Volumetric neuropathology assessment

Abstract

The invention relates to systems, methods, and computer-readable media for automated assessment of volumetric changes in brain regions of interest in preclinical animal models. Digital images of brain tissue cross-sections are processed using a convolutional neural network (CNN) to isolate regions of interest, followed by a transformer-based artificial intelligence model utilizing axial and patched attention for pixel classification. Segmentation accuracy is enhanced through a topological data analysis (TDA) algorithm, which refines results by constructing and evaluating simplicial complexes to ensure anatomical continuity. The refined segmentation data is used to calculate three-dimensional volumes based on the Cavalieri principle, and the results are displayed or stored for further analysis. The system offers improved accuracy, efficiency, and consistency over manual methods, supporting preclinical evaluation of therapies targeting neuropathology. Applications include tracking volumetric changes over time and facilitating high-throughput analyses in neurodegenerative disease research.

Description

VOLUMETRIC NEUROPATHOLOGY ASSESSMENTCROSS-REFERENCE TO RELATED APPLICATIONSThis nonprovisional application claims priority to U. S. Provisional Application No. 63 / 724,785, entitled “Volumetric Neuropathology Assessment.” Filed November 25, 2024 by the same inventors.BACKGROUND OF THE INVENTION1. FIELD OF THE INVENTIONThe described embodiments relate generally to automated volumetric analysis of brain regions in preclinical models. Specifically, the described embodiments relate to systems and methods for using transformer-based artificial intelligence and topological data analysis to segment, refine, and calculate three-dimensional brain region volumes from two-dimensional cross-sectional images.2. BRIEF DESCRIPTION OF THE RELATED ARTThe rising prevalence of neurodegenerative diseases, such as Huntington disease (HD), coupled with the need to evaluate novel treatments, underscores the demand for accurate, rapid, and resource-efficient methodologies for neuropathological assessment in preclinical animal models. Neurodegenerative diseases, characterized by the progressive loss of neuronal structure and function, impose significant societal and economic burdens. As therapeutic interventions increasingly aim to slow or reverse neuropathology rather than simply alleviate symptoms, the ability to assess treatment efficacy with precision has become increasingly important. However, existing methodologies and tools for volumetric analysis present various limitations in terms of efficiency, accuracy, and scalability, complicating their widespread use for preclinical evaluations.T raditional stereological methods for regional brain volume assessment involve manual, labor-intensive procedures that limit throughput and introduce potential biases. These approaches typically require generating equidistant cross-sectional brain images spanning a region of interest, staining them to highlight cytoarchitecture, and imaging them for further analysis. Researchers must manually trace the boundaries of the brain region in each cross-section to calculate two-dimensional (2D) areas These 2D measurements are then integrated using the Cavalieri principle to estimate the three-dimensional (3D) volume of the region. While effective in principle, this method is time-consuming, prone to subjectivity, and lacks the scalability required for large datasets or high-throughput studies.Advances in technology have sought to address these limitations but remain inadequate for fully automating the neuropathological assessment process. For example, NEUROINFO withBRAINMAKER by MBF SCIENCES integrates experimental data with the Allen Mouse Brain Atlas, providing regional overlays and estimating the anterior-posterior location of crosssections. However, it requires manual adjustment of the overlays to match experimental images, a process that must be repeated for each cross-section. This approach does not automate volumetric calculations or resolve the labor-intensive aspect of tracing brain region boundaries, leaving significant gaps in efficiency and reproducibility.Similarly, the Segment Anything Model (SAM) by Meta is a general-purpose image segmentation model with a very large parameter count. While broadly flexible, its computational intensity can limit practicality for high-resolution histology workflows typical in preclinical research. Running such large models increases processing time and resource usage, which can be inefficient for high-throughput volumetric analyses. These constraints underscore the value of a purpose-built approach for neuropathology.These limitations across traditional methods, domain-specific tools like NEUROINFO with BRAINMAKER, and general-purpose models like SAM underscore the need for a purpose-built solution that can integrate automated segmentation, volume estimation, and anatomical boundary refinement with improved efficiency and scalability Such a solution must overcome the inherent subjectivity and labor demands of manual methods while addressing the computational inefficiencies of existing technologies.BRIEF SUMMARY OF THE INVENTIONThe invention leverages artificial intelligence to automate and enhance the stereological volume assessment process, addressing inefficiencies and subjectivities associated with traditional manual tracing methods. Central to the invention is the application of advanced self-attention mechanisms within transformer-based models, which excel in analyzing and segmenting complex image data. Self-attention mechanisms allow the system to selectively focus on relevant image features, significantly improving segmentation accuracy The invention incorporates these mechanisms alongside algorithms derived from topological data analysis (TDA), creating a robust framework for automated volumetric brain assessment.Initially, the invention employs a convolutional neural network (CNN) to process cross-sectional images of brain tissue, identifying and isolating the region of interest. This preprocessing step reduces the computational load of subsequent stages by ensuring that only the relevant portion of each image is analyzed. Once isolated, a transformer-based Al model classifies each pixel within the region of interest, determining whether it belongs to the target region The model utilizes axial attention and patched attention techniques to refine segmentation and ensure accuracy, particularly in complex or irregularly shaped regions.To further enhance the precision and anatomical validity of the segmentation, the invention applies a TDA-based refinement algorithm. This algorithm constructs and evaluates simplicial complexes representing connected groups of pixels, ensuring that the final segmentationmaintains anatomical continuity. The refinement process addresses potential inaccuracies in the Al’s initial output, producing a highly accurate and anatomically coherent representation of the brain region.The invention is particularly focused on assessing striatal atrophy in Huntington disease (HD) model mice. The striatum is a critical region for evaluating HD progression, as it is among the first brain areas to undergo neurodegeneration. In HD, the medium spiny neurons that comprise the majority of the striatum experience significant cell death, resulting in progressive reductions in striatal volume. These changes are consistently observed in premanifest and symptomatic HD patients, as well as in animal models such as Hu97 / 18 humanized transgenic mice and Q175FDN knock-in mice. By automating the analysis of striatal volume in these models, the invention facilitates the efficient evaluation of therapeutic interventions aimed at mitigating neurodegeneration.While the initial focus is on the striatum, the invention is adaptable to other brain regions such as the frontal cortex and corpus callosum, which also exhibit atrophy in HD and other neurodegenerative diseases. Extending the methodology to additional regions involvestraining the Al models on relevant datasets and refining the segmentation algorithms to account for unique anatomical features. Beyond HD, the invention has broad applicability to a range of neurological conditions characterized by regional brain atrophy, including Alzheimer’s disease, Parkinson’s disease, and amyotrophic lateral sclerosis (ALS). The ability to accurately and efficiently quantify volumetric changes in affected brain regions provides a powerful tool for validating potential therapies across these conditions.The invention improves upon traditional and existing methodologies by providing a highly efficient and accurate approach to neuropathological assessments in preclinical research By automating the manual tracing process, the invention eliminates significant labor requirements, enabling the processing of large datasets in significantly reduced timeframes. This allows for high-throughput analysis, particularly in studies requiring volumetric assessments of brain regions across numerous samples.A key technical advantage of the invention is a parameter-efficient transformer-based segmentation module that uses fewer trainable parameters than general-purpose segmentation models while achieving the accuracy demonstrated herein. The reduced parameter count minimizes computational demands, enabling the invention to process high-resolution images more rapidly and efficiently while maintaining segmentation precision This efficiency makes the invention well-suited for large-scale studies and resource-constrained applications Additionally, the invention combines advanced self-attention mechanisms and topological data analysis (TDA) to achieve accurate segmentation and anatomical boundary refinement. These features ensure that the automated outputs are aligned with anatomical structures, improving reproducibility and reducing the variability that is inherent in manual methods. By addressing both computational and procedural inefficiencies, the invention provides a standardizedapproach to volumetric analysis, enhancing the scalability and reliability of neuropathological assessments.The invention also reduces resource consumption by lowering the reliance on manual labor and expertise, which are both costly and time-intensive. This reduction enables resources to be reallocated to other areas of research and development. Furthermore, the automated nature of the process increases the efficiency of animal model studies, reducing the number of animals required to achieve statistically significant results. Scalability is another key advantage, as the computational algorithms underpinning the invention can be applied to new datasets with minimal modifications. This scalability makes the invention suitable for widespread adoption across research institutions and studies, including large-scale preclinical trials involving multiple therapeutic candidatesFinally, the broad applicability of the invention underscores its transformative potential in neuropathological research. With appropriate training and validation, the methodologies can be adapted to assess volumetric changes in brain regions associated with a wide array of neurological diseases. By addressing the limitations of traditional stereological methods and enabling accurate, rapid, and resource-efficient volumetric analysis, the invention represents a significant advancement in the field of neuropathological assessment.The invention encompasses methods, systems, and computer-readable media for assessing volumetric changes in brain regions of interest in preclinical animal models The method involves receiving a plurality of digital images of cross-sections of brain tissue spanning the brain region of interest. These images contain structural features enabling segmentation, which are processed using a convolutional neural network (CNN) to identify and isolate the brain region of interest. A transformer-based artificial intelligence model, utilizing axial attention and patched attention mechanisms, classifies pixels as belonging to the brain region or not, thereby improving segmentation accuracy. To ensure anatomical continuity, the segmented brain region is refined using a topological data analysis (TDA) algorithm that constructs and evaluates simplicial complexes representing connected pixel groups. A three-dimensional volume of the brain region is calculated using the Cavalieri principle, and the calculated volume is outputted for display or storage in a memory.In some embodiments, preprocessing techniques are applied to the digital images, such as contrast enhancement and intensity normalization, prior to CNN processing. The transformerbased model may be trained on datasets clustered by structural similarity of brain regions, enhancing segmentation accuracy. The TDA refinement process can exclude disconnected pixel groups below a predefined size threshold, ensuring anatomically valid outputs. The Cavalieri principle calculates volumes by summing cross-sectional areas multiplied by the interval between sections, with the resulting volumes optionally displayed as graphical overlays for visualization. Additional features include the ability to store segmentation data for audit purposes or to output time-series volumetric changes for longitudinal studies.The system embodiments comprise a memory for storing digital brain tissue cross-section images and a processor configured to perform the steps of segmentation, refinement, and volumetric analysis. The processor executes the CNN for isolating brain regions and the transformer-based model for pixel classification, while applying TDA algorithms to refine segmentations. The system may also include an interface for receiving user inputs specifying the brain region of interest and displaying graphical overlays of the volumetric analysis. Further, the memory may store clustered datasets based on morphological variations, while the processor applies histogram equalization to enhance structural features or outputs additional metrics for preclinical evaluationsEmbodiments further include a computer-readable medium storing instructions that cause a processor to implement the methods described. The instructions enable preprocessing, clustering, threshold-based refinement, and volumetric calculations, as well as the storage of segmentation data and longitudinal analysis metrics. These embodiments provide a fully automated solution for volumetric brain analysis, ensuring efficient, reproducible, and anatomically accurate assessments for preclinical research.BRIEF DESCRIPTION OF THE DRAWINGSFor a fuller understanding of the invention, reference should be made to the following detailed description, taken in connection with the accompanying drawings, in which:FIG 1 shows a series of collected, stained, and imaged sections spanning the striatum for a single brain.FIG 2 is an example of target striata being segmented within the wanted cross section. FIG 3 shows target striata and brain boxed in the imaged mounted brain section.FIG 4 shows a grid classification of striatum or not within the image.FIG 5 shows a U* transformer framework according to an embodiment of the invention. FIG 6 shows a series of distinct striatum clusters.FIG 7 is a diagrammatic illustration of discrete path connectedness.FIG 8 is a series of the RCNN extracting the left striata from each imaged cross section. FIG 9 is a box and whisker plot representing the accuracy of the U* Transformer based on the cross-section number.FIG 10 is an example input and output of the U* Transformer.FIG 11 is a box and whisker plot representing the accuracy of the U* Transformer combined with the TDA algorithm based on the cross-section number.FIG 12 is an example input and output of the TDA algorithm according to an embodiment of the invention.FIG 13 shows the comparison between manual and automated NAT assessments of striatal volumes, demonstrating consistency across methods and the ability of the NAT to detect genotypic differences in Huntington disease model mice.DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTThe dataset used in the development and validation of the Neuropathology Assessment Tool (NAT) was constructed from brain tissue obtained from two established mouse models of Huntington disease (HD): the Hu97 / 18 humanized transgenic HD mice and the Q175FDN knock-in HD mice. These models were specifically selected due to their well-documented neuropathological features that closely resemble the progressive atrophy observed in human HD. Both mouse models were aged to 12 months to ensure sufficient manifestation of neuropathological changes, particularly within the striatum, a brain region central to HD pathology.Following aging, the mice were perfused intracardially. This process involved replacing blood with a fixative solution to preserve the structural integrity of the brain tissue. A 4% paraformaldehyde solution was used as the fixative, as it effectively cross-links proteins and stabilizes tissue morphology for subsequent imaging and analysis. After perfusion, the brains were post-fixed overnight in the same paraformaldehyde solution to ensure thorough fixation. Subsequent to fixation, the brains were cryoprotected by immersion in a 30% sucrose solution This step prevents the formation of ice crystals during freezing, which could otherwise damage the delicate neural structures. Once cryoprotected, the brains were sectioned into 25 pm-thick free-floating slices using a microtome. To ensure comprehensive coverage of the striatum, the sections were spaced 200 m apart. This spacing strikes a balance between capturing sufficient structural detail and minimizing the total number of sections required for volumetric analysis. The sections were stained for neuronal nuclei using a metal-enhanced diaminobenzidine (DAB) detection method. DAB staining is particularly advantageous for neuropathological studies as it provides high contrast and resolution, enabling clear visualization of cytoarchitectural features. The stained sections were then mounted onto glass slides for imaging. Imaging was performed under standardized conditions to ensure uniformity across the dataset, capturing high-resolution digital images of each stained section. These images represent a comprehensive dataset spanning the striatum for each brain, as shown in Figure 1, which illustrates a series 10 of collected, stained, and imaged sections covering the striatum for a single brain.Both the left and right striatum were manually traced from each section by experienced researchers to establish a ground truth for cross-sectional area determination This manual tracing process involves delineating the boundaries of the striatum within each image, a labor-intensive and time-consuming task that is prone to subjectivity and variability. These manuallytraced areas served as the reference data for evaluating the accuracy and efficiency of the automated processes implemented in the NAT.T o automate the tracing process, the invention utilizes image segmentation techniques. Image segmentation involves partitioning an image into distinct regions based on specific features or characteristics. For example, when applied to the cross-section of an HD mouse brain, segmentation explicitly delineates the boundaries of the striatum and other regions of interest. This is depicted in Figure 2, which shows the desired segmentation of the wanted striata 12 within a representative cross-section. By automating this process, the NAT minimizes variability and improves efficiency compared to manual tracing.The NAT leverages advancements in artificial intelligence (Al) algorithms and incorporates topological data analysis (TDA) to achieve precise and reliable segmentation. These technologies allow for accurate identification of the striatum and other relevant structures, overcoming the limitations of traditional manual methods. The dataset described herein forms the foundation for training and validating these advanced algorithms, enabling the NAT to deliver automated, high-throughput volumetric analysis of brain regions with minimal human interventionCreation of the Al ModelThe Neuropathology Assessment Tool (NAT) employs an advanced artificial intelligence (Al) model designed to accurately and efficiently perform segmentation of brain regions, particularly the striatum, in two-dimensional cross-sectional images. The NAT's Al framework consists of three core components: the Faster Region-Based Convolutional Neural Network (Faster R-CNN), a custom-developed U* Transformer, and a refinement algorithm incorporating topological data analysis. This section describes the creation of the first two components, which underpin the system's automated segmentation capabilities.Faster Region-Based Convolutional Neural Network (R-CNN)The Faster R-CNN serves as the first processing stage of the NAT. Its primary function is to "crop" or isolate the desired regions of an image, such as the striatum, before proceeding to detailed segmentation. This preprocessing step ensures that only the relevant portions of the image are analyzed, reducing computational demands and enhancing segmentation accuracy. A representative example of the Faster R-CNN's output is shown in Figure 3, where the striatum and surrounding brain area are boxed 14, 16 within an imaged brain section 18.This implementation of the Faster R-CNN addresses two primary considerations. First, it improves accuracy by limiting the segmentation algorithm's focus to the region of interest. Preliminary research revealed that segmentation algorithms achieve higher accuracy when tasked with analyzing smaller, targeted areas rather than entire brain sections. For example, if an algorithm were given an image of the entire brain, there exists a nonzero probability that a pixel within the cortex could be misclassified as part of the striatum. By cropping the image tothe striatal region, such misclassifications are effectively eliminated. Second, the Faster R-CNN contributes to the scalability of the NAT Smaller cropped images require fewer computational resources to process, allowing the system to operate more efficiently and making the technology accessible on standard computing systems.The Faster R-CNN is particularly adept at handling larger image datasets due to its design. However, the incorporation of image cropping not only enhances segmentation accuracy but also reduces computational overhead, ensuring faster and more resource-efficient evaluations. This preprocessing step lays the foundation for the NAT's downstream components, which operate on the isolated striatal regions.Advancements in Image Segmentation Accuracy Using TransformersMaximizing segmentation accuracy is a central objective of the NAT. Recent advancements in the field of Al have demonstrated the efficacy of Transformer architectures in improving accuracy through self-attention mechanisms. Self-attention enables the model to analyze all available information in an image and selectively focus on the most relevant features. Building upon this principle, the NAT initially employed an adaptation of the Visual Transformer (VIT) to implement a classification grid approach. This adaptation divides an input image into smaller grid segments and classifies each segment as either belonging to the striatum or not. A visual representation of this approach is shown in Figure 4, which illustrates the classification of image grids 24, 26 into striatum 20 (solid fill) and non-striatum 22 (hatched) regions.While the VIT adaptation demonstrated the potential of gridded classification for improving accuracy, it faced limitations in terms of performance and computational efficiency. Specifically, the integration of the ViT adaptation into the NAT framework posed challenges related to accuracy and resource demands, leading to its replacement during subsequent investigations. Despite these challenges, the concept of gridded classification informed the development of a more advanced Transformer-based model, underscoring the importance of feature isolation for accurate segmentation.Development of the U* TransformerThe U* Transformer was developed as the core artificial intelligence model for the Neuropathology Assessment Tool (NAT), addressing limitations in prior segmentation models and leveraging advanced Transformer architecture to improve segmentation accuracy and computational efficiency. A detailed representation of an embodiment of the U* Transformer framework is provided in Figure 5, which outlines the entire workflow from input image processing to final segmentation output.The process begins with the input image 30, a cross-sectional image of brain tissue stained and prepared for neuropathological analysis. The input is fed into the initial feature extraction layers 32, which are convolutional in nature. These layers identify key structural features suchas texture, intensity gradients, and patterns associated with the brain region of interest. These features are essential for guiding subsequent segmentation steps.The extracted features are then passed through the downsampling path, where the spatial dimensions are progressively reduced while retaining semantically relevant information. This stage is computationally efficient and essential for isolating high-level features critical for accurate segmentation. Skip connections 34 link the earlier high-resolution feature maps from the downsampling path to corresponding layers in the up-sampling path, ensuring that finegrained details lost during downsampling are reintroduced later in the process This integration allows the model to maintain anatomical precision while benefiting from the computational efficiency of a hierarchical feature extraction approachThe core of the U* Transformer architecture incorporates advanced self-attention mechanisms, including axial attention and patched attention. Axial attention focuses computational resources on individual axes, such as rows or columns, enabling the model to analyze spatial relationships efficiently without processing the entire image simultaneously. This feature reduces computational complexity, making the U* Transformer well-suited for high-resolution input images Patched attention, on the other hand, subdivides the image into smaller patches, allowing the model to examine localized regions in greater depth. This approach enhances the model’s ability to identify complex boundaries, such as those delineating the striatum from surrounding tissues.Once the features have been processed through the self-attention mechanisms, the up-sampling path begins. This phase, depicted in Figure 5 as layers following the skip connections, progressively reconstructs the spatial dimensions of the image while integrating both high-resolution details from the skip connections and abstract features from the Transformer layers. This ensures that the segmentation output is both anatomically accurate and computationally efficient.The output layer of the U* Transformer classifies each pixel in the processed image, resulting in a grid 40 where pixels are labeled as either striatum 20 (darker shade) or not striatum 22 (lighter shade) This grid provides a clear representation of the segmented region, which is then overlaid onto the original image for validation and further analysis. The final segmented output 42 highlights the striatum as a distinct region, ready for volumetric calculations using the Cavalieri principle.The design of the U* Transformer improves upon prior architectures like the Visual Transformer (ViT) by addressing their limitations in handling high-resolution images and complex anatomical features. The inclusion of axial attention allows the model to efficiently process large datasets without requiring excessive computational resources, while patched attention ensures precise analysis of smaller, localized regions. These innovations, combined with the hierarchical feature integration enabled by skip connections, result in a segmentation model that is both efficient and robust, capable of handling the variability inherent in neuropathological datasets. Fig. 5effectively illustrates these advancements, showcasing the streamlined workflow and the integration of key components within the U* T ransformer framework.The U* Transformer represents a significant advancement in the NAT's Al framework. By combining these advanced attention mechanisms with the insights gained from prior research, the model achieves precise segmentation of the striatum while maintaining computational efficiency. This development marks a key step in the automation of stereological volume assessments, enabling accurate, high-throughput analysis of brain regions with minimal human intervention.Training the U* TransformerThe training of the U* Transformer, the core artificial intelligence (Al) model within the Neuropathology Assessment Tool (NAT), was an essential phase in its development. This process ensured that the model could accurately segment the striatum in two-dimensional cross-sectional images of mouse brains. Training involved the use of two datasets derived from established Huntington disease (HD) mouse models, pre-processing the images for optimal feature extraction, and iteratively refining the model's parameters to maximize segmentation accuracy.Dataset Preparation for Training and ValidationThe primary dataset used to train the U* Transformer was generated from the Hu97 / 18 humanized transgenic HD mouse model, while a secondary dataset derived from Q175FDN knock-in HD mice was reserved for validation. The use of distinct datasets for training and validation ensured that the model's performance was evaluated on images it had not encountered before, providing an unbiased measure of its accuracy.During training, the U* T ransformer underwent continuous refinement of its parameters through iterative learning. This process allowed the model to adaptively adjust its functions, thereby improving its ability to recognize and trace the striatum. To further validate its accuracy, the validation dataset was introduced after training was completed. This dataset consisted of brain images the Al had not previously processed, ensuring that the model's accuracy reflected its generalizability across unseen dataComputational Resources and Model RefinementDespite the advanced design of the U* Transformer, the computational demands of training initially exceeded the capacity of standard individual computers. To address this, training was conducted using the supercomputer at the University of Central Florida (UCF) Newton Advanced Research Computing Center. The availability of this high-performance computing resource enabled the training process to be completed in only a few days This acceleration facilitated rapid iterative improvements to the U* Transformer based on observations during trainingThe use of the supercomputer also enabled the handling of large and complex datasets, which were integral to the training process. The computational power provided by the UCF Newton Center allowed for high-resolution image processing and efficient parameter optimization, ensuring that the U* Transformer was properly trained to achieve its intended function.Image Pre-Processing and ClusteringPre-processing of the images was another essential aspect of training. The striatum exhibits three distinct morphological shapes, which vary depending on the cross-sectional position within the brain. These shapes were observed to transition from a rib-eye appearance to an oval shape and, finally, to a curved kidney bean shape To account for these variations, the dataset was divided into three distinct subsets based on the shape of the striatum. Each subset was used to train a separate instance of the U* Transformer, enabling the model to specialize in recognizing each morphological category. The clustering process was accomplished using the Structural Similarity Index Matrix (SSIM), a metric that quantifies the similarity between images.Once the datasets were clustered, additional pre-processing steps were applied to enhance image quality. Histogram equalization was performed to standardize the contrast of the images, redistributing intensity values to improve the visibility of structural features. This technique is particularly useful for microscopy-stained images, where variations in staining and imaging conditions can obscure important details. By enhancing contrast, histogram equalization ensured consistent feature extraction across all images.The validation set was also clustered using SSIM to align with the clusters in the training dataset. This step ensured that validation images matched the morphological categories encountered during training, enabling a more precise evaluation of the model's performance. The clustered validation set was then used to validate the segmentation accuracy of each U* Transformer instance trained on the corresponding subset of the dataset Segmentation Validation and OptimizationThe training and validation process demonstrated that the U* Transformer could reliably and accurately segment the striatum across a variety of shapes and imaging conditions. By clustering the images and training specialized models for each cluster, the NAT achieved enhanced segmentation performance. The combination of advanced training methodologies, computational resources, and image pre-processing ensured that the U* Transformer was effectively optimized for its intended application.The results of this training phase established the foundation for the integration of the U* T ransformer into the NAT framework, enabling high-accuracy automated segmentation of brain regions in preclinical models of neurodegenerative disease. A representative illustration of the distinct striatum clusters used for training is provided in Figure 6 reference numerals 44, 46 and 48.Refinement of the Neuropathology Assessment Tool (NAT)The refinement stage of the Neuropathology Assessment Tool (NAT) employs a robust algorithm based on topological data analysis (TDA). TDA provides a mathematical framework for analyzing complex, high-dimensional data by studying its intrinsic geometric and topological structure. Unlike traditional Al methods, which often rely on explicit feature extraction or statistical correlations, TDA captures relationships between data points to identify patterns and structures, particularly in cases where the geometry and connectivity of the data play a significant role This makes TDA highly effective for refining image segmentation results, where anatomical continuity and accuracy are paramount.Data Analysis and Metric Space RepresentationAt the core of the refinement process is the concept of a metric space. A metric space (M,d), such that M is a set with a function d: M x M > R+, the distance function such that for any a,b,c e M, the following properties are adhered to:d(a,b) > 0 where d(a, b~) = 0 <=> a = bd a, b) = d(b, a)d a,c) < d(a,b) + d(b,c)When defining a metric space (M,d) upon a segmented image in 2D space with an image of size H x W H, W <= N where M H x W, such that V peM, p =x e H,y e W, the distance function can be defined asV p1(p2eM, such that pt= xlly1),p1= x2,y2)', d(p1,p2) = √(x2- x1)2+ (y2- y1)2Using this, a form of discrete path connectedness can be used to define two connected spaces Specifically, in the context of images, subsets can be defined as connected based on either 4-connectivity or 8-connectivityThis means that a subset I £ M is connected if Vpl,p2 e 1, 3{pl,p3,...,pn,p2} £ I such that d(pi,pi + 1) < V2 for i = 1,...,n - 1. Figure 7A shows all elements of M such that each pixel in white represents an element in M based on its (x,y) positioning in the image. When using discrete path connectedness, there are two discrete path connected subsets of M represented by different shading (Figure 7B).These simplices are used to construct simplicial complexes, collections of interconnected simplices that represent the spatial relationships between data points in the segmented image. The simplicial complexes generated in the NAT capture the geometry and connectivity of the striatum, enabling precise refinement of the segmentation.The distance formula is a fundamental tool in defining simplices, which are basic building blocks in TDA. Using the same distance formula for metric space (M, d) defined above, a simplex can be defined by forming connections between points that are within a specified radius r. For example, a 0-simplex corresponds to a single point, a 1 -simplex is formed by connecting twopoints that are within the distance r, a 2-simplex is created by connecting three points to form a triangle if all pairwise distances are less than r, and higher-dimensional simplices are formed similarly. This approach allows for the construction of simplicial complexes, which are collections of simplices that capture the spatial relationships between data points in a way that reflects their underlying topological structure.Using these concepts from TDA, a final refinement algorithm was created for NAT. Again, the anatomy of the striatum is heavily considered. The striatum is a singular piece initiated by finding all discrete connected spaces, using the same method as described above, followed by only processing the largest one. This ensures that the final output is as close to anatomically correct as possible. From there, each pixel that is classified as striatum is treated as its own distinct data point, and simplicial complexes are formed around each one. The radius for the simplicial complexes stops increasing once all points are connected by a simplex From there, each actual simplex is filled in. Anything that is within the area outlined by the U* Transformer, and is within the main discrete connected area, is filled in.Results of the Neuropathology Assessment Tool (NAT)The performance of the Neuropathology Assessment Tool (NAT) was evaluated using several key endpoints to measure its accuracy, efficiency, and ability to detect neuropathological changes in preclinical models of Huntington disease (HD). These endpoints were assessed across multiple stages of the NAT pipeline, including region extraction by the Faster R-CNN, pixel-wise segmentation by the U* Transformer, and refinement through the TDA-based algorithmEvaluation of the Faster R-CNN for Region ExtractionThe first endpoint involved determining the ability of the Faster Region-Based Convolutional Neural Network (R-CNN) to accurately extract the striatum from cross-sectional brain images. The Faster R-CNN processed each image by isolating and cropping the striatal region to reduce computational complexity and improve segmentation accuracy. Quantitative analysis revealed an average accuracy of 99.7% for the Faster R-CNN in extracting the striatum, as visually confirmed by overlaying the extracted boundaries on the original images. A representative example of the extracted left striatum from an image cross-section is shown in Figure 8 as lightbounding boxesPerformance of the U* Transformer for SegmentationFollowing region extraction, the U* Transformer was evaluated using a pixel-by-pixel binary accuracy metric. This metric assessed whether each pixel in the extracted region was correctly classified as part of the striatum or not. To ensure robustness, the validation dataset included 1,369 images that had not been used during the training phase. The U* T ransformer achieved an average accuracy of 93.6% across the validation set. A box-and-whisker plot illustrating segmentation accuracy across cross-sections is shown in Figure 9, while Figure 10 providesan example of the input and output images from the U* Transformer, demonstrating the quality of the segmentation.Refinement Algorithm PerformanceThe refinement algorithm, incorporating topological data analysis (TDA), was then applied to improve the accuracy of the segmentation produced by the U* Transformer. The algorithm enhanced anatomical continuity and eliminated segmentation artifacts. After applying the refinement algorithm, the accuracy of the segmented images increased to an average of 93.8%, as shown in the updated box-and-whisker plot in Figure 11. Visual comparisons before and after refinement highlighted a noticeable improvement in segmentation quality, as depicted in Figure 12.Assessment of NAT in Measuring Striatal VolumesThe final evaluation focused on the NAT's ability to measure striatal volumes and detect genotypic differences in preclinical HD models. The tool was tested on Q175FDN heterozygous (HET) HD mice and wild-type (WT) littermate controls, with results compared to manual assessments of striatal volume. Statistical analysis demonstrated no significant difference between NAT-derived and manually measured striatal volumes (p = 0.28), as shown in Figure 13. A. The overall difference between manual and NAT measurements was only 2.7%, indicating a strong agreement between the two methodsIn line with historical data, the NAT detected a significant reduction in striatal volume in Q175FDN HET mice compared toWT controls, with p-values of 6E-6 for manual measurements and 9E-6 for NAT measurements, as illustrated in Figure 13. B. The manual method identified an 11 24% atrophy, while the NAT reported 10.37% atrophy, further validating its accuracy. Comparisons within genotype groups (HD to HD and WT to WT) revealed no significant differences (p = 0.46 and p = 0.28, respectively)A simple linear regression model was used to evaluate the relationship between manual and NAT-generated volumetric data. The model produced an R-squared value of 0.945, indicating that 945% of the variance in NAT measurements could be explained by manual tracings The model was statistically significant (p = 43E-35), with a slope confidence interval of [0.98, 1.12] and an intercept confidence interval of [-0.63, -0.43], further confirming the consistency between the two methods. These results are summarized in Figure 13. C.Summary of ResultsThe results confirm that the NAT achieves high levels of accuracy, efficiency, and reliability across all stages of its pipeline. The Faster R-CNN effectively isolates the striatal region, while the U* Transformer provides robust segmentation that is further refined by the TDA-based algorithm. The NAT's ability to measure striatal volumes and detect genotypic differences aligns closely with manual assessments, offering a faster and less resource-intensive alternative forneuropathological analysis. These findings demonstrate the NAT's potential as a valuable tool for preclinical studies in neurodegenerative diseases.The integration of the Faster R-CNN, U* Transformer, and TDA-based refinement algorithm forms the Neuropathology Assessment Tool (NAT), which is specifically designed for evaluating neuropathology in preclinical models. The NAT accurately assesses striatal volumes and detects neuropathological changes in Q175FDN HD model mice. Analyses demonstrate a high degree of agreement between NAT-derived and manual volumetric measurements, with no significant differences and an overall variation of only 2.7%. The NAT also detects genotypespecific differences in striatal atrophy, including a significant distinction in striatal volumes between Q175FDN HET and WT mice, aligning with established data.The NAT offers efficiency advantages over traditional manual methods, reducing the time and resources required for volumetric analysis while maintaining precision and reliability. Linear regression analysis of NAT and manual data shows a strong correlation, with an R-squared value of 0.945, indicating that 94.5% of the variance in manual tracing-derived volumes is explained by NAT measurements. The results underscore the accuracy and robustness of the NAT as a reliable alternative to manual stereological methodsThe NAT enables reliable assessment of neuropathological endpoints, such as striatal atrophy, which are critical in preclinical trials. By automating volumetric analysis and maintaining consistency across datasets, the NAT supports preclinical evaluations with greater efficiency and accuracy. These capabilities streamline the identification and evaluation of experimental therapies, contributing to advancements in the study and treatment of neurodegenerative diseases such as Huntington disease.REFERENCE NUMERALSNo. Element Description Figure(s) Context / Location in Text 10 Series of collected, stained, and Fig 1 Dataset showing multiple imaged brain sections spanning the histological sections used for striatum for a single brain volumetric analysis12 Segmented region corresponding to Fig. 2 Area delineated in segmentation the target striatum example image14 Left striatum region (ROI) identified Fig. 3 Boxed area showing left-side by Faster R-CNN striatum16 Right striatum region (ROI) identified Fig. 3 Boxed area showing right-side by Faster R-CNN striatum18 Entire imaged mounted brain section Fig. 3 Background brain section from which striatal ROIs are cropped 20 Pixel region classified as “striatum” Fig 4, 5 Black-fill cells or pixels labeled (positive class) as belonging to striatum22 Pixel region classified as “not Fig. 4, 5 Light-shaded cells or pixels striatum” (negative class) outside striatum region24 Input grid for classification prior to Fig. 4 Left panel showing unclassified Transformer operation grid segments26 Output grid showing classified regions Fig. 4 Right panel depicting labeledafter T ransformer operation striatum vs non-striatum cells30 Input image to U* Transformer Fig. 5 Starting image entering convolutional feature extractor 32 Initial convolutional feature extraction Fig. 5 CNN layers that identify layers textures, gradients, and edges 34 Skip connection between encoder Fig 5 Re-introduces high-resolution and decoder feature maps features into up-sampling path 36 Intermediate up-sampling / decoding Fig. 5 Part of reconstruction path layer combining encoded and skip- connected features38 Output reconstruction block prior to Fig. 5 Produces final spatially-restored classification feature map40 Pixel classification grid (not striatum = Fig. 5 Visualization of pixel-wise hatched) prediction results42 Final segmented output overlaid on Fig. 5 Depicts the U* Transformer’s original image segmentation result for visualization44 Cluster A striatum morphology (ribFig. 6 First morphological cluster in eye shape) SSIM-based training set 46 Cluster B striatum morphology (oval Fig. 6 Second morphological cluster shape)48 Cluster C striatum morphology Fig. 6 Third morphological cluster(kidney-bean shape) representing posterior sectionsCOMPUTER AND SOFTWARE TECHNOLOGYThe present invention may be embodied on various platforms. The following provides an antecedent basis for the information technology that may be utilized to enable the invention. Embodiments of the present invention may be implemented in hardware, firmware, software, or any combination thereof. Embodiments of the present invention may also be implemented as instructions stored on a machine-readable medium, which may be read and executed by one or more processors. A machine-readable medium may include any mechanism for storing or transmitting information in a form readable by a machine (e.g., a computing device). For example, a machine-readable medium may include read only memory (ROM); random access memory (RAM); magnetic disk storage media; optical storage media; flash memory devices; electrical, optical, acoustical or other forms of propagated signals (e.g., carrier waves, infrared signals, digital signals, etc.), and othersFurther, firmware, software, routines, instructions may be described herein as performing certain actions. However, it should be appreciated that such descriptions are merely for convenience and that such actions in fact result from computing devices, processors, controllers, or other devices executing the firmware, software, routines, instructions, etc. The machine-readable medium may be, for example, but not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer readable storage medium would include the following: an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM),a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In the context of this document, a computer readable storage medium may be any non-transitory, tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device. Storage and services may be on premise or remote such as in the “cloud” through vendors operating under the brands, MICROSOFT AZURE, AMAZON WEB SERVICES, RACKSPACE, and KAMATERA.A machine-readable signal medium may include a propagated data signal with machine-readable program code embodied therein, for example, in baseband or as part of a carrier wave. Such a propagated signal may take any of a variety of forms, including, but not limited to, electro-magnetic, optical, or any suitable combination thereof. A machine-readable signal medium may be any machine-readable medium that is not a computer readable storage medium and that can communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device. However, as indicated above, due to circuit statutory subject matter restrictions, claims to this invention as a software product are those embodied in a non-transitory software medium such as a computer hard drive, flash-RAM, optical disk or the like.Program code embodied on a machine-readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wire-line, optical fiber cable, radio frequency, etc, or any suitable combination of the foregoing. Machine-readable program code for carrying out operations for aspects of the present invention may be written in any combination of one or more programming languages, including an object-oriented programming language such as Java, C#, C++, Visual Basic or the like and conventional procedural programming languages, such as the “C” programming language or similar programming languages. Additional languages may include scripting languages such as PYTHON, LUA and PERLAspects of the present invention are described below with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by machine-readable program instructions. GLOSSARY OF CLAIM TERMSAxial attention means a self-attention mechanism in transformer-based models that processes image data one axis at a time, rather than all at once. Instead of attending to every pixel simultaneously, the model sequentially focuses along each dimension of a two-dimensional image (for example, analyzing all rows, then all columns). This partitioned approach significantly reduces computational complexity, transforming a large segmentation task intomanageable one-dimensional passes. In the context of the Neuropathology Assessment Tool (NAT), axial attention allows the system’s transformer model to efficiently examine spatial patterns along horizontal and vertical directions within brain slice images. By isolating attention to one axis, the model captures linear anatomical features and boundary continuities with high precision. This focused scrutiny along each axis not only enhances segmentation accuracy but also curtails processing time and memory usage — a critical advantage when handling high-resolution histological images in preclinical research The axial attention mechanism thereby ensures that relevant structural details of a brain region are extracted methodically and integrated into a cohesive segmentation of the region of interest.Brain region of interest means a specific anatomical area of the brain selected for focused analysis in a preclinical study due to its relevance to a disease or treatment. It is the target area that the Neuropathology Assessment Tool (NAT) is designed to isolate and measure. For example, in Huntington disease research the striatum is a common region of interest because it undergoes early neurodegeneration. Identifying a brain region of interest involves using imaging and segmentation techniques to delineate that region’s boundaries on a series of brain cross-sections. Once isolated, this region can be quantitatively analyzed — for instance, its volume can be calculated over multiple slices to gauge atrophy or growth. In the context of this invention, the brain region of interest defines the scope of automated analysis: all subsequent image processing steps, from initial detection by a convolutional neural network to pixel-wise classification by the transformer model, are concentrated on this anatomically defined area to assess pathological changes or treatment effects.Cavalieri principle means a stereological method for estimating the volume of a three-dimensional object by summing the areas of its two-dimensional cross-sections, multiplied by the distance between those sections. Applied in this invention, the Cavalieri principle provides a rigorous way to calculate brain region volumes from a stack of histological slices. After the Neuropathology Assessment Tool (NAT) segments the region of interest in each cross-sectional image, the area of that segmented region in each slice is determined. These areas are then aggregated: each area is multiplied by the known interval (thickness or spacing) between consecutive slices, and the results are summed to yield an approximate total volume. This mathematical approach ensures that volume measurements are objective and reproducible. The principle assumes slices are evenly spaced and parallel, conditions typically met in preclinical histological studies. By leveraging segmented area data generated by NAT’s algorithms, the Cavalieri principle enables the invention to produce accurate volumetric assessments of brain structures, bridging two-dimensional image analysis with three-dimensional quantitative outcomes.Connected pixel groups mean clusters of pixels in a segmented image that are contiguous according to a specified connectivity rule. Pixels are considered connected when they directly touch one another either along edges (four-neighborhood connectivity) or at corners (eight-neighborhood connectivity). Such a group represents a continuous region in the binary segmentation output. In this invention, connected pixel groups correspond to portions of the brain region identified by the Neuropathology Assessment T ool (NAT). An anatomically correct segmentation of a structure like the striatum should form one large connected pixel group rather than fragmented pieces. The NAT’s refinement algorithm evaluates these groups to ensure the segmented region of interest is one coherent cluster. Any segmentation output is scanned for connectivity: the principal connected group (the largest cluster matching the expected region) is preserved, while any small isolated groups (which may arise from noise or misclassification) are discarded This practice maintains anatomical continuity, guaranteeing that the final segmented region aligns with the continuous, singular nature of the actual brain structure under analysis.Convolutional neural network (CNN) means a type of deep learning model specialized for analyzing visual data through layered convolution operations. A CNN automatically learns hierarchical features from images (for instance, detecting edges, shapes, and complex patterns) by applying filters that slide over the input. In this invention, a CNN is employed as an initial processing stage to handle raw cross-sectional brain images. The CNN is trained to recognize and extract the brain region of interest (such as the striatum) from the larger brain slice image based on distinctive structural features. By scanning the image, the CNN produces a localized output — often a bounding box or mask — that isolates the target region from surrounding tissues. This focused isolation serves as a preprocessing step that significantly reduces the computational load for subsequent steps. By confining attention to a smaller area, the invention’s transformer-based model can operate more efficiently. Thus, the CNN component of the Neuropathology Assessment Tool (NAT) lays the groundwork for accurate segmentation by quickly and reliably identifying the relevant anatomical region in each image.Cross-sections of brain tissue mean thin, planar slices of brain obtained from animal models for histological examination. Typically produced by cutting the brain at regular intervals (e.g., using a microtome), each cross-section preserves the anatomical details present at that specific level. These slices are often mounted on slides and stained to highlight cellular structures or regional boundaries. In this invention, such cross-sections are imaged to create high-resolution digital images that serve as input data for analysis By spanning the entire brain region of interest at consistent sectioning intervals, the collection of cross-sections provides a basis for volumetric measurement The Neuropathology Assessment Tool (NAT) processes each cross-sectional image to segment the target region (for example, delineating the striatum in each slice). With a complete series of segmented cross-sections, the invention can apply the Cavalieri principle to estimate the three-dimensional volume of the region. Thus, standardized cross-sections of brain tissue form the foundational dataset that the automated system analyzes to quantify neuropathological changesDigital images mean electronic visual representations of brain cross-sections captured through imaging devices such as microscopes or scanners. These images are composed of pixels that encode intensity or color information corresponding to microscopic anatomy of the tissue In the context of this invention, digital images are the direct input to the Neuropathology Assessment Tool (NAT). They portray the stained brain sections with all relevant structural detail needed for computational analysis Prior to analysis, images may undergo standardization steps like contrast enhancement and intensity normalization (for example, via histogram equalization) to ensure consistent quality across the dataset. The NAT’s algorithms then operate on these images to identify and segment the region of interest. Because digital images can be stored, processed, and transmitted easily, they enable the automation pipeline to function efficiently. The fidelity of digital images — preserving fine features such as cellular patterns or boundaries between brain regions — is crucial, as the accuracy of segmentation and volume estimation directly depends on the quality of the input imagery.Disconnected pixel groups mean isolated clusters of pixels in a segmented image that do not form part of the main contiguous region of interest. These small clusters fail to meet the defined connectivity criteria (for instance, they might be separated from the main region by more than one pixel gap) and are usually considered segmentation artifacts or noise. In this invention, the occurrence of disconnected pixel groups typically indicates spurious misclassifications — such as tiny spots erroneously labeled as the target region. The Neuropathology Assessment Tool (NAT) addresses this by applying a refinement step grounded in topological analysis: after the initial segmentation, any pixel group below a certain size or isolated from the principal connected region is identified and removed. By excluding these disconnected fragments, the final segmentation output is made to reflect a single, continuous anatomical structure, as expected in reality. This practice prevents minor errors from skewing volume calculations and ensures that the segmented output remains anatomically faithful to the actual brain region of interest.Faster Region-Based Convolutional Neural Network (Faster R-CNN) means an advanced CNN architecture tailored for rapid object detection and region proposal in images. It combines a deep convolutional network with a Region Proposal Network to swiftly identify regions of interest and then refine their boundaries In this invention, a trained Faster R-CNN serves as the first module in the Neuropathology Assessment Tool (NAT) pipeline. Its role is to scan each cross-sectional brain image and “crop out” the specific brain region of interest (such as the striatum) from the larger image. By doing so, the model produces a tightly bounded sub-image that contains predominantly the target region and minimal surrounding tissue. This targeted extraction yields two major benefits: it ensures that subsequent segmentation by the transformer model focuses only on relevant anatomy, and it substantially reduces computational load by shrinking the image area to be processed The Faster R-CNN’s high efficiency and accuracy in isolating the region of interest (achieving near-perfect extraction intesting) lay a reliable foundation for the downstream pixel-level segmentation and volume analysis.Graphical overlays mean visual markers or annotations superimposed on an image to highlight specific features or results. In the context of this invention, graphical overlays are used to present the outcomes of the automated analysis directly on the original brain section images. For example, after the Neuropathology Assessment Tool (NAT) segments a brain region and calculates its volume, the system can draw the outline of the segmented region on the image in a distinct color. It may also overlay text or graphical indicators showing quantitative results, such as the measured volume or percentage of atrophy. These overlays allow researchers to easily verify and interpret the Al’s results by seeing exactly which areas were identified as the region of interest and how extensive that region is. Importantly, overlays do not alter the underlying image data; they are a visual aid, often implemented in software, that can be toggled on or off. By providing immediate visual validation of segmentation accuracy and volumetric calculations, graphical overlays enhance the transparency and usability of the NAT’s output in a laboratory setting.Histogram equalization means an image preprocessing technique that improves global contrast by redistributing the frequency of pixel intensity values. Essentially, it spreads out the most frequent intensity levels in an image, making dark areas lighter and bright areas darker as needed to achieve a more uniform intensity distribution. In this invention, histogram equalization is applied to images of stained brain tissue sections before segmentation takes place. The process enhances the visibility of subtle structural details — such as boundaries between gray and white matter or variations in cell density — that might otherwise be obscured by uneven staining or lighting conditions By normalizing the contrast across all images in the dataset, the Neuropathology Assessment Tool (NAT) ensures that its convolutional neural networks and transformer models receive input data with consistent quality. This consistency leads to more reliable feature extraction and segmentation, as the algorithms can detect anatomical structures without bias from variable image exposure or staining intensity. Overall, histogram equalization is a crucial preparatory step that standardizes images for optimal analysis by the automated system.Hu97 / 18 humanized transgenic HD mice mean a specialized mouse model engineered to carry human genes associated with Huntington disease (HD), used for studying the disease’s pathology in a laboratory setting. In these mice, a mutant human huntingtin gene (with pathogenic CAG repeat expansions) is introduced into the mouse genome, causing the animals to develop neuropathological features that closely resemble human HD Notably, Hu97 / 18 mice exhibit progressive neuronal degeneration and brain changes, including the characteristic shrinkage of the striatum. In this invention, data from Hu97 / 18 mice were instrumental in developing and training the Neuropathology Assessment Tool (NAT). High-resolution images of brain sections from these mice, with their striatal region delineated (either manually orthrough initial algorithms), were used to teach the Al models how to recognize and segment the region of interest. The humanized aspect of this model ensures that the patterns learned by NAT (such as tissue texture and degeneration patterns) are relevant to human disease By validating the tool on Hu97 / 18 mice, the inventors demonstrated that NAT can accurately detect and measure HD-related neuropathology in a preclinical modelLinear regression means a statistical modeling technique used to describe the relationship between a dependent variable and one or more independent variables by fitting a linear equation to observed data. In the context of this invention, linear regression is employed as a validation tool to compare volumetric measurements produced by the Neuropathology Assessment Tool (NAT) against those obtained via traditional manual methods. Specifically, after NAT automatically computes the volume of a brain region (such as the striatum) for a set of samples, those values are plotted against the corresponding volumes measured by human experts tracing the same regions. A linear regression analysis then assesses how well the two sets of measurements agree. Key outcomes include the slope of the best-fit line (ideally close to 1 0 if NAT’s results match manual results in scale) and the correlation coefficient (indicating the strength of agreement). A strong linear relationship with minimal deviation demonstrates that NAT’s automated approach is consistent with established manual measurements. This statistical confirmation is crucial for verifying the tool's accuracy and reliability in a quantitative manner.Longitudinal preclinical studies mean research investigations carried out in animal models over an extended period, with multiple time points of data collection to observe how a disease progresses or how a treatment affects that progression. Such studies might, for example, track the same cohort of animals as they age or as they undergo a therapeutic intervention, measuring relevant biological markers at various intervals. In this invention, the Neuropathology Assessment Tool (NAT) greatly facilitates longitudinal studies by providing a fast and repeatable way to quantify changes in brain anatomy, like regional volume, at each time point. Because the analysis is automated and consistent, the volumes of a brain region (such as the striatum) measured at different times can be compared reliably without the variability that comes from different people or different sessions of manual tracing. This consistency is critical when assessing disease trajectories (like progressive striatal atrophy in Huntington disease models) or evaluating whether a drug slows or reverses such atrophy. By enabling high-throughput and standardized volumetric assessments, NAT strengthens the rigor and efficiency of longitudinal preclinical research.Metric space means a fundamental mathematical framework in which distances between all pairs of points in a set are defined according to specific rules (satisfying non-negativity, identity of indiscernibles, symmetry, and triangle inequality). In simpler terms, it provides a way to quantify how far apart any two points are. In this invention, a metric space is established for the pixels in a two-dimensional segmented image of brain tissue. Each pixel’s location can beconsidered a point in this space (with coordinates, for example, given by its row and column indices), and a standard distance function such as the Euclidean distance is used to measure separation between pixels. By treating the segmented region as a collection of points in a metric space, the Neuropathology Assessment Tool (NAT) can apply topological data analysis techniques: it determines which pixels are “neighbors” or within a certain radius of each other, and it identifies connected clusters and potential gaps. This rigorous spatial representation underpins the construction of simplicial complexes in the refinement process, ensuring that the segmentation’s spatial relationships mirror the true anatomical layout of the tissue Morphological variations mean differences in shape, size, or structural appearance of a particular anatomical region observed across a set of images. Within a brain region like the striatum, morphology can vary due to several factors: the position of the cross-section (an anterior slice might show the region in a different shape than a posterior slice), individual anatomical differences between subjects, or changes induced by disease progression. In this invention, recognizing and accounting for morphological variations is essential to achieving robust segmentation. The dataset of brain images is therefore analyzed and grouped according to similar shape characteristics of the region of interest. For example, images might be clustered so that those with a “round” striatal profile are separated from those with a more “oblong” profile. By clustering the data based on these morphological patterns — often using metrics like the Structural Similarity Index to quantify shape resemblance — the Neuropathology Assessment Tool (NAT) can train specialized models or adjust its algorithms for each cluster. This stratification ensures that the Al is fine-tuned to handle the spectrum of anatomical variation, thereby improving overall segmentation accuracy and reliability across diverse samples and slice levels.Neuropathology Assessment Tool (NAT) means the comprehensive system of hardware and software modules described in this invention for automating neuropathological volume analysis. It integrates multiple advanced techniques — image processing, artificial intelligence, and topological algorithms — into a single workflow that can take raw brain section images and produce quantitative volumetric data. Specifically, the NAT encompasses: (1) a convolutional neural network stage to detect and isolate a brain region of interest from each cross-sectional image; (2) a custom transformer-based segmentation model (the U* T ransformer) which applies axial and patched attention to classify each pixel within that isolated region as belonging to the target structure or not; (3) a topological data analysis refinement step that constructs simplicial complexes from the segmented output to enforce anatomical continuity and eliminate artifacts; and (4) a volumetric calculation engine that applies the Cavalieri principle to the refined segmentation across all slices. The NAT is designed for use in preclinical research (such as with mouse models of neurodegenerative disease) to objectively measure pathological changes like regional atrophy By replacing manual tracing with high-speed, reproducible computations, the Neuropathology Assessment Tool accelerates and standardizes the assessment of brain tissue changes related to disease or therapeutic intervention.Patched attention means a variant of the self-attention mechanism in which an image is partitioned into smaller sub-regions (patches) that are analyzed in a localized fashion. Instead of computing attention weights over an entire large image at once, the transformer model processes the image as a collection of patches, focusing on one patch (and its immediate context) at a time. This approach permits the model to capture fine-grained details within each patch while still understanding the broader pattern by aggregating information across patches. In this invention, patched attention is employed by the Neuropathology Assessment Tool’s transformer model to improve segmentation performance on complex brain images. By confining the attention scope to patch-level regions (for instance, dividing a histological image into a grid of smaller tiles), the model can better distinguish subtle features that define the boundary of a brain region of interest. It reduces the dilution of important local signals that might occur if the entire image were treated uniformly. Ultimately, patched attention complements axial attention by ensuring that both local details and global structure are accounted for, resulting in a more accurate classification of pixels in challenging or intricate portions of the image.Pixels mean the individual tiny picture elements that make up a digital image, each representing a single point of the image with a specific color or intensity value. In a microscopic image of brain tissue, a pixel corresponds to a very small area of the tissue slice, reflecting details such as the presence of stained cells or fibers at that location. In this invention, pixels are the fundamental units that the Neuropathology Assessment Tool (NAT) evaluates during segmentation. The transformer-based model examines pixels and classifies each one as either part of the brain region of interest or not, effectively creating a binary mask of the region. Because volume calculations are ultimately derived from counting how many pixels belong to the region (and knowing the physical size each pixel represents), the accuracy of classifying pixels directly affects the precision of the volumetric outcome. Every processing step — from image preprocessing to CNN localization and transformer segmentation — is geared toward making the correct determination at the pixel level. Thus, pixels serve as the atomic level of information in the NAT, and handling them correctly is essential for faithful reconstruction of the region’s structure and volume.Preclinical evaluation means the stage of research in which potential therapies or interventions are tested in laboratory models (such as cell cultures or animal models) before any trials in humans. The goal of preclinical evaluation is to gather evidence of efficacy and safety that can justify moving a therapy into clinical trials. In the context of this invention, preclinical evaluation specifically refers to assessing experimental treatments using animal models of neurological disease by measuring their impact on brain pathology. The Neuropathology Assessment Tool (NAT) assists in this process by providing a means to quantitatively measure changes in brain region volumes, which is a key indicator of neuropathological progression or improvement. For example, if a candidate drug aims to slow neurodegeneration, NAT can automatically determine whether treated animals show lessstriatal atrophy over time compared to controls. By automating volumetric analysis, NAT increases throughput and consistency, allowing researchers to evaluate more subjects and conditions than would be feasible with manual methods. This yields robust data on how an intervention affects disease-related brain changes prior to human testing, thereby strengthening the preclinical evidence base.Q175FDN knock-in HD mice mean a genetically modified mouse line that carries a pathogenic mutation in its own huntingtin gene (knock-in of an expanded CAG repeat analogous to the human Huntington disease mutation). Unlike transgenic models that introduce separate genetic constructs, knock-in models like Q175FDN incorporate the mutation into the native gene locus, resulting in a disease progression that closely mimics the genetic and pathological context of human HD. Q175FDN mice develop hallmark features of Huntington disease overtime, notably including progressive striatal atrophy (loss of volume in the striatum) along with behavioral deficits. In this invention, Q175FDN mice were used as a critical testbed to validate the accuracy and utility of the Neuropathology Assessment Tool (NAT). The NAT processed brain images from these mice to automatically segment the striatum and measure its volume. The tool’s ability to detect known genotypic differences (for instance, distinguishing volumes between mutant and wild-type or treated and untreated groups) in the Q175FDN model demonstrated its sensitivity to real disease-related changes. Thus, this mouse model provided a rigorous preclinical scenario in which to prove that the NAT could reliably measure neuropathology consistent with established biological expectations.Segmentation means the process of partitioning an image into meaningful regions by classifying each pixel or group of pixels based on visual characteristics. In image analysis, segmentation results in a delineation of structures of interest separate from background or other structures. In this invention, segmentation refers specifically to isolating a chosen brain region of interest (for example, the striatum) from the surrounding brain tissue in each cross-sectional image. The Neuropathology Assessment Tool (NAT) achieves this through a two-tiered approach: first, a convolutional neural network roughly identifies the location of the region, and then a transformer-based model performs fine-grained pixel-level classification within that area. The outcome is a binary mask or outline indicating exactly which pixels belong to the target region This segmented output is crucial, as it forms the basis for subsequent volume calculation and analysis. By automating segmentation, the invention removes the need for manual tracing of brain regions, thereby improving consistency (eliminating subjective variability) and speed. Accurate segmentation is foundational to the NAT’s function, ensuring that only the relevant anatomical structure is quantified and analyzed in preclinical studies. Simplicial complexes mean mathematical constructs formed by connecting multiple simplices (basic geometric elements such as points, line segments, and triangles) to represent the shape or topology of a dataset. In practice, a set of data points is made into a network by linking points that lie close to each other, thereby capturing how the points cluster or form holes. In thisinvention, a simplicial complex is built from the pixels segmented as the brain region of interest. Each pixel is treated as a vertex (point in space), and pixels that are within a certain distance of each other are connected (forming edges or filled triangles among them). The resulting complex encodes the spatial continuity of the segmented region. The Neuropathology Assessment Tool (NAT) uses these complexes to verify and refine its segmentation output: a correctly segmented anatomical region will produce a single, well-connected complex, whereas any missing sections or spurious fragments would show up as breaks or separate clusters in the complex. By analyzing the connectivity of the complex, the NAT identifies and fills minor gaps and removes isolated outliers, ensuring that the final segmentation is one continuous region that faithfully reflects the actual anatomy.Striatal atrophy means the pathological loss of volume or shrinkage of the striatum, a condition commonly associated with neurodegenerative diseases such as Huntington disease. As striatal neurons degenerate and die, the overall size of this brain region diminishes, which can be quantified by measuring volume reductions compared to healthy controls or earlier time points. Striatal atrophy is a hallmark of Huntington disease, often correlating with the severity of motor and cognitive symptoms, and it can also appear in other disorders that affect basal ganglia integrity In this invention, striatal atrophy is precisely measured using the Neuropathology Assessment T ool (NAT). By automatically segmenting the striatum in serial brain sections and calculating its volume, the tool provides an objective readout of atrophy. This allows researchers to detect even subtle differences in striatal size between experimental groups (for instance, treated vs. untreated disease model animals) or over the course of disease progression in longitudinal studies. Quantifying striatal atrophy via an automated method enhances the reliability of preclinical evaluations, since it minimizes human bias and error in assessing how much the striatum has degenerated under various conditions.Striatum means an anatomical region of the brain that plays a central role in motor control and certain cognitive functions, and is a key site of pathology in Huntington disease. In mammals, the striatum is a subcortical structure composed primarily of the caudate nucleus and putamen (in rodents these components form one continuous region). It is part of the basal ganglia system and is rich in medium spiny neurons, which are especially vulnerable in Huntington disease. Early degeneration of these neurons leads to a noticeable reduction in striatal volume, making the striatum a critical biomarker region for disease progression. In this invention, the striatum is the principal brain region of interest for demonstrating the capabilities of the Neuropathology Assessment Tool (NAT). The tool automatically identifies and measures the striatum in cross-sectional images of mouse brains, using the striatum’s distinctive anatomical position and staining pattern. By focusing on the striatum, the NAT can effectively evaluate neurodegeneration: a smaller segmented striatum volume as output indicates the extent of atrophy The techniques developed, though exemplified on the striatum, can be adapted to other brain regions in similar analyses of disease impact.Structural features mean identifiable characteristics or patterns within an image that reflect underlying anatomical structures. These features can include variations in texture, intensity, shape, or organization of tissue elements that distinguish one brain region from another. For example, a particular region might have a higher cell density and therefore appear darker in a stained section, or it might have a unique arrangement of fibers that produces a distinct texture. In this invention, structural features in digital images guide the segmentation process performed by the Neuropathology Assessment Tool (NAT) The convolutional neural network and transformer models are trained to recognize the structural cues that define the boundaries of the region of interest (such as the striatum) During analysis, they leverage differences in features — like the contrast between the striatum and adjacent white matter or cortex — to accurately delineate the target region. Maintaining sensitivity to genuine structural features (while ignoring artifacts) ensures that the automated segmentation aligns with true histological boundaries. In short, the term covers the myriad visual indicators in the tissue image that the NAT uses to identify and isolate the brain structure being measuredStructural Similarity Index Matrix (SSIM) means a computational metric for comparing the similarity of two images based on their luminance, contrast, and structural patterns. SSIM yields a quantitative score that indicates how closely one image resembles another in terms of these features. In this invention, SSIM is used not for image quality assessment but to identify groups of brain images with analogous anatomical structures. The Neuropathology Assessment Tool (NAT) calculates SSIM between cross-sectional brain images (particularly focusing on the region of interest in each) and uses those similarity scores to cluster images that share a common morphology. For instance, slices that cut through a similar part of the striatum will produce high SSIM values when compared, signifying they have the same general shape and internal structure. By grouping images into clusters of high structural similarity, NAT can train or apply specialized segmentation models fine-tuned to each cluster’s characteristics. This approach ensures that variations in the shape or appearance of the brain region across different slices or specimens are accounted for, thereby enhancing segmentation accuracy across the entire dataset.Topological data analysis (TDA) means an analytical framework that applies principles of topology — a branch of mathematics concerned with spatial properties that persist through continuous deformations — to derive insights from complex data. Rather than focusing only on numerical values or local features, TDA examines the overall shape, connectivity, and voids in data. In this invention, TDA is used to enhance and validate the segmentation results of the Neuropathology Assessment Tool (NAT). After an initial identification of the brain region of interest in the images, the segmented pixels are analyzed using topological constructs like simplicial complexes. By doing so, the tool assesses how these pixels are connected and whether they form the single contiguous structure expected of an intact anatomical region For example, TDA methods can detect if the segmentation has inadvertently left a gap (a topological hole) or created an extra fragment separate from the main region. The algorithmcan then correct such issues by filling in missing pieces or removing outliers, guided by topological consistency. Through TDA, the NAT ensures that its automated segmentation is not just pixel-accurate, but also anatomically coherent, preserving the continuity and wholeness of the biological structure understudy.Transformer-based artificial intelligence model means a deep learning architecture that relies on self-attention mechanisms to interpret input data, rather than on traditional convolutional or recurrent layers. Transformers were originally developed for sequence data (like natural language), but they have been adapted for image analysis by treating images as sequences of patches or pixels. These models excel at capturing long-range relationships within data by allowing each element (each patch or pixel) to attend to every other element via learned relevance weights. In this invention, a transformer-based model is the core component responsible for detailed segmentation of the brain region of interest. After the region is isolated in an image by the CNN stage, the transformer processes that isolated sub-image using advanced attention techniques (including axial and patched attention) to determine which pixels belong to the target structure. The transformer's self-attention architecture enables it to simultaneously consider both global context and fine local details when classifying pixels, leading to a precise delineation of complex anatomical shapes. Thus, the term highlights that the segmentation model in NAT is built on the transformer paradigm, providing a flexible and powerful approach to image segmentation.U* Transformer means a custom-designed transformer-based Al model developed specifically for this invention to perform high-precision image segmentation. The name denotes a unique transformer architecture (marked by an asterisk) that integrates multiple attention strategies to meet the demands of neuropathological image analysis The U* Transformer combines axial attention (for efficient scanning of images along one axis at a time) and patched attention (for focused analysis on local image patches) within its network layers. This hybrid attention approach allows the model to capture both the global outline of a brain region and the finegrained details of its boundaries with minimal computational overhead In practice, the U* Transformer takes the cropped region of interest (supplied by the Faster R-CNN stage) and classifies each pixel in that region as part of the target structure or background. It was engineered to be highly parameter-efficient, using dramatically fewer parameters than generic vision models, thus enabling faster training and inference on high-resolution images. The U* Transformer's performance is central to the Neuropathology Assessment Tool (NAT): it delivers accurate, pixel-level segmentations that are then refined and quantified to produce the final volumetric assessment of the brain region.Visual Transformer (ViT) means a type of transformer-based neural network model specifically tailored for image analysis, which demonstrated that self-attention mechanisms alone (without convolutional filters) can achieve high performance in computer vision tasks. The ViT pioneered the approach of slicing an image into patches and feeding those as a sequenceinto a transformer, effectively treating image recognition similarly to a language processing problem. It showed that by using self-attention across these patches, the model can learn global image features and perform classification or other tasks comparably to conventional CNNs. In this invention, the ViT serves as an inspirational predecessor to the U* Transformer used in the Neuropathology Assessment Tool. The success of ViT confirmed that transformer architectures could excel at interpreting images, thereby informing the inventors’ decision to design a custom transformer for segmentation While the U* Transformer is optimized for pixel-wise segmentation of brain anatomy (rather than image-wide classification), its design philosophy — leveraging the power of self-attention for vision — draws from the proven principles of the Visual Transformer. Thus, reference to ViT highlights the modern Al foundation on which the NAT’s segmentation model is built.The advantages set forth above, and those made apparent from the foregoing description, are efficiently attained Since certain changes may be made in the above construction without departing from the scope of the invention, it is intended that all matters contained in the foregoing description or shown in the accompanying drawings shall be interpreted as illustrative and not in a limiting sense.

Claims

[0001] What is claimed is:

1. A system for assessing volumetric changes in a brain region of interest in preclinical animal models, the system comprising:a. one or more processors; andb. a non-transitory memory storing instructions that, when executed by the one or more processors, cause the one or more processors to perform operations comprising:i. receiving a plurality of digital images of cross-sections spanning a brain region of interest, each image including structural features enabling segmentation;ii. executing a convolutional neural network to identify the brain region of interest within each image based on the structural features and to generate a region-of-interest crop;iii. executing a transformer-based artificial intelligence model on the region- of-interest crop to classify, for each pixel, membership in the brain region of interest and to produce an initial segmentation, the transformer-based artificial intelligence model utilizing axial attention and patched attention; iv. applying a topological data analysis refinement to the initial segmentation by constructing and evaluating simplicial complexes over pixels connected under a defined connectivity to generate a refined segmentation having anatomical continuity;v calculating a three-dimensional volume of the brain region of interest by applying the Cavalieri principle to the refined segmentation using a known section spacing; andvi. outputting the refined segmentation and the calculated three-dimensional volume for display and storage.

2. The system of claim 1, wherein the operations further comprise performing histogram equalization and intensity normalization as image preprocessing prior to executing the convolutional neural network.

3. The system of claim 1, wherein the convolutional neural network comprises a Faster Region-Based Convolutional Neural Network configured to produce the region-of- interest crop4 The system of claim 1, wherein the transformer-based artificial intelligence model applies axial attention along rows and along columns in separate passes.

5. The system of claim 1, wherein the transformer-based artificial intelligence model applies patched attention on a grid of non-overlapping patches within the region-of- interest crop6. The system of claim 1, wherein the defined connectivity comprises 4-connectivity or 8-connectivity.

7. The system of claim 1, wherein the topological data analysis refinement identifies connected components in the initial segmentation, retains a largest connected component as the brain region of interest, and removes smaller connected components8. The system of claim 1, wherein the topological data analysis refinement fills interior voids bounded by the initial segmentation to produce the refined segmentation.

9. The system of claim 1, wherein calculating the three-dimensional volume comprises summing per-slice areas derived from the refined segmentation and multiplying by the section spacing between adjacent cross-sections.

10. The system of claim 1, wherein the non-transitory memory stores intermediate outputs comprising the region-of-interest crop, the initial segmentation, the refined segmentation, and the calculated three-dimensional volume together with audit metadata identifying model versions and processing parameters.

11. The system of claim 1, further comprising a user interface configured to receive a selection of the brain region of interest from a set comprising a striatum, a frontal cortex region, and a corpus callosum region, and to display graphical overlays showing boundaries of the refined segmentation12. The system of claim 1, wherein the non-transitory memory stores a dataset clustered by structural similarity of the brain region of interest and the transformer-based artificial intelligence model is trained using cluster-specific subsets identified by a Structural Similarity Index.

13. The system of claim 1, wherein a training dataset for the transformer-based artificial intelligence model comprises images from Hu97 / 18 humanized transgenic Huntington disease mice and Q175FDN knock-in mice14. The system of claim 1, wherein the one or more processors are further configured to compute a time-series of volumes for a subject and to output longitudinal volumetric changes.

15. The system of claim 1, wherein the one or more processors are further configured to output per-slice area values and left-striatum and right-striatum volumes when the brain region of interest includes bilateral structures.

16. The system of claim 1, wherein the one or more processors are further configured to store the section spacing and a slice index for each image and to validate that the refined segmentation corresponds to a single connected component for the brain region of interest.

17. A non-transitory computer-readable medium storing instructions that, when executed by one or more processors, cause the one or more processors to perform operations comprising:a. receiving a plurality of digital images of cross-sections spanning a brain region of interest;b. executing a convolutional neural network to identify the brain region of interest and to generate a region-of-interest crop;c executing a transformer-based artificial intelligence model on the region- of-interest crop to classify pixels and to produce an initial segmentation, the transformer-based artificial intelligence model utilizing axial attention and patched attention;d. applying a topological data analysis refinement to generate a refined segmentation by constructing and evaluating simplicial complexes over pixels connected under a defined connectivity and by retaining a largest connected component;e calculating a three-dimensional volume of the brain region of interest by applying the Cavalieri principle to the refined segmentation using a known section spacing; andf. outputting the refined segmentation and the calculated three-dimensional volume for display and storage18. The non-transitory computer-readable medium of claim 17, wherein the operations further comprise performing histogram equalization and intensity normalization prior to executing the convolutional neural network.

19. The non-transitory computer-readable medium of claim 17, wherein the convolutional neural network comprises a Faster Region-Based Convolutional Neural Network.

20. The non-transitory computer-readable medium of claim 17, wherein the defined connectivity comprises 4-connectivity or 8-connectivity and the operations further comprise removing connected components smaller than a threshold expressed as a pixel count or area.

21. The non-transitory computer-readable medium of claim 17, wherein the operations further comprise clustering the digital images using a Structural Similarity Index andtraining the transformer-based artificial intelligence model using cluster-specific subsets.

22. The non-transitory computer-readable medium of claim 17, wherein the operations further comprise generating a time-series of volumes for a subject and displaying graphical overlays showing boundaries of the refined segmentation.

23. A method for assessing volumetric changes in a brain region of interest in preclinical animal models, the method comprising:a. receiving a plurality of digital images of cross-sections spanning the brain region of interest, each image including structural features enabling segmentation;b executing a convolutional neural network to identify the brain region of interest within each image and to generate a region-of-interest crop; c. executing a transformer-based artificial intelligence model on the region-of- interest crop to classify, for each pixel, membership in the brain region of interest and to produce an initial segmentation, the transformer-based artificial intelligence model utilizing axial attention and patched attention; d. applying a topological data analysis refinement to the initial segmentation by constructing and evaluating simplicial complexes over pixels connected under a defined connectivity, identifying connected components, retaining a largest connected component, and filling interior voids to generate a refined segmentation;e. calculating a three-dimensional volume of the brain region of interest by applying the Cavalieri principle to the refined segmentation using a known section spacing; andf. outputting the refined segmentation and the calculated three-dimensional volume for display or storage.

24. The method of claim 23, further comprising performing histogram equalization and intensity normalization before executing the convolutional neural network.

25. The method of claim 23, wherein executing the convolutional neural network comprises executing a Faster Region-Based Convolutional Neural Network.

26. The method of claim 23, wherein the defined connectivity comprises 4-connectivity or 8-connectivity and the method further comprises removing connected components smaller than a size threshold expressed as a pixel count or area.

27. The method of claim 23, further comprising displaying graphical overlays indicating boundaries of the refined segmentation and the calculated three-dimensional volume.

28. The method of claim 23, further comprising generating and outputting a time-series of volumes for longitudinal analysis for a subject.

Images