Neuromethods: Basic Techniques for Evaluating the Nervous System in Nonclinical Studies

Abstract

The first session of the 2025 European Society of Toxicologic Pathology (ESTP) Congress reviewed routine and specialized methods for microscopic evaluation of neural tissues during nonclinical studies. Three longer presentations reviewed brain sampling approaches in safety assessments, including an example to accentuate topographical analysis and integration of toxicology data; specific brain and spinal cord sampling for molecular and protein analyses; and an overview of technical aspects of intraparenchymal drug delivery. Four shorter talks discussed the uses, advantages, disadvantages, and interpretation of several special neurohistological techniques (stains and immunohistochemical markers) for assessing test item-associated responses. Common special methods used (when warranted) for nonclinical studies include Fluoro-Jade or silver stains for detecting neuronal death, Luxol fast blue (LFB) for examining myelin, anti-glial fibrillary acidic protein (GFAP) to demonstrate reactive astrocytes, and anti-ionized calcium-binding adaptor molecule 1 (IBA1) to highlight reactive microglia and macrophages, though alternatives methods were described. The last presentation discussed artificial intelligence as an aid in detecting subtle toxicant-induced lesions during digital pathology analyses (using the Olney lesion [acute neuronal vacuolation and necrosis in the cerebral cortex] as an example). Taken together, talks in this session provided a cohesive overview of traditional and innovative approaches to facilitate microscopic evaluation for potential neurotoxicity in nonclinical studies.

The 2025 annual congress of the European Society of Toxicologic Pathology (ESTP) was held as one component of the 5th Cutting Edge Pathology Congress. This joint symposium was organized as an ESTP meeting convened in parallel with a combined meeting of the European College of Veterinary Pathologists (ECVP) and the European Society of Veterinary Pathology (ESVP). The Congress was held at the Real Centro Universitario Maria Cristina in San Lorenzo de El Escorial (Madrid), Spain from August 28 to 31, 2024.

The ESTP congress was entitled “Neuropathology: The Vast Pink Wonderland.” This theme was chosen for three reasons. First, neurological conditions are widely prevalent around the world and therefore present substantial economic and sociological burdens on families, communities, and nations faced with providing care for patients with severe and/or life-threatening, often chronic diseases. Second, our ever-growing understanding of the pathogenesis for such disorders has provided a burgeoning armamentarium available to specifically target the molecular and cellular bases of many neurological disorders, offering the hope of considerable amelioration or even cures for numerous conditions previously refractory to traditional treatments. Finally, from a histopathology perspective, the “vast pink wonderland” acknowledges the varying shades of pink visible in conventional hematoxylin and eosin (H&E)-stained neural tissue sections. The “Neuropathology” topic was addressed in a series of sequential sessions covering methods for nervous system sampling, processing, and evaluation; characteristics of neurodegenerative diseases in patients and many nonclinical models (in vitro, ex vivo, and in vivo); developmental neurotoxicity; and recent innovations in neural tissue analysis. The materials presented during the ESTP congress included presentations (full-length and brief talks), case reports, panel discussions with audience interactions, and keynote lectures. The speakers and audience members gathered from around the world.

The current paper is a synopsis of the principles and practices discussed during the first session of the ESTP congress, which was dedicated to a review of routine and various special techniques (collectively “neuromethods”) used in sampling, processing, and evaluating neural tissues during nonclinical studies of novel test items (or test articles), mainly chemicals and therapeutic candidates of different product classes. This session contained 7 oral presentations and a final audience participation event. Subjects covered in this Neuromethods session are arranged below in the order in which they were presented.

Approaches to Nervous System Sampling

The neuroanatomical heterogeneity in the central nervous system (CNS), and especially the brain, may be broad over very short distances, yielding an intricately complex series of circuits and layers with divergent neurophysiological and neurochemical properties. The organization of the nervous system in context of function poses multiple unique challenges in the evaluation of potential neurotoxicity. Therefore, nonclinical study designs must incorporate carefully devised strategies for histopathological sampling of neural tissues to achieve meaningful results.

Optimal Neural Sampling Depends on the Scenario

The first talk, presented by Dr Deepa Bandi Rao, focused on nervous system sampling for histopathology evaluation in routine safety assessments. The lecture began with a broad overview on tissue sampling in toxicology studies that then tapered to a focus on nervous system tissues. In response to an initial request from the Center for Drug Evaluation and Research (CDER) at the US Food and Drug Administration (FDA), an ad hoc task force from the Society of Toxicologic Pathology (STP) was convened to compile a recommended tissue list for histopathology examination in repeat-dose toxicology and carcinogenicity studies. The list included approximately 42 tissues. ¹³ An additional set of nine tissues were later included following a response to Bregman et al ¹³ 2003 via a Letter to the Editor by FDA authors. ²² In general, this core list of approximately 50+ tissues constitutes what is currently evaluated in conventional safety toxicology studies. Although tissues reflecting nervous system components (brain, spinal cord, peripheral nerve, and eyes [retina] with optic nerves) were identified individually, the vast innervation distributed to all other organ systems necessitates that the pathologist evaluate all organs with a sense of awareness regarding internal neural elements, so that interpretation of findings in the nervous system can be undertaken in an integrated manner.

Brain atlases typically identify approximately 600 neuroanatomical areas, thereby emphasizing awareness of various neuroanatomical areas (or subsites) during histopathology evaluation.^30,31 Historical approaches to traditional brain sampling practices were presented in the talk wherein the underlying rationale for the transition from the evaluation of 3 brain slices to at least 7 brain slices in rodents ³⁰ was underscored. Specifically, the screening of at least 10% of the 600 neuroanatomical areas as reflected by the 60 to 70 identifiable neuroanatomical areas in 7 brain slices/sections (rodents) in safety toxicology studies^30,31 was highlighted. Brain sampling in rodents and non-rodents was compared to emphasize that brain slices (trimming planes) are not equivalent to brain sections when comparing small and large brains. While 7 brain slices equate to 7 brain sections in small animals (rats and mice), this is not so with large animals (rabbits, dogs, minipigs, nonhuman primates [NHP]) due to the large brain size and correspondingly larger individual neuroanatomical areas. In fact, in large animal brains, at least 2 sections (see Sections 3A and 3B in Figure 2 of Bolon et al ⁹ ) are typically needed to adequately assess the brain fields available in 1 hemi-brain slice (Figure 1). Thus, compared with rodents, 7 sections in large brains of non-rodents reflect substantially less than the basic 10% screening that is desirable for safety assessment studies.

Figure 1.

This image represents an example of a single sub-gross 4-mm-thick brain slice (left) and its associated sections stained with Hematoxylin & Eosin (right) from a non-human primate. Note that the entire hemi-brain slice does not fit into a single standard tissue cassette (gray plastic) used for routine embedding. It takes approximately four standard size tissue cassettes to embed a single brain slice (bilateral), or two standard tissue cassettes to embed a hemi-brain slice. One tissue cassette translates to one brain section. Trimming edges off brain slices to fit into tissue cassettes is not recommended—the objective during brain trimming should be to retain the capability to track and retrieve additional neuroanatomical areas following initial histopathology evaluations. Although large-size tissue cassettes (that allow complete brain/hemi-brain slice embedding) are available, these are rarely used in nonclinical toxicologic pathology but can be justified for use in intraparenchymal direct delivery studies. This figure is extracted from Figure 2 in Bolon et al. ⁹ by courtesy of Sage.

Multiple approaches for effective neural sampling have been published across various species over the years^{5 -7,18,25 -28,30,31} including recommended best practices generated by an STP working group (2013) that is generally followed by many global organizations engaged in nonclinical safety testing.^9,11 While these serve as excellent guidelines, protocol development should necessarily integrate all available information to identify strategic sampling poised to detect potential neurotoxicity in nonclinical toxicology studies. For example, a study design that follows any of the above published guidelines but does not include potential target neuroanatomical areas (for example, based on primary/secondary pharmacology data for the test compound) fails to meet the objectives of a safety assessment study. Specifically, for non-rodents, sampling should be aligned to include brain trimming strategies that balance sufficient evaluation of various neuroanatomical areas of the brain, selective approaches specific for a given test compound, as well as flexibility to range across tiered evaluations that aligns with the guiding principle on “Reduction” to minimize the number of animals based on the “Three Rs.” ¹⁴

Because neuronal cell populations within neuroanatomical areas are physiologically interacting with each other via defined neuroanatomical pathways, and because it is not always possible to capture all components of every specific pathway in a given safety toxicology study, neuropathology evaluation can be especially challenging and is impacted by sampling to a considerable extent. Moreover, with limited sampling, a lesion within a specific neuroanatomical subsite may easily be misinterpreted as a sporadic or spontaneous finding (as when it is not seen in any of the other available sections) when evaluated out of context of topographical analysis.

Dr Rao’s talk concluded with the depiction of an example wherein minimal nerve fiber degeneration in the spinal trigeminal tract in the brain (Figure 2) was eventually linked with morphological and functional effects along the trigeminal (cranial nerve V) neuroanatomical pathways. Although the finding of oligodendrocyte apoptosis (reflecting nerve fiber degeneration) can be subtle and potentially missed with digital pathology evaluations, ³² Dr Rao stated that this finding was specific and consistent along its neuroanatomical trajectory in both rodents and non-rodents as well as with varying therapeutic modalities (small molecules, large molecules, and biologics including adeno-associated virus [AAV] vectors gene therapy targeting various sensory ganglia). The example emphasized the importance of topographical analysis, a need to evaluate additional neuroanatomical areas or additional neuromethods as deemed by the evaluating pathologist, and necessary strategic approaches to integrate all available study data when interpreting findings in neuropathology evaluations (especially those based on limited sampling). The lecture was concluded with a call to action, outlining future directions to bridge existing challenges and gaps toward building strategies for neuropathology evaluations.

Figure 2.

This image shows nerve fiber degeneration (arrows) within the spinal trigeminal tract (dotted lines) in a coronal/transverse section of the brainstem in a rat (unperfused) available for evaluation in a routine safety assessment study. When noted, additional histopathology evaluations along the trigeminal pathways are warranted. 20× objective.

Proper Neural Sampling Supports Integrated Anatomical, Biochemical, and Functional Analyses

The second talk, given by Dr Xavier Palazzi, extended the first lecture by defining sampling approaches that optimize collection of neural tissues to permit an integrated histopathological evaluation and biochemical analysis of a discrete neural structure with a known function. The premise for this approach is that molecules of interest are bilaterally symmetrical in distribution and activity, thereby yielding bilaterally symmetrical changes in molecular (nucleic acid) biology, function, and structure. Such biochemical assays are used to assess biodistribution or target engagement by test items, and their presence in regions exhibiting microscopic findings represents a key part of comprehensive neurotoxicity testing. In particular, biodistribution studies conducted in the context of ICH S12 (International Council for Harmonisation of Technical Requirements for Pharmaceuticals for Human Use) to assess pharmacology and toxicology supporting early-phase clinical trials of gene therapy products require a specific methodology to sample fresh brain and spinal cord and allow for subsequent routine and investigative pathology as well as molecular biology investigations. ²¹

The sampling approach for this parallel evaluation depends on several factors. At necropsy, animals may be euthanized and tissues removed at once, or animals may be perfused with ice-cold physiological saline to slow cellular processes that can degrade ribonucleic acids ¹⁰ ; perfusion fixation with fixative is avoided lest it destroy the molecules of interest. Brain is removed from the skull, placed in a chilled brain matrix or on a chilled cutting board, and sliced rapidly using chilled instruments to produce a tissue slice in the proper orientation (usually a series of transverse full coronal slices). Unfixed tissue punches for nucleic acid (vector and/or outcome of the transduction process) analysis from structures in one hemisphere are then removed and placed in test tubes on ice (to be stored for further analysis) after which the tissue slice from the opposite undisturbed hemisphere is immersed in a suitable fixative (usually commercial neutral buffered 10% formalin [NBF] containing 1% methanol as a stabilizer but sometimes methanol-free 4% formaldehyde [MFF], known colloquially as “paraformaldehyde”). For spinal cord and nerve, long segments acquired from representative sites (e.g., cervical and lumbar divisions of the spinal cord, the middle of the sciatic nerve trunk) are parsed into shorter segments, after which adjacent segments are dedicated to biochemical analysis. For ganglia, a common strategy used for evaluating dorsal root ganglia (DRG) is to dedicate one ganglion at a given spinal cord level for histopathological evaluation while allocating the other to nucleic acid or protein analysis for AAV vector or transgene distribution. ⁴ However, recent research suggests that a better sampling approach is to assign both members of the ganglion pair for a given spinal cord level for microscopic evaluation and both members of the pair for the adjacent level for molecular analysis (unpublished data). The reason for this latter approach is that AAV-related ganglionic pathology does not always occur bilaterally, so there is no guarantee that molecular data from one ganglion in the pair can be correlated with data from the microscopic evaluation of the opposite ganglion.

Neural Sampling After Direct Delivery of Test Items to the Nervous System

The third talk, delivered by Alexandra Duetting and Dr Annette Romeike, focused on technical aspects and pathology findings in nonclinical studies assessing direct drug delivery, focusing on introduction into the brain parenchyma in NHPs. Direct delivery is necessary for certain test item classes as systemically administered agents (e.g., antisense oligonucleotides [ASOs], gene therapy vectors) may be restricted from entering by the blood-brain barrier (BBB). Intrathecal (IT) injection into the cerebrospinal fluid (CSF) via the cisterna magna (intracisterna magna [ICM] route), lateral ventricles (intracerebroventricular [ICV] route), or lumbar cistern (intra-lumbar cistern [ICL] route) is an established means of thwarting the BBB.^2,3,15,19,36 Nonetheless, substances introduced into CSF often fail to penetrate deeply into the dense, lipid-rich brain parenchyma. To overcome this impediment, intraparenchymal (IPa) delivery to circumscribed regions using stereotactic injection or infusion is an increasingly popular choice for test item administration.

The impact of the route of direct CNS delivery has been clearly demonstrated by recent forays using AAV-based gene therapy vectors. Delivery by ICM or ICV injection can result in broad distribution of the AAV construct throughout the brain, but direct IPa infusion usually leads to higher local exposure of desired brain regions. All three routes require anesthesia. The cisterna magna may be entered using a conventional needle, but delivery to the ventricles or parenchyma requires osteotomy to penetrate the skull to provide a portal for the catheter. In non-rodents, up to 2 mL of dosing solution may be delivered into the cisterna magna or ventricles while IPa administration should typically be limited to 100 µL/hemisphere with a slow infusion rate (≤ 10 µL/min). Gradient infusions (gradually rising number of µL/min) are usually tolerated better than constant infusions. ⁸

Neural sampling for IPa and IT studies is based on the principles noted above but is generally adjusted on a case-by-case basis to optimize evaluation of the target region and effects that might occur along the injection trajectory. For IPa delivery to the brain, procedure-related microscopic findings are modest (minimal to mild) and typically include edema, hemorrhage, and localized parenchymal reactions including linear injection tracks, tissue loss (necrosis [“cavitation”]) at the site of test item deposition, and gliosis (mainly microglial cell reactions) in nearby tissue. ⁸ Some test items (e.g., AAV vectors) may exacerbate these IPa-related procedural findings to a modest degree while also invoking other changes such as mononuclear cell infiltrates around small blood vessels or along the margins of necrotic areas. ⁸ For IT delivery, procedure-related changes include linear injection tracks in the dorsal brain (brainstem for the ICM route, cerebral cortex for the ICV route) or lumbar or sacral spinal cord (for the ICL route); brainstem and spinal cord findings arise from inadvertent parenchymal damage when the injection apparatus is passed through the intended site of administration (a CSF-filled space) and into the underlying tissue. In general, test items delivered by IT injection (e.g., AAV vectors, ASOs) are often accompanied by modest mononuclear cell (or less often mixed cell) infiltrates in the meninges, more frequently near the site of administration. ² Nerve fiber degeneration may be observed whenever direct CNS delivery is performed, but in our experience this finding is most common for IT injections and is often associated with procedure-related injury at or near the site of administration (mainly to the spinal cord and spinal nerve roots).

Special Neurohistological Methods to Highlight Microscopic Findings in Neural Tissues

The baseline stain for evaluating neural tissues in nonclinical studies is hematoxylin and eosin (H&E). This method is suitable for detecting obvious structural changes but is less sensitive for detecting subtle alterations in subcellular structures and various chemical components of neural tissues. Many special neurohistological procedures have been developed to highlight various neural constituents, and some of them have become mainstays for assessing CNS and peripheral nervous system (PNS) constituents in health and disease. This Neuromethods session continued with four mini-talks discussing core neurohistological techniques that are utilized during nonclinical studies—when warranted—as a supplement to H&E in evaluating neural tissues for potential neurotoxic changes. These special methods are used to more effectively identify and characterize potential neurotoxic effects and target cells/tissues, and in some cases may reveal changes that cannot be observed in H&E-stained sections.

Silver stains to detect neurotoxicity were described by Dr Kristel Kegler and Dr Klaus Weber. Silver-based methods have been used historically as a histochemical tool to demonstrate axonal features and distribution ²⁹ as well as to detect neuroaxonal degeneration.^33,35 Among all silver-based protocols developed to date, the most used silver staining methods in laboratory animals (and humans) are modified Bielschowsky, modified methenamine-silver, Bodian, Gallyas (GAL), and Campbell–Switzer. From a practical perspective, the axonal affinity for silver depends not only on the protocol but also on the lesion to be evaluated. For example, silver staining methods are selectively applied in detecting interrupted axonal tracts between CNS-specific regions (e.g., Bielschowsky), in identifying senile plaques and neurofibrillary tangles present in aging or in Alzheimer’s disease (e.g., Campbell-Switzer), in quantifying axonal damage and loss in multiple sclerosis (e.g., modified Bielschowsky), and for screening irreversible neuroaxonal injury caused by neurotoxic substances (e.g., amino cupric silver). Advantageously, silver stains can be combined with various immunohistochemical (IHC) markers to co-localize altered structural proteins within axons to specific functions, thereby providing support for specific mechanisms of neurotoxicity. Hence, neuropathologists and neurotoxicologists should be aware of differences among various silver methods and their targets (Table 1) as well as the suitability of co-localization approaches to avoid ambiguities when interpreting silver staining results in animal models for CNS degenerative diseases or when assessing test item-related neuroaxonal toxicity in nonclinical studies.

Table 1.

Selected silver stains used to detect specific features within the central nervous system (CNS).

Feature stained	Golgi silver	Bielschowsky	Modified Bielschowsky	Campbell-Switzer	Holmes	Gallyas	Bodian
Neuronal body	yes	yes	no	no	no	no	no
Axons	yes	yes	yes	yes	yes	yes	yes
Degenerated neurons	no	yes	yes	no	no	yes	yes
Degenerated axons	no	yes	yes	yes	yes	yes	yes
Special features
Pick bodies	—	yes	yes	yes	—	no	yes
AD-NFTs Down-NFTs DNTC-NFTs	—	yes	yes	yes	yes	yes	yes
PSP/CBD-neurons	—	not consistent	not consistent	no	—	yes	yes
PSP/CBD-glia	—	not consistent	not consistent	no	—	yes	not consistent
AD-SPs	—	yes	yes	yes	—	yes	yes
AD-diffuse deposits	—	not consistent	not consistent	yes	—	no	no
Lewy bodies	—	yes	yes	yes	—	no	yes
GCIs	—	yes	yes	yes	—	yes	yes
Argyrophilic granules	—	yes	yes	no	—	yes	yes

Abbreviations: AD, Alzheimer’s disease; CBD, corticobasal degeneration; DNTC, diffuse NFTs with calcification; GCIs glial cytoplasmic inclusions; NTTs, neurofibrillary tangles; PSP, progressive supranuclear palsy; SP, senile plaques; —, not tested/unknown. Data from modified methamine-sliver stain was not included in this table.

Methods for demonstrating glial responses were discussed by Dr Stefanie Arms and Dr Enrico Vezzali. Careful selection of IHC markers for various glial cell types is frequently important in characterizing morphological changes and their pathogenesis in neural tissues, ¹ and may also be useful in evaluating lesion reversibility and assigning adversity. Common IHC markers for glial cells used in nonclinical studies are glial fibrillary acidic protein (GFAP) and sex-determining region Y (SRY)-related HMG box gene 9 (SOX9) for astrocytes, oligodendrocyte transcription factor 2 (Olig2) for oligodendrocytes, and ionized calcium-binding adaptor molecule 1 (IBA1) for microglial cells. ²⁰ Additional markers are available for these glial cell classes (Figure 3) ²⁰ ; some markers are suitable for all nonclinical species (e.g., GFAP, IBA1) while others are expressed in some but not all species (e.g., CD68, which can be detected in rodents but not dogs ³⁴ ). The nature of a glial reaction may need to be assessed using several markers for a single glial cell type. For example, astrocytes expressing both GFAP and SOX9 are reactive (ie, astrocytosis) while astrocytes that are positive for SOX9 but negative for GFAP are maintaining their normal homeostatic (“resting”) state. ³⁴ Similarly, microglial cells in a homeostatic state express P2ry12 purinergic receptor P2Y (P2Y12), reactive microglia are positive for Cluster of Differentiation 68 (CD68), and both homeostatic and reactive microglia are labeled by IBA1. For routine nonclinical studies, GFAP and IBA1 are the workhorse methods due to their well-characterized staining patterns and the robust nature of their antigens, which permits their recognition in conventionally processed (formalin-fixed, paraffin-embedded) neural tissues. In general, these cell type−specific markers are used when warranted (e.g., gliosis is observed in H&E-stained sections, the test is anticipated or known to be neurotoxic) and not for general toxicity studies. ⁷

Figure 3.

Overview on immunohistochemical markers for detecting macroglia (astrocytes and oligodendrocytes) and microglia. Markers highlighted in crimson are often considered, where warranted, for use in nonclinical safety testing. ALDH1L1 indicates aldehyde dehydrogenase 1 family member L1; CD68, Cluster of Differentiation 68; GFAP, glial fibrillary acidic protein; GSTπ, glutathione S-transferase pi; IBA1, ionized calcium-binding adapter molecule 1; MBP, myelin basic protein; NG2, neural/glial antigen 2; Olig2, oligodendrocyte transcription factor 2; P2Y12, purinergic receptor P2Y12; PDGFRα, platelet-derived growth factor receptor alpha; S100, non-acronym for a protein soluble in 100% saturated ammonium sulfate solution at pH 7.0; SOX9, sex-determining region Y (SRY)−related HMG box transcription factor gene 9; SOX10, SRY-related HMG box transcription factor gene10.

Myelin stains were addressed by Dr Brad Bolon and histology experts Amber Moser and Elizabeth Chlipala. Fine cytoarchitectural details of myelin are hard to appreciate in H&E-stained sections, so special stains are often utilized to better characterize myelin integrity if damage is anticipated to myelin sheaths and/or myelinating cells (oligodendrocytes in the CNS, Schwann cells in the PNS, and olfactory ensheathing cells for the olfactory nerve [cranial nerve I]). Myelin procedures that may be used as supplemental studies for nonclinical studies include acidophilic dyes (e.g., toluidine blue [used with hard plastic sections]); lipoprotein-binding dyes (e.g., Luxol fast blue [LFB]); lipid impregnation with metals (e.g., Marchi’s, which uses osmium tetroxide for en bloc staining before embedding); and IHC methods to highlight various myelin antigens (e.g., myelin basic protein [MBP] and peripheral myelin protein 22 [PMP22]). ¹² The two most common choices for assessing myelin integrity for neurotoxicity screening studies are LFB and MBP. Rarely, neurotoxicity assessments may require transmission electron microscopy (TEM) to discern the laminar arrangement in myelin sheaths. The inclusion of a myelin method may be driven by a regulatory expectation that myelin should specifically be investigated as a potential target while the choice of myelin method often depends on the appearance of myelin-rich regions in H&E-stained sections.

Artificial Intelligence as an Aid to Neuropathological Evaluation

The last presentation describing incorporation of artificial intelligence (AI) to enhance detection of a neuropathological finding was delivered by Erio Barale-Thomas. Identification of the Olney lesion, a well-characterized but very subtle and transient acute neuronal vacuolation and necrosis that occurs in the retrosplenial cortex of the rat brain,^17,23 was used as an example to emphasize pitfalls related to the use of an AI workflow rather than the more familiar dangers inherent in routine histology processing and microscopic evaluation. In particular, successful AI-based digital pathology analysis requires close collaboration among the pathologists, histologists, and computational scientists.

Implementation begins with appropriate sampling and processing of the neural sample. Industry standard techniques^9,11 are often helpful, although unusual regions of interest may necessitate alternative trimming protocols. Profiles of bilaterally symmetrical structures should be highly homologous on both sides of the section and for different animals. Whole slide scans should have sufficient resolution to permit discrimination of the key diagnostic features; for the Olney lesion, scans should be acquired using a 40× objective (yielding a resolution of approx. 0.25 μm/pixel) due to the small size of the cell changes that define this finding (Figure 4). The real challenge to AI incorporation in this instance was the need for inter-laboratory cooperation to perform the initial manual microscopic evaluation and the subsequent algorithm assembly, training, and validation.

Figure 4.

Representative examples of the Olney lesion over time from the digital pathology data set used for AI model development. This subtle lesion is characterized by punctate cytoplasmic neuronal vacuolation within hours of toxicant exposure, evolving to necrotic neurons (classic “red dead” cells with condensed nuclei and hypereosinophilic cytoplasm) by 2 to 4 days following exposure with clearance of the affected cells within a few more days. The first column shows the vacuolation at the early time point (6 hours post-dosing) and the second column, the necrosis at the late time point (7 days post-dosing). The top row shows perfusion-fixed samples while the bottom row shows immersion-fixed samples (with more artifactual white spaces in the brain parenchyma). Neuronal vacuole, true positive (1); neuronal necrosis (2); artifact, white space (3); artifact, dark neuron (4). Scanned at 40×.

Three key challenges were encountered in developing the AI solution to reliably detect the Olney lesion. The first was to choose between a high-magnification approach for accurate cell counting versus a statistical approach to assessing the presence of the lesion on each slide and for a given treatment group. The second challenge was to be able to discriminate and, when warranted, quantify related findings that precede or occur in tandem with classic “red dead” neurons (which feature shrunken, often angular cell profiles with hypereosinophilic cytoplasm and shrunken nuclei). In this regard, the AI solution had to recognize dark neurons as an artifact, ²⁴ necrotic neurons that lacked a nucleus, and various classes of neuronal vacuolation. Finally, the AI solution had to be able to delineate affected neuron fields with considerable precision to increase confidence that the algorithm was functioning properly. In the end, these challenges were overcome, and the proposed AI-based solution is currently being used to assess digitized images of brain sections from several organizations.

Conclusions

In summary, strategic sampling and processing methods for neural tissues are instrumental in the success of anatomical assessments during nonclinical safety testing for potential neurotoxicity. The traditional morphology-based screening methods are often suitable for general toxicity testing, but incorporating more sampling and/or special neurohistological methods (ie, “expanded neurohistopathology evaluation”^7,16) are frequently helpful in identifying and characterizing potential neurotoxic effects. In the future, inclusion of AI-based solutions in digital pathology workflows is likely to further enhance the sensitivity and speed with which complex neural tissues can be examined. This session synopsis paper provides a solid launching point for individuals seeking guidance on points to consider as well as references to examine when seeking to inject advanced neuromethods into their nonclinical safety program.