. Magnified inset images show the individual marker used. Scale bar: 50 µm. Inset images correspond to the regions marked in white boxes. c, d Corresponding 3D semantic segmentations generated using ilastik’s pixel classification workflow (headless prediction on a distributed computering cluster). Yellow: pRPS6- neurons; cyan: pRPS6+ neurons; magenta: blood vessels. Background suppression demonstrates striking accuracy. e, f Local cell density maps generated from the TPFM reconstructions in a and b.

Specific cytoarchitectural alterations in the population of neurons expressing the phospho-Ribosomal Protein S6Ser235/236 (pRPS6+) were quantitatively assessed with respect to neurons stained with anti-NeuN antibody through machine-learning-based image enhancement and automatic cell segmentation (Fig. 3c, d). In particular, semantic segmentation of TPFM image reconstructions, automatic recognition and quantitative characterization of neuronal bodies were carried out using ilastik4, a user-friendly, open-source software for bioimage analysis. Leveraging the power of supervised machine learning, ilastik simplifies the often complex and labor-intensive task of image segmentation by allowing users to interactively train random forest image classifiers through a highly intuitive interface. Despite being considerably slower than state-of-the-art deep learning-based tools for automatic cell centroid localization, such as the GPU-implemented BCFind-v240, ilastik enables the faithful semantic segmentation of neuronal bodies and, thus, the quantitative evaluation of their volume and shape. In this work, automatic neuron detection was achieved by means of two separate machine learning workflows supported by ilastik, i.e., pixel and object classification. Following the interactive manual training, both ilastik’s workflows were run in headless mode on a high-performance computing cluster comprising four nodes.

An average out-of-bag error41 of 2.8% was achieved on the training set, which was generated by three expert operators involved in the manual image annotation task. Additionally, median F1 scores of 0.78 (IQR = 0.15) and 0.78 (IQR = 0.21) were obtained for the pRPS6- and pRPS6+ cell body classes, respectively, from 30 image Regions Of Interest (ROIs) randomly selected for validation (15 ROIs of 150 µm x 150 µm from each subject, Fig. S5; see Table S3 in Supplementary Information for full ROI-wise performance metrics). These results reflect the random forest classifier’s ability to perform an accurate segmentation of both target neuronal subpopulations in their respective class probability channels, as well as to reliably reconstruct local 3D microvascular structures, thanks to effective suppression of background tissue signal.

The subsequent object classification stage allowed us to refine the raw cell detections, improving the accuracy of the original binarized class probabilities, through the suppression of false cell instances and the robust recognition of multiple tightly packed cells that had been unavoidably merged during the binarization of the semantic maps returned by the upstream image enhancement workflow. This was crucial for preventing local underestimates of the actual cellular density and, thus, for the correct representation of spatial and inter-subject variations in the brain tissue cytoarchitecture. To visualize the cluster of pRPS6+ neurons in the S1 specimen and the remarkably higher prevalence of pRPS6-expressing cells observed in the S2 specimen, we generated neuronal density maps for both the subjects (Fig. 3e, f).

The method detected pRPS6-expressing neurons in both tissue samples but the percentage of pRPS6+ neurons relative to the total number of neurons, identified by the NeuN staining, resulted remarkably higher in S2 (35.4%) compared to S1 (3.1%) (Table S4). This difference is in line with the finding that genetic testing revealed only a probably benign VUS in S1, while S2 was found to carry a pathogenic variant in MTOR reported to cause strong hyperactivation of the mTOR pathway39. Interestingly, in S2 we also observed a statistically significant increase in the neuronal volume (Bonferroni-corrected p < 0.001; Hellinger distance, H = 0.308, inter-median distance, |ΔM| = 2884.3 μm³; Fig. 4a and b) and ellipticity (p < 0.001, H = 0.282, ΔM = 0.4; Fig. 4c and d) of pRPS6+ neurons, compared to S1. This morphological alteration pattern resembles the one already highlighted by neuropathological evaluations of FCD brain tissue, which revealed the presence of dysmorphic giant neurons exhibiting RPS6 hyperphosphorylation39. S2 also showed a slight increase in the cell body volume and ellipticity of neurons not expressing the pRPS6 protein (pRPS6-), compared to S1 (p < 0.001, H = 0.154, |ΔM| = 579.7 μm³; p < 0.001, H = 0.122, |ΔM| = 0.1; Fig. 4a–d). This finding may be ascribed to the differences in cell types and distribution in different brain areas, as tissues used in this study have been sampled from the parietal lobe for S1 and the frontal lobe for S2. We also detected statistically significant differences between S1 and S2 regarding their local neuronal density (within a radius of 150 μm). In detail, compared to S1, S2 exhibited a lower cell density of neurons not expressing pRPS6 (p < 0.001, H = 0.606, |ΔM| = 11 cells / 10⁶ µm³; Fig. 4e and f) and a higher density of pRPS6+ cells (p < 0.001, H = 0.612, |ΔM| = 4 cells/10⁶ µm³; Fig. 4e and f). Thus, the pronounced increase in the percentage of pRPS6-expressing neurons reported above for subject S2. Differences in cell distribution may also explain the modest but statistically significant difference between S1 and S2 that emerged for the clustering indices of pRPS6- neurons (p < 0.001, H = 0.192, |ΔM| = 0.05 a.u.; Fig. 4g and h). On the other hand, it is more relevant to emphasize the peculiar distribution of the clustering indices of pRPS6+ neurons in subject S1 compared to that of S2 (Fig. 4g and h). Indeed, differently from the narrow data distributions observed in S2 (indicating, to a greater or lesser extent, a uniform spatial arrangement across the acquired TPFM image reconstructions), in S1 the pRPS6+ cell clustering indices show a wide distribution spreading from negative data with a marked peak at the lowest value of the distribution, specifically associated with the presence of completely isolated cells, to high values corresponding to local clusters of neurons. This manifests in statistically significant differences with respect to the pRPS6+ clustering data distribution of subject S2 (p < 0.01, H = 0.332, |ΔM| = 0.05 a.u.; Fig. 4h) and to the reference clustering indices of neurons pRPS6- evaluated for that subject (p < 0.01, H = 0.453, |ΔM| = 0.07 a.u.; Fig. 4h). All statistical results are reported in Table S5.

Fig. 4: Quantitative observed cytoarchitectural analysis on pediatric human brain specimens.
Fig. 4: Quantitative observed cytoarchitectural analysis on pediatric human brain specimens.
Full size image

Quantitative cytoarchitectural analysis of neuronal body morphology (a, b: cell volume; c, d: cell ellipticity) and spatial organization (e, f: local cell density; g, h: clustering index). PD probability density; ***p < 0.00017, **p < 0.00167, Mann–Whitney U-Test (Bonferroni correction); H: Hellinger distance between the compared data distributions.

Discussion and conclusions

Volumetric reconstruction with sub-cellular resolution can provide detailed insights into the complexity of human brain architecture and pathology. In this work, we introduce a comprehensive workflow to perform 3D reconstruction and automated quantitative cellular analysis on deparaffinized human brain samples. We developed a versatile deparaffinization protocol for human brain tissue blocks, compatible with different optical clearing and multi-labeling methods. Volumetric reconstructions were analyzed to extract quantitative information about the morphology and spatial distribution of different classes of neurons, through a machine-learning-based image segmentation workflow (Fig. 1).

Optical clearing methods are widely used on fresh or archival formalin-fixed tissues. However, a substantial fraction of human samples stored in biobanks are embedded in a paraffin matrix to preserve tissue integrity. Efficient clearing and 3D imaging require paraffin removal, which otherwise acts as a physical barrier, impeding the diffusion of clearing agents, thus leading to uneven optical transparency that compromises the image quality. Residual paraffin can also hinder the accurate labeling of structures of interest. Harsh deparaffinization methods, such as prolonged exposure to organic solvents or high temperatures, can induce epitope masking effects and/or antigen damage, affecting the accuracy of subsequent analyses.

The mild deparaffinization method proposed here allows for the complete removal of paraffin from human brain tissue blocks by combining a water bath’s wet heat to eliminate outer paraffin and multiple xylene incubations for inner paraffin removal. Unlike the approaches presented for murine models, processing human brain tissue required increasing the frequency and reducing the duration of individual xylene washes, performed using a rotatory shaker. This approach ensured the complete removal of inner paraffin while preserving the tissue’s macrostructure and antigenicity, as demonstrated by subsequent staining.

There are significant differences in tissue architecture and optical properties between adult and pediatric brain samples, as well as between tissues from various regions of the human brain42,43,44,45,46. Additionally, factors such as the type of fixative (formalin or paraformaldehyde (PFA)), the duration of fixation, and pre- and postmortem changes, such as pH fluctuations and tissue oxygenation, can vary significantly from one brain to another. Further variability arises from donor age and the collection modality, whether the tissue is obtained postmortem or during surgery, both of which can influence tissue integrity and molecular preservation. Although steps like paraffin melting time and the number of xylene washes require optimization on a case-by-case basis (Table S1), our deparaffinization approach effectively enabled paraffin removal from a wide range of specimens, demonstrating the protocol’s versatility. This includes normal and abnormal patient-derived human brain samples from both adult and pediatric cases and obtained through postmortem examination or surgical resection. Two methodologies were tested for clearing and labeling: the SHORT (SWITCH-H2O2-antigen  Retrieval-TDE) tissue transformation technique15,26 and the organic solvent-based iDISCO22. Both methods achieved high tissue transparency and specific, homogeneous co-staining with multiple antibodies, allowing us to visualize different neuronal populations and characterize pathological samples. However, the SHORT technique was more effective at quenching lipofuscin autofluorescence and proved more efficient for labeling pediatric samples with respect to iDISCO. Therefore, the quantitative evaluation of neuronal subpopulations was performed on SHORT-cleared slabs.

Neuronal and glial cell numbers are important attributes in the characterization of distinct functional regions of the nervous system. These numbers are, moreover, susceptible to pathologies, pharmacological treatments, and genetic alterations47,48,49,50. The vast majority of reports on cell counts in the neuroscience literature are based on stereological methods51,52,53. In these approaches, the number of cells is accurately measured in a small but unbiased portion of the volume of interest, and the numbers for the whole target region can then be extrapolated under the assumption that the sample is representative. Since stereological neuron counting requires the involvement of a human operator to identify cells in the stained tissue54, the procedure is inherently laborious, time-consuming, and rather inefficient, particularly for cell-sparse regions55.

However, in recent years, image analysis has been revolutionized by machine learning (ML). Thanks to the growing adoption of ML-based image segmentation, 3D fluorescence imaging techniques, such as LSFM and TPFM, are nowadays evolving from being just a qualitative observational modality to a valuable analytical tool that enables the automatic extraction of quantitative cytoarchitectural information from 3D mesoscopic reconstructions of tissue samples imaged with sub-cellular resolution. This was so far unfeasible for human operators to perform with adequate reliability and within practical time frames. The manual labeling effort required to robustly train a supervised image classification algorithm with sufficiently rich image feature data may, however, still be considerable, if not impossible. In this regard, the adoption of the ilastik toolkit4 for interactive ML-based image analysis has proved of the utmost importance, providing us with a convenient user interface supporting fast interactive training with immediate feedback on the achieved segmentation accuracy. Indeed, trained pixel and cell body classifiers were ready within a few working days. This capability to swiftly deploy and execute ML-based bioimage analysis workflows, as shown in the present work, will be crucial to building meaningful large-scale human brain cytoarchitectural datasets for future neuroanatomical studies.

Among the different fluorescence microscopy technologies, LSFM has the advantage of enabling rapid reconstruction of large, centimeter-sized samples, with spatial resolution in the micrometer range56,57. However, this level of resolution is not always sufficient for accurately assessing neuronal morphology and dimensions. Moreover, LSFM is susceptible to stripe artifacts, which arise from residual opaque structures that remain in the sample despite high-quality optical clearing. These artifacts can significantly compromise fluorescence uniformity in the image and, in more severe cases, obscure critical features of interest, potentially leading to quantification issues58,59. In contrast, TPFM is a point-scanning technique that does not suffer from shadowing artifacts and enables high-resolution reconstruction, down to the nanometer scale60. However, the acquisition time is significantly longer compared with the LSFM setup. Therefore, the choice of the imaging setup should be determined by the specific goals of the biological analysis.

Nowadays, tissue clearing is often combined with expansion microscopy to overcome the diffraction limits of conventional setup and achieve super-resolution volumetric imaging of nanoscale structures, such as synapses and organelles, within large, volumetric samples. In this study, the superior resolution of the TPFM setup prompted the decision to perform the quantitative evaluations on mesoscopic reconstructions obtained with this system from two surgically removed pediatric brain specimens, previously imaged by LSFM. The genetic testing on these specimens revealed a likely benign VUS in PTEN in S1 and a known pathogenic variant in MTOR which hyperactivates the mTOR pathway in S2. For this reason, we selected the cell populations to analyze based on the expression of the pRPS6, a marker of mTOR signaling activation, overexpressed in most FCDII36. We first evaluated the percentage of pRPS6+ neurons relative to the total number of neurons identified by the NeuN staining and found it to be significantly higher in the sample carrying the pathogenic variant in MTOR (S2, ~35%) with respect to that carrying the likely benign VUS in PTEN (S1, ~3%). The low percentage of pRPS6+ neurons observed in the S1tissue could be ascribed to the fact that the likely benign VUS in PTEN identified in Patient S1 does not cause mTOR pathway hyperactivation or that its effect is not highlighted by RPS6 phosphorylation. The downstream signaling of PTEN can be mediated by two complexes, mTORC1 and mTORC2, which have different downstream targets. Considering that RPS6 is a downstream effector of the mTORC1 complex, we cannot rule out that the PTEN variant in Patient S1 if indeed pathogenic, exerts its effect independently of pRPS6, by preferentially activating the mTORC2 complex. Selective deregulation of mTORC2 caused by mutant PTEN has been previously documented in a model of human glioblastoma cell lines61 and in PTEN-deficient mice62. Our analysis has also been able to detect morphological alterations and, in particular, a statistically significant increase in the neuronal volume and ellipticity of pRPS6+ neurons in S2 with respect to S1, corroborating a well-known characteristic of FCD brain tissue that exhibits a high presence of dysmorphic giant neurons where RPS6 is hyperphosphorylated. Additionally, we observed differences in cell clustering indices: in S1, pRPS6+ cells displayed a broad distribution, including both isolated neurons and local clusters, whereas in S2, pRPS6+ cells exhibited a uniform, non-clustered distribution. Future studies are needed to further investigate the possible correlation of these pRPS6+ cells clusters with epileptogenic foci, as well as the variance of cell density distributions in MCDs. Nevertheless, the protocol we set up provides new insights into clarifying pathophysiological alterations of brain anomalies, such as focal MCDs. In fact, most data used to define histopathological alterations of FCD samples come from 2D images, but such images do not allow for the reconstruction of all brain networks altered in these conditions. Cells carrying pathogenic variants may establish connections with others, either mutants or WT, located on different focal planes. Thus, tracing the connections of mutant cells along different focal planes will improve our knowledge of how they interact with the surrounding tissue. The protocol we designed also has the potential to disclose new morphological cell markers that may impact the diagnosis and classification of focal MCDs.

In conclusion, our workflow offers valuable tools for investigating human brain tissues, highlighting the untapped potential of FFPE tissue archives for advanced volumetric analysis. This could help in clarifying pathophysiological alterations and uncover new morphological markers, potentially impacting the diagnosis and classification of neurological diseases.

Materials and methods

Human samples ethics approval and consent to participate

  1. 1.

    Pediatric human brain samples: Informed consents were obtained for all the samples involved in this study. The study was approved by the board: Pediatric Ethics Committee of the Tuscany Region (Italy) for Meyer General Hospital in the context of the DESIRE FP7 EU project and its extension by the DECODE-EE and Human Brain Optical Mapping projects: HBOM-Meyer, approval number 283/2021. Pediatric human brain samples were removed during surgical procedures for the treatment of drug-resistant epilepsy in children with focal MCDs or during autoptic procedures. Samples were obtained after informed consent, according to the guidelines of the Pediatric Research Ethics Committee of the Tuscany Region. Upon collection, samples were placed in neutral buffered formalin (pH 7.2–7.4) (Diapath, Martinengo, Italy) and stored at 4 °C until the deparaffinization and clearing process.

  2. 2.

    Healthy human brainstem: tissue was obtained from the Body Donation Program “Donation to Science” at the University of Padova. Before death, participants provided their written consent for the use of their entire body for any educational or research purpose in which the anatomy laboratory is involved, as required by the Italian Law 10-2-2020 n. 10 and to the Declaration of Helsinki. All procedures were carried out in accordance with the law, and the authorization documents (in the form of handwritten testaments) are kept in the files of the Body Donation Program. Upon collection, samples were fixed in 4% formalin solution, embedded in paraffin and stored at room temperature (RT) until the deparaffinization and clearing process.

  3. 3.

    Human hippocampus: samples with histopathological features of HS and cryptogenic epilepsy were obtained from adult patients who underwent surgery as treatment for drug-resistant epilepsy at the Fondazione Istituto Neurologico Carlo Besta (Milan, Italy). The resection was performed for strictly therapeutic reasons, after informed consent, and the extent of the excision was planned preoperatively based on both the epileptogenic zone localization and the risk of postsurgical deficits. Written informed consent was obtained for the surgical procedure and to use brain surgical specimens and clinical data for research purposes, in compliance with institutional and national ethical guidelines; the protocol for the study was approved by the Carlo Besta ethics committee (EpiBesta protocol n.83, 14-04-2021).

Deparaffinization of brain tissue

Deparaffinization was achieved by first removing as much paraffin surrounding the tissue block as possible with a blade. To completely remove the remaining paraffin, samples were placed in a water bath at 60 °C for 30–60 min until the paraffin was macroscopically eliminated. To speed up this step and completely clean samples from the paraffin, tissue blocks were positioned on a porous membrane inside a 50 ml tube floating in the water bath. Following the paraffin melting, tissue blocks were incubated 6–8 times in 98% xylene for 30 min at RT, using a rotary shaker. At this point, tissue blocks should appear translucent. Afterwards, samples were washed with 100%, 95%, and 70% ethanol diluted with water at RT for 30 min respectively, and further washed 3 times with 1× PBS for 10 min at RT. After embedding in 4% agarose, tissue sections with a thickness of 500 ± 50 μm were obtained using a vibratome (Leica VT1000 S) and stored at 4 °C in PBS + 0.01% w/v NaN3.

SHORT tissue transformation method

Single deparaffinized slices with a thickness of 500 μm were processed with the SHORT tissue transformation method15,26. The specimens were first incubated in a Switch-Off solution (50% v/v 1× PBS (pH 3), 25% v/v 0.1 M hydrochloric acid (HCl), 25% v/v 0.1 M potassium hydrogen phthalate (KHP), and fresh 4% v/v glutaraldehyde) at 4 °C in gentle shaking for 24 h. The solution was replaced with the Switch-On (1× PBS (pH 7.4) and fresh 1% v/v glutaraldehyde) and the incubation was performed at 4 °C in gentle shaking for 24 h. Afterwards, slices were washed 3 times for 2 h each in 1× PBS at RT and reactive glutaraldehyde was inactivated by o/n (overnight) incubation with an inactivation solution (1× PBS (pH 7.4), 4% w/v acetamide, 4% w/v glycine, pH 9.0) at 37 °C in a water bath. Following inactivation, the samples were washed in PBS at RT and then incubated in the clearing solution (200 mM sodium dodecyl sulfate (SDS), 20 mM sodium sulfite (Na2SO3), and 20 mM boric acid (H3BO3), pH 9.0 for 3–6 days (Table S1) at 55 °C. After the clearing step, samples were extensively washed in 1× PBS + 0.1% Triton X-100 (PBST) at 37 °C for 24 h. Next, the slices were treated with 30% H2O2 for 45 min and antigens were unmasked with the preheated antigen retrieval solution (10 mM Tris base (v/v), 1 mM EDTA solution (w/v), 0.05% Tween 20 (v/v), pH 9) for 10 min at 95 °C. After cooling down to RT, the specimens were washed in deionized (DI) water for 5 min each and then equilibrated with PBS for 1 h. The primary antibodies were diluted in PBST + 0.01% w/v NaN3 (Tables S1 and S2), and sample incubation was performed at 37 °C in gentle shaking. After washing with PBST at 37 °C for 24 h, the stained samples were incubated in gentle shaking with the secondary antibodies at 37 °C in PBST + 0.01% w/v NaN3 (Tables S1 and S2). Next, the stained samples were extensively washed with PBST at 37 °C, equilibrated first in 30% TDE/PBS (v/v), then in 68% TDE/PBS (v/v), and finally placed in the sandwich with 68% TDE. Detailed step-by-step procedures are available at protocols.io (https://doi.org/10.17504/protocols.io.kqdg3qmjev25/v).

iDISCO method

Single deparaffinized slices with a thickness of 500 μm were processed with a modified iDISCO protocol22. Specimens were dehydrated with increasing concentrations of methanol MeOH/H2O (20%, 40%, 60%, 80%, and 100%), with gentle rotation for 1 h each at RT. Samples were then incubated in 66% dichloromethane (DCM) / 33% MeOH overnight at RT with gentle rotation. Afterwards, samples were washed with 100% MeOH for 1 h twice, and autofluorescence was bleached through incubation with 5% H2O2 in MeOH at 4 °C overnight, without shaking. After bleaching, samples were rehydrated with decreasing concentrations of MeOH/H2O (80%, 60%, 40%, 20%), followed by one PBS and two PBST (1× PBS and 0.2% v/v Triton X-100) washes, all with gentle rotation for 1 h each at RT. On the same day, samples were incubated first with the Permeabilization solution (1× PBS, 0.2% v/v Triton X-100, 20% v/v DMSO and 0.3 M glycine), followed by incubation in Blocking solution (1× PBS, 0.2% v/v Triton X-100, 0.2% w/v gelatin porcin skin and 0.01% w/v NaN3), both for 24 h, at 37 °C with gentle rotation. Primary antibodies (Table S1) were diluted in PBSTW-Heparin (1× PBS, 0.2% v/v TWEEN-20 and 0.01 mg/ml Heparin) for 6 days, at 37 °C with gentle rotation. Samples were then washed for 1 day in PBSTw-Heparin and then incubated with the secondary antibodies (Tables S1 and S2) in PBSTw-Heparin for 4 days, at 37 °C with gentle rotation. After 1 day of washes in PBSTw-Heparin, samples were dehydrated as previously described and incubated for 3 h in 66% DCM / 33% MeOH. Finally, they were washed in 100% DCM for 30 min twice and incubated and stored in DBE (Dibenzyl ether) at RT.

TPFM imaging

A custom-made TPFM system63,64 was used to acquire high-resolution 3D image reconstructions of the brain tissue slices. The system employs a Chameleon (Coherent, US) tunable mode-locked Ti:Sapphire laser (pulse rate: 90 MHz, pulse width: 120 fs) as the excitation source. The light beam is laterally scanned by means of an optically coupled pair of galvanometric mirrors (LSKGG4/M, Thorlabs, US) and focused onto tissue samples using a 20× objective (Nikon CFI90 20XC Glyc, JP; numerical aperture (NA): 1.0, working distance (WD): 8.2 mm), with a resulting field of view of 621 μm × 621 μm. Two-photon excitation was achieved by operating the laser source in the near-infrared region at a wavelength of 800 nm. 655/40 nm, 618/50 nm, and 482/35 nm single-band bandpass filters (BrightLine®, Semrock, US) were respectively used to detect the fluorescence emission of neuronal bodies (NeuN), pRPS6+ neurons, and blood vessels (second harmonic autofluorescence). The filtered fluorescence emission is collected and amplified by three GaAsP photomultiplier tubes (H7422, Hamamatsu Photonics, JP) and digitized with 8-bit precision. The sequential imaging of adjacent overlapping 3D stacks is enabled by a closed-loop XY stage (U-780 PILine XY Stage System, Physik Instrumente, GER) for the lateral translation of tissue specimens, and a closed-loop piezoelectric stage (ND72Z2LAQ PIFOC Objective Scanning System, Physik Instrumente, GER) for the axial displacement of the objective. This, in turn, allows for the acquisition of high-resolution tiled image reconstructions of entire brain slices. The nominal lateral overlap of ~50 μm between adjacent tiles, adopted in the present work, ensured an optimal fusion with a precise superimposition of corresponding tissue structures. In this study, 512 × 512 pixel images were acquired with a z-step of 2 μm, resulting in a final voxel size of 1.21 μm × 1.21 μm × 2 μm. The uneven illumination (vignetting) of the separate images was corrected via the retrospective CIDRE method65, in order to prevent the grid-like shading artifacts that would otherwise arise when fusing the adjacent stacks. The flat-field intensity models related to the wavelengths of interest (655 nm, 570 nm, and 482 nm) were identified using a reference dataset comprising 7630 uncorrelated RGB images acquired with the employed objective.

LSFM imaging

A custom-made inverted LSFM setup described in a previous work of our group25,34 was also used to image brain tissue samples. This custom LSFM setup includes two identical 12× objectives (LaVision Biotec LVMI-Fluor 12× PLAN, GER; NA: 0.53, WD: 8.5 to 11 mm), oriented at 90° relative to each other and spaced so that their FOVs overlap in the center. These objectives are inclined at 45° with respect to the horizontal sample holder plane and alternately play excitation and detection roles. The microscope is equipped with four laser diode modules (Cobolt, SE) characterized by the following wavelengths and nominal output powers: (405 nm, 100 mW); (488 nm, 60 mW); (561 nm, 100 mW); and (638 nm, 180 mW). These excitation sources emit Gaussian beams which are adjusted in width by dedicated telescopes before being combined by means of three dichroic mirrors. This combined beam is then divided by a 50:50 beam splitter. Outgoing light is then introduced in two identical excitation pathways in which each beam is firstly modulated in intensity, phase, and wavelength using an acousto-optical tunable filter (AOTF, AOTFnC-400.650-TN, AA Optoelectronic, FR). The two AOTFs also enable the independent shuttering of each illumination pathway to avoid introducing stray light originating from the excitation beam during image acquisition. Secondly, the beam is scanned by means of a galvo mirror (6220H, Cambridge Technology, US) and then conveyed into a scan lens (#45-353, Edmund Optics, US; focal length (FL) = 100 mm, achromat) producing the planar light-sheet illumination. The generated light sheet is finally directed to the pupil of the 12× objective through an excitation tube lens (#45-179, Edmund Optics, US; FL = 200 mm, achromat), to illuminate a single plane of the tissue sample. Each objective is mounted on a motorized stage (PI L-509.14AD00, Physik Instrumente, GER) in order to move its focal plane position. On the other hand, the sample is translated using a 3-axis motorized stage system (two PI M-531.DDG and a PI L-310.2ASD, Physik Instrumente, GER; overall motion range: 30 cm × 30 cm × 2.5 cm) and imaged while being shifted along the horizontal x-axis, through the fixed FOVs of the two objectives. The fluorescence emitted by the sample and collected by the detection objective is separated from the reflected excitation light by means of a multiband dichroic beamsplitter (Di03-R405/488/561/635-t3-55×75, Semrock, US) before being directed by the detection tube lens onto an sCMOS camera (Orca-Flash4.0 v3 Hamamatsu, JP; pixel size: 6.5 μm × 6.5 μm, pixel array: 2048 × 2048). Five sets of band-pass filters were mounted in front of each camera on a motorized filter wheel (FW102C, Thorlabs, US) to selectively image the differently labeled structures within tissue samples. Cameras operate in confocal detection mode by having the rolling shutter sweep in synchrony with the galvo scan of the light sheet66,67. To prevent excitation, stray light from interfering with the active rolling-shutter rows on the acquiring camera, a time delay of half a frame is introduced between the detection and excitation modes of each optical path. Following the completion of an image stack, the tissue sample is shifted along the y-axis (by a range below the FOV side: ≈1109 μm), and the sequence is repeated until the whole tissue volume has been acquired. In this work, a y-step of 0.9 mm was adopted for a lateral image overlap of more than 100 μm. An x-step of 5.15 μm, instead, was set to obtain a native voxel depth of 3.64 μm.

Microscopy image post-processing

Since the cameras of the custom-made LSFM setup acquire images of brain tissue samples while these pass through the focal plane of the detection objectives with a 45° relative inclination, LSFM image stacks had to be processed to obtain isotropic volumetric images properly aligned with the sample holder. Therefore, raw stacks underwent a composite affine transformation (scaling, shearing, and rotation) by means of a custom Python tool which generated images oriented according to the reference system of the laboratory, characterized by an isotropic voxel size of 3.64 μm (i.e., the original voxel depth produced by the stage step motion).

Adjacent, overlapping TPFM and LSFM image stacks were then aligned and fused using the open-source ZetaStitcher tool (https://github.com/lens-biophotonics/ZetaStitcher)34. ZetaStitcher efficiently identifies the optimal alignment of adjacent image tiles by evaluating the FFT-based cross-correlation of overlapping regions at selected depths, enabling the high-throughput stitching of large volumetric microscopy datasets. The nominal lateral tile overlaps of ≈50 and ≈100 μm, respectively adopted for adjacent image stacks acquired with the custom TPFM and LSFM setups, guaranteed that no spurious doubling of corresponding tissue structures took place within overlapping areas.

Microscopy image segmentation

TPFM pixel classification: ilastik’s pixel classification workflow was used to generate 3D semantic segmentations of the TPFM image reconstructions, associating each voxel with predefined semantic class labels. Four separate classes were adopted: pRPS6- neurons, pRPS6+ neurons, blood vessels, and background. In detail, cell bodies expressing only the NeuN protein, thus solely associated with red fluorescence emission around 655 nm, were labeled as pRPS6-, whereas the pRPS6+ class was adopted when labeling neurons emitting light around 570 nm as well. pRPS6- and pRPS6+ cell detections were therefore mutually exclusive. ilastik’s random forest classifier, once trained by expert human operators, estimated the probability of each voxel belonging to each of the above classes, based on a set of multiscale voxel-wise features related to RGB intensity, edgeness, and texture. The adopted features more specifically comprised the output of Gaussian smoothing filters of varying full width at half maximum (multiscale intensity), the output of a range of Laplacian of Gaussian kernels (multiscale edgeness), and, finally, the eigenvalues of scaled Hessian matrices of second-order partial derivatives (multiscale texture). The spatial scales of interest were thoroughly selected according to the expected size of neuronal bodies. In detail, the feature scale space was sampled assuming minimum and maximum soma diameters of 15 and 25 μm, respectively, with a step size of 5 μm. The corresponding scales σ (i.e., the standard deviation of the smoothing filters performing the image scale selection) were then derived as follows:

$${{{{\rm{\sigma }}}}}_{{{{\rm{D}}}}}={{{\rm{D}}}}/\left(2\sqrt{3}\right)$$

where \({{{{\rm{\sigma }}}}}_{{{{\rm{D}}}}}\) is the optimal scale (in μm) where the local response to the presence of a binary sphere of diameter \({{{\rm{D}}}}\) is maximized (Lindeberg et al., 1992). Scale values in pixels, to be specified within ilastik’s feature selection widget, were finally obtained by taking into account the anisotropic voxel size of 1.21 μm × 1.21 μm × 2 μm (i.e., dividing by its lateral and longitudinal sizes).

The training of the classifier was conducted by three different operators on a random subset of TPFM image stacks, sampling and manually annotating one stack from each tiled reconstruction composing our dataset. The performance of the classifier trained for the automatic volumetric segmentation of pRPS6- and pRPS6+ neuronal subpopulations was systematically evaluated using a balanced set of 30 150 μm × 150 μm ROIs, selected within the TPFM acquisitions from the two study subjects (15 ROIs from each subject). The extracted ROIs were evenly distributed across the entire lateral extent of the tiled image reconstructions and across varying tissue depths, in order to enhance the representativeness of the validation set. This design aimed to account for potential variations in classifier performance due to differences in illumination, contrast, cellular distribution, and cell density. Specifically, the 2D ROIs were extracted from the 3D binarized segmentation outputs generated by ilastik and subsequently reviewed and corrected by an expert operator using QuPath68.] The corrected annotations were then quantitatively compared to the original automated binary masks by computing precision, recall, and F1 scores for each target class.

Neuronal body classification: pRPS6- and pRPS6+ cell probability maps returned by pixel classification were separately fed to the second object classification workflow and initially binarized by hysteresis thresholding. No preliminary smoothing was performed. In contrast to standard thresholding, this approach allowed on one hand to suppress noisy pixels within the background and, on the other hand, to preserve the actual shape of neuronal cell bodies by dilating the original high-confidence seed detections. Specifically, high and low thresholds of 80% and 50% were imposed on the input probability maps. Connected-component analysis69 was then applied to label connected pixel regions belonging to the foreground class and detect individual 3D objects.

Object classification through object-level features, following the binarization of the pixel-wise semantic probability maps, aimed at preventing the inclusion of spurious neuronal bodies and, secondly, at addressing the incorrect merging of densely packed cells into a single object. Six properties were considered in this regard, namely: size of the segmented object, radii of the object (i.e., the eigenvalues returned by the principal components analysis of the object pixels’ coordinates), mean object intensity, intensity variance, convex hull volume, and object convexity (defined as the ratio between the object volume and the one of its convex hull). The number of object features was purposely kept low to avoid overfitting regimes leading to poor generalization capabilities, as the size of the training set coincided with the number of annotated cells (significantly lower than the number of pixels annotated in the previous workflow). Again, the object classifier training was performed by three separate operators. Once trained, this classifier enabled us to automatically discriminate single cells, merged objects (with the number of fused cell bodies), and likely false detections.

Quantitative cell population analysis: Quantitative data returned by ilastik’s object segmentation workflow were further processed to assess the specific spatial distribution and morphology of pRPS6- and pRPS6+ cells. Four properties were investigated, namely: cell volume, cell ellipticity, local cell density, and clustering degree. Cell volume was directly provided by ilastik (being one of the features employed for object classification), whereas cell ellipticity was estimated as the ratio between the shortest and longest radii of each segmented object. The subsequent statistical analysis of these two features only focused on single-cell detections, in order to avoid any bias due to merged neuronal bodies.

On the other hand, local cell density and spatial clustering were evaluated as previously described67. Specifically, the spatial clustering of neurons was evaluated using the 3D Ripley K-function (Ripley, 1976):

$${{{\rm{K}}}}({{{\rm{r}}}})={{{\rm{n}}}}({{{\rm{r}}}})/{{{\rm{\lambda }}}}$$

where \({{{\rm{n}}}}({{{\rm{r}}}})\) represents the number of neighboring cells within a radius \({{{\rm{r}}}}\), and \({{{\rm{\lambda }}}}\) is the local cell density, evaluated in a sphere of diameter 300 μm surrounding the centroid of each neuronal body.

This reference volume is indeed substantially larger than the average cell but still smaller than the size of the analyzed TPFM image reconstructions. Under the hypothesis of a completely random spatial distribution, the expected value of \({{{\rm{K}}}}({{{\rm{r}}}})\) corresponds to the sphere volume:

$${{{\rm{E}}}}[{{{\rm{K}}}}({{{\rm{r}}}})]=4{{{\rm{\pi }}}}{{{{\rm{r}}}}}^{3}/3$$

The basic principle of the formulated 3D cell clustering index, \({{{\rm{I}}}}\), is that local violations of such a random condition can be quantified for each neuron by integrating the deviations from the above-expected value over a specific range of distances:

$${{{\rm{I}}}}=\int _{{{{{\rm{r}}}}}_{\min }}^{{{{{\rm{r}}}}}_{\max }}{{{\rm{K}}}}\left({{{\rm{r}}}}\right)\,-{{{\rm{E}}}}\left[{{{\rm{K}}}}\left({{{\rm{r}}}}\right)\right]{{{\rm{dr}}}}$$

In the present work, a minimum and maximum radius of \({{{{\rm{r}}}}}_{\min }=10\,\)μm (the approximate average soma radius), and \({{{{\rm{r}}}}}_{\max }\) \(=100\) μm (below the radius of the reference sphere, used for local density estimation) were respectively selected, along with a discrete integration step of 10 μm.

Local cell density and spatial clustering of each segmented cell object were concurrently computed by means of an optimized Python tool leveraging multiprocessing and SciPy’s k-d tree implementation for fast nearest-neighbor lookup70 to shrink the computation time required to handle our dataset of cell body detections. Regarding the spatial mapping of these two neuronal distribution properties, merged cell objects were carefully taken into consideration by transforming their original centroid into a proper number of random coordinates surrounding it (e.g., two 3D coordinates for an object classified as being formed by two cell bodies). Such coordinates were specifically obtained by first defining a set of random unit vectors, and then multiplying their length by distance values sampled from a uniform random distribution, bounded by the diameters corresponding to the 5% and 95% percentiles of the volume of single-body detections (18.9 μm and 21.2 μm, respectively, assuming a spherical shape). This mechanism aimed at preventing the underestimation of local density (and thus clustering) values, which would be understandably significant if the presence of these merged objects, corresponding in fact to tightly packed cell bodies, were neglected.

Statistics and reproducibility

To ascertain if the aforementioned neuronal cell measures were subject-specific, that is, dependent on the pertinent diseases, or influenced by the presence of phosphorylated S6 ribosomal proteins, data were initially assessed using the nonparametric Scheirer-Ray-Hare test (Scheirer et al., 1976). This was carried out by means of a custom R script (R Core Team, 2021). Subsequently, nonparametric Mann–Whitney U tests for independent samples were performed for pairwise comparisons across subjects and channels (i.e., pRPS6- versus pRPS6+ neurons). Pairwise Mann–Whitney tests were implemented through an original Python script, using SciPy’s stats module. All results were deemed statistically significant if their corresponding p-value was below 0.05. Bonferroni correction was applied in the case of the pairwise Mann–Whitney U tests, adjusting the above significance threshold in relation to the overall number of possible comparisons (i.e., \({{{\rm{\alpha }}}}/6\approx 0.008\)). Dissimilarity between the compared data distribution pairs, \({{{\rm{p}}}}({{{\rm{x}}}})\) and \({{{\rm{q}}}}({{{\rm{x}}}})\), was described in terms of their Hellinger distance, \({{{\rm{H}}}}({{{\rm{p}}}},{{{\rm{q}}}})\)|\(0\le {{{\rm{H}}}}({{{\rm{p}}}},{{{\rm{q}}}})\le \,1\) (Hellinger, 1909). For a pair of discrete probability distributions, \({{{{\rm{p}}}}}_{{{{\rm{d}}}}}=({{{{\rm{p}}}}}_{1},\,...,{{{{\rm{p}}}}}_{{{{\rm{n}}}}})\,\) and \({{{{\rm{q}}}}}_{{{{\rm{d}}}}}=({{{{\rm{q}}}}}_{1},\,...,{{{{\rm{q}}}}}_{{{{\rm{n}}}}})\), this metric can be computed as:

$${{{\rm{H}}}}({{{\rm{p}}}},{{{\rm{q}}}})=\sqrt{1-{\sum }_{{{{\rm{i}}}}=1}^{{{{\rm{n}}}}}\sqrt{{{{{\rm{p}}}}}_{{{{\rm{i}}}}}{{{{\rm{q}}}}}_{{{{\rm{i}}}}}}}$$

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.