System for automated histopathological image analysis using in-depth, multi-stage extraction of cellular features

Abstract

A system for automated histopathological image analysis using deep, multiscale cellular feature extraction, comprising the following: a digitization unit for microscope slides, configured to receive a histological glass slide and to produce a high-resolution digital image using an optical imaging arrangement consisting of a light source, a lens assembly and an image sensor; a preprocessing unit that is operationally coupled with the digitization unit for microscope slides and is configured to perform normalization of the color, color separation and artifact suppression on the digital image; a patch generation unit configured to segment the digital image into a multitude of image patches with different magnification levels; a hierarchical feature extraction unit consisting of a multitude of convolutional processing layers arranged in parallel paths, each path being configured to process image sections according to a specific spatial resolution in order to extract cellular and structural features; a feature aggregation unit configured to combine feature representations derived from a multitude of parallel paths using an adaptive weighting mechanism; and a classification unit configured to receive the aggregated feature representation and generate a diagnostic output corresponding to one or more pathological classes.

Description

Technical field of the invention

[0001] The present disclosure relates generally to the field of computer-aided pathology and medical image analysis, and in particular to a system and associated device architecture for automated histopathological image interpretation using deep-learning-based, multi-scalar cellular feature extraction. The invention integrates advanced digital imaging, hierarchical feature display, and predictive classification to support precise and reproducible diagnosis of tissue abnormalities. Background of the invention

[0002] Histopathological examination remains the gold standard for diagnosing various diseases, including cancer. However, the manual evaluation of histological specimens is time-consuming, subjective, and prone to discrepancies between examiners. Conventional image processing techniques are unable to effectively capture complex cell morphologies across different spatial resolutions. While newer deep learning approaches show promise, existing systems have limitations regarding the integration of contextual information across different scales, the consideration of staining variability, and the robust classification of heterogeneous datasets. Therefore, there is a need for a technologically advanced system that combines the extraction of features across different scales with automated inference mechanisms to improve diagnostic precision and workflow efficiency.

[0003] Histopathological image analysis is an essential component of modern diagnostics, particularly in oncology. The microscopic examination of stained tissue sections allows for the identification of cellular abnormalities, the determination of tumor grade, and the assessment of disease progression. Traditionally, this analysis is performed manually by trained pathologists who visually examine slides stained with dyes such as hematoxylin and eosin (H&E) using light microscopes. While manual examination remains the clinical gold standard, it is inherently limited by subjectivity, variability in interpretation, and the increasing workload resulting from the growing global demand for pathology services.Variability between and within examiners often leads to inconsistent diagnoses, particularly in borderline or complex cases where subtle morphological differences must be interpreted with high precision. Furthermore, manual analysis is time-consuming and requires extensive expertise and sustained attention, resulting in fatigue-related errors and delays in diagnosis, thus negatively impacting patient outcomes.

[0004] To overcome these limitations, computer-aided approaches based on classical image processing and machine learning methods were developed early on. These systems typically used manually developed feature extraction methods, in which domain-specific features such as texture descriptors, color histograms, morphological parameters, and edge-based characteristics were manually designed and extracted from digitized histopathological images. Techniques such as grayscale co-occurrence matrix (GLCM), local binary patterns (LBP), and wavelet transformations were frequently employed to characterize tissue textures and cell structures. These features were then fed to traditional classifiers such as support vector machines (SVM), k-nearest neighbors (k-NN), or random forests for classification tasks.Although these approaches represented a significant improvement over purely manual analysis, they were fundamentally limited by their reliance on predefined feature sets that often failed to capture the full complexity and heterogeneity of histopathological patterns. The performance of such systems depended heavily on expertise in feature engineering, and their generalizability to different datasets, staining protocols, and imaging conditions remained limited.

[0005] With the advent of digital pathology and whole-seat imaging, it became possible to scan entire histological specimens at high resolution, generating gigapixel images suitable for computer-aided analysis. This development enabled the use of more sophisticated techniques for processing large image datasets. However, the sheer size and complexity of whole-seat images presented new challenges regarding computational efficiency, storage capacity, and the effective representation of information of varying scales. Early approaches to processing whole-seat images often relied on patch-based methods, in which the image was divided into smaller, independently analyzable areas.While this strategy reduced computational effort, it led to problems related to the loss of contextual information, as local patches often lacked the broader tissue context necessary for accurate interpretation. Consequently, classification models based solely on patch features frequently exhibited lower diagnostic accuracy, particularly in cases where global architectural patterns played a crucial role.

[0006] The advent of deep learning, particularly convolutional neural networks (CNNs), marked a fundamental shift in histopathological image analysis. CNN-based models demonstrated the ability to automatically learn hierarchical feature representations directly from raw image data, eliminating the need for manual feature development. Architectures such as AlexNet, VGGNet, ResNet, and Inception were adapted for medical imaging tasks like tumor detection, tissue classification, and segmentation. These models achieved significant performance improvements compared to traditional methods because they can capture complex spatial patterns and nonlinear relationships within the data. However, standard CNN architectures are inherently designed for a fixed input resolution, which is a significant limitation in histopathology, where diagnostically relevant features exist at various spatial scales.For example, nuclear atypia is visible microscopically, while glandular organization and tissue architecture require analysis at lower magnification. Conventional CNN models often struggle to capture both fine and coarse-grained features simultaneously, leading to suboptimal results when analyzing data at different scales.

[0007] To overcome this limitation, various multi-scale and multi-resolution approaches have been proposed. These include pyramid-based representations, which process images at multiple resolutions, and multi-branch network architectures, which use separate convolutional paths for different scales. While these approaches show promise, they increase the complexity of model design and training. A significant drawback is the challenge of effectively integrating features across different scales, as simple concatenation or averaging may fail to preserve the relative importance of features from different resolutions. Furthermore, multi-branch architectures often require substantial computational resources, making their use in real-time clinical environments difficult.Another limitation lies in the lack of adaptive mechanisms for dynamically prioritizing relevant features based on context information, which leads to inefficient use of the learned representations.

[0008] Another significant challenge in histopathological image analysis is the variability of staining and imaging conditions. Differences in staining protocols, scanner calibration, and tissue preparation can lead to substantial variations in color distribution and image quality across different datasets. Although stain normalization techniques have been developed to address this issue, many existing methods rely on fixed reference templates or linear transformations that may not adequately capture nonlinear variations. Consequently, models trained on one dataset often exhibit reduced performance when applied to images from other sources, highlighting the domain shift problem. This lack of robustness limits the practical applicability of existing solutions in diverse clinical settings.

[0009] The segmentation of histopathological structures such as cell nuclei and glands represents another area where existing solutions reach their limits. Traditional segmentation methods based on thresholding, watershed techniques, and contour recognition often struggle with overlapping structures, irregular shapes, and varying contrast levels. Deep learning-based segmentation models such as U-Net and its variants have significantly improved segmentation accuracy; however, they too are prone to errors in dense cell regions, with artifacts, or with fuzzy boundaries. Furthermore, these models typically require large amounts of annotated data for training, the acquisition of which is labor-intensive and time-consuming due to the necessary expert annotations.

[0010] The interpretability and explainability of automated systems remain crucial challenges. Many deep learning models function as "black boxes," delivering predictions without revealing the underlying decision-making process. In clinical practice, pathologists need not only accurate predictions but also interpretable, verifiable, and trustworthy results. Existing approaches to improving interpretability, such as class activation maps and attention mechanisms, offer some visualization but often lack the necessary granularity or consistency to be fully reliable in clinical decision-making.

[0011] Computational efficiency and scalability pose additional challenges for existing solutions. Processing high-resolution whole-specimen images requires significant computing power, memory bandwidth, and storage capacity. Many modern deep learning models are resource-intensive and require specialized hardware, such as high-performance GPUs, for both training and inference. This limits their use in resource-constrained environments, such as smaller laboratories or developing countries. Furthermore, real-time analysis remains difficult, especially when multiple specimens need to be processed simultaneously.

[0012] Another drawback of current systems is their limited integration into clinical workflows. Many existing solutions function as standalone tools requiring manual data transfer, preprocessing, and interpretation, leading to inefficiencies and potential sources of error. Seamless integration with laboratory information systems, electronic health records, and diagnostic reporting systems is often lacking, hindering widespread adoption. Furthermore, regulatory and validation challenges persist, as automated systems must demonstrate consistent performance across diverse patient populations and disease patterns.

[0013] In summary, despite significant advances in histopathological image analysis through digital imaging and deep learning, existing solutions continue to face technical and practical limitations. These include inadequate multiscale feature representation, challenges in feature integration, sensitivity to staining variability, limitations in segmentation accuracy, lack of interpretability, high computational costs, and insufficient integration into clinical workflows. These drawbacks underscore the need for a more robust and efficient system capable of utilizing multiscale information, adapting to varying imaging conditions, and providing precise, interpretable, and scalable diagnostic support. Summary of the invention

[0014] The present invention describes a system and an associated device for automated histopathological image analysis using comprehensive, multiscale cellular feature extraction. The system comprises a digital slide acquisition unit, a preprocessing unit for stain normalization and artifact removal, a hierarchical feature extraction unit based on a multi-branch convolutional neural network architecture, a feature aggregation unit for fusing cellular features across multiple magnification levels, and a classification unit for generating diagnostic results. The system further includes a device with an integrated imaging platform, embedded computing hardware, and a display interface for real-time visualization and decision support.

[0015] The present invention aims to provide a technically advanced system for automated histopathological image analysis that overcomes the limitations of manual interpretation and conventional computer-aided methods by extracting in-depth, multiscale cellular features, thereby achieving higher diagnostic accuracy and consistency. The invention enables the precise identification and classification of pathological features by capturing both detailed cell morphology and the overall tissue architecture within a unified analytical framework.

[0016] A further objective of the invention is to provide a system for extracting hierarchical features from histopathological images across different spatial resolutions. This enables the simultaneous analysis of nuclear, cytoplasmic, and structural patterns, which are crucial for precise disease characterization. The invention aims to integrate these multiscale features through an adaptive aggregation mechanism that preserves contextual relevance and improves classification performance in complex and heterogeneous tissue samples.

[0017] A further objective of the invention is to provide a preprocessing mechanism for normalizing staining variations and removing image artifacts. This improves the robustness and generalizability of the system for datasets from different laboratories, scanners, and staining protocols. The invention aims to ensure consistent performance regardless of the image acquisition conditions.

[0018] A further objective of the invention is to provide a system with an efficient computing architecture for the scalable and time-saving processing of high-resolution whole-specimen images. The invention aims to reduce the computational effort while simultaneously ensuring high analytical accuracy, thereby enabling practical use in clinical environments with varying computing resources.

[0019] A further objective of the invention is to provide an integrated device with a slide scanning unit, a processing unit, and a visualization interface. The device is configured for real-time image acquisition, processing, and diagnostic output. The invention aims to provide a compact and structurally optimized device that enables seamless integration into the workflows of pathology laboratories.

[0020] A further objective of the invention is to provide a system that, in addition to classification results, also enables the spatial localization of pathological areas through segmentation and visualization, thus supporting pathologists in identifying diagnostically relevant areas. The invention aims to improve interpretability and support clinical decision-making by presenting visually intuitive results.

[0021] A further objective of the invention is the integration of adaptive learning functions into the system, thereby continuously improving model performance through training with diverse and annotated datasets. The invention aims to increase generalizability and ensure high accuracy across various disease types and patient populations.

[0022] A further aim of the invention is to provide a system that minimizes variability between different examiners by delivering standardized and reproducible diagnostic results, thereby improving the consistency of histopathological assessment. The invention aims to support pathologists by reducing subjective interpretations and providing objective analytical insights.

[0023] A further objective of the invention is seamless integration into existing clinical information systems, including laboratory information systems and digital pathology platforms, to enable efficient data management, optimized reporting, and interoperability. The invention aims to integrate automated analysis into the clinical workflow without significant disruption.

[0024] A further objective of the invention is to provide a scalable and extensible framework that can be adapted to various histopathological applications, including cancer detection, classification, and prognostic assessment, thereby expanding its applicability in several areas of medical diagnostics. The invention aims to offer a comprehensive solution for automated, reliable, and highly efficient histopathological image analysis. BRIEF DESCRIPTION OF THE IMAGE

[0025] These and other features, aspects and advantages of the present invention will be better understood if the following detailed description is read with reference to the accompanying drawing, in which the same symbols represent the same parts: Fig. Figure 1 shows a block diagram of a system for automated histopathological image analysis using deep multiscale cellular feature extraction.

[0026] Furthermore, those skilled in the art will recognize that the elements in the drawing are simplified and not necessarily drawn to scale. For example, the flowcharts illustrate the process by highlighting the main steps to facilitate understanding of the present disclosure. With regard to the construction of the device, one or more components may be represented in the drawing by conventional symbols. The drawing may show only those specific details relevant to understanding the embodiments of the present disclosure, so as not to clutter the drawing with details that are already apparent to those skilled in the art from the description contained herein. Detailed description of the invention

[0027] To facilitate understanding of the principles of the invention, reference is made below to the embodiment shown in the drawing, which is described using specific terms. It is understood, however, that this does not limit the scope of protection of the invention. Rather, modifications and further developments of the depicted system, as well as further applications of the inventive principles shown therein, are conceivable, insofar as they would normally occur to a person skilled in the art in the field of the invention.

[0028] It will be clear to those skilled in the art that the foregoing general description and the following detailed description are exemplary and explanatory of the invention and are not to be understood as a limitation of it.

[0029] References to “an aspect”, “another aspect”, or similar phrases in this description mean that a particular feature, structure, or property described in connection with the embodiment is included in at least one embodiment of the present disclosure. Therefore, phrases such as “in one embodiment”, “in another embodiment”, and similar expressions in this description may, but do not necessarily, all refer to the same embodiment.

[0030] The terms "includes," "comprehensive," or similar expressions denote non-exclusive inclusion. Thus, a procedure or method containing a list of steps does not only include those steps but may also include further steps not explicitly listed or inherent in the procedure or method. Likewise, the statement "includes..." for one or more devices, subsystems, elements, structures, or components, without further limitations, does not preclude the existence of other devices, subsystems, elements, structures, or components.

[0031] Unless otherwise defined, all technical and scientific terms used herein have the same meanings generally known to those skilled in the art in the field to which this invention belongs. The systems, methods, and examples described herein serve only for illustration and are not to be understood as limiting.

[0032] Embodiments of the present disclosure are described in detail below with reference to the attached drawing.

[0033] Fig.Figure 1 shows a block diagram of a system for automated histopathological image analysis using deep multiscale cellular feature extraction. The system 100 comprises: a slide digitizing unit (102) configured to receive a histological slide and generate a high-resolution digital image using an optical imaging arrangement consisting of an illumination source, a lens assembly, and an image sensor; a preprocessing unit (104) operationally linked to the slide digitizing unit and configured to perform stain normalization, color separation, and artifact suppression on the digital image; and a patch generation unit (106) configured to segment the digital image into a multitude of image patches with multiple magnification levels.a hierarchical feature extraction unit (108) with multiple convolutional processing layers arranged in parallel paths, each path configured to process image sections according to a specific spatial resolution to extract cellular and structural features; a feature aggregation unit (110) configured to combine the feature representations obtained from the multiple parallel paths using an adaptive weighting mechanism; and a classification unit (112) configured to receive the aggregated feature representation and generate a diagnostic output according to one or more pathological classes.

[0034] In one embodiment, the slide digitizing unit (102) further comprises a motorized table configured to move the histological slide along orthogonal axes with submicrometer-accurate positioning, and a focus control unit configured to dynamically adjust the focus depth to capture tissue sections of varying thicknesses and thus ensure uniform image sharpness across the entire slide.

[0035] In one embodiment, the preprocessing unit (104) comprises a color deconvolution processor configured to split the color components into different channels representing different tissue components, followed by a normalization processor configured to adjust the statistical distribution of color intensities to a predefined reference distribution using histogram transformation techniques.

[0036] In one embodiment, the preprocessing unit (104) further comprises a noise reduction processor configured to apply spatial filtering with a Gaussian kernel and an artifact removal processor configured to detect and eliminate non-tissue areas, folds and deposits using morphological operations and threshold-based segmentation.

[0037] In one embodiment, the patch generation unit (106) is configured to extract overlapping image sections at predefined scales corresponding to high, medium, and low magnification. Each section contains spatial coordinate information for subsequent mapping to the original digital image.

[0038] In one embodiment, the hierarchical feature extraction unit (108) comprises a first folding path configured to process high-resolution image sections for the extraction of nuclear morphology features such as shape, interface irregularity, and chromatin distribution; a second folding path configured to process medium-resolution image sections for the extraction of cell arrangement features such as clustering and glandular formation; and a third folding path configured to process low-resolution image sections for the extraction of global tissue architectural features.

[0039] In one embodiment, each convolution path comprises a sequence of convolution layers, normalization layers, nonlinear activation functions, and pooling layers arranged such that the spatial dimensionality is progressively reduced while retaining discriminative feature information.

[0040] In one embodiment, the feature aggregation unit (110) comprises an attention-based processor configured to assign adaptive importance weights to the feature vectors obtained from each spatial resolution, the weights being determined on the basis of the context relevance of the features within a given image area.

[0041] In one embodiment, the feature aggregation unit (110) further comprises a fusion processor configured to concatenate the weighted feature vectors using fully connected layers and transform them into a unified representation.

[0042] In one embodiment, the classification unit (112) comprises a plurality of fully interconnected layers, followed by a probabilistic output layer configured to generate class probabilities according to predefined diagnostic categories, the classification unit further comprising a calibration processor configured to adjust the probability outputs to improve prediction reliability.

[0043] The described system is implemented through a coordinated arrangement of hardware modules that process histopathological samples and image data. The slide digitization unit utilizes an optical imaging setup with a structured illumination source, a precision lens, and a solid-state image sensor mounted on a motorized translation stage with integrated focusing. This enables the precise acquisition of tissue sections at various depths. The preprocessing unit consists of dedicated electronics and image processing hardware for color separation, intensity normalization, and noise reduction using integrated filter components and morphological processing circuitry. The patch generation unit uses memory buffers and addressable segmentation hardware to divide high-resolution image data into areas of different scales while maintaining spatial coordinates.The hierarchical feature extraction unit is implemented in parallel hardware pipelines for convolutional processing. Each pipeline includes arrays of processing elements, normalization circuits, activation modules, and pooling components designed to handle different spatial resolutions. The feature aggregation unit comprises hardware-based weighting and fusion circuits with attention-driven control logic and interconnect structures that combine feature signals of different scales into a unified representation. The classification unit utilizes dedicated hardware computational layers and probabilistic output circuits to generate diagnostic decisions, as well as calibration components to adjust the output signals for improved reliability.

[0044] The system for automated histopathological image analysis using deep, multiscale cellular feature extraction employs a tightly integrated sequence of data acquisition, preprocessing, hierarchical feature calculation, cross-scale fusion, and predictive inference. Each stage is technically configured to preserve diagnostically relevant information across various spatial resolutions while ensuring high computational efficiency. After a histological slide is placed in the digitization unit, the motorized stage moves the slide in a raster scan pattern, while the optical unit captures a sequence of high-resolution image tiles. A focus control mechanism continuously adjusts the depth of field through local contrast maximization, ensuring that each captured tile is optimally sharp.The individual tiles are then combined into a coherent digital whole-specimen representation using a coordinate-based image registration technique, while preserving spatial indices for subsequent mapping and reconstruction.

[0045] The preprocessing unit receives the digitized whole-specimen image and performs staining normalization via color deconvolution. The image is decomposed into individual staining channels corresponding to different biochemical components of the tissue. This decomposition is achieved by transforming the color space into an optical density range and then projecting it onto predefined staining vectors. The resulting channels are normalized by histogram alignment to a reference distribution to minimize staining variability between specimens. Following normalization, noise reduction is applied using a Gaussian convolution filter, which reduces high-frequency noise while preserving structural boundaries.In parallel, artifact removal is performed, in which non-informative areas such as background, air bubbles, and tissue folds are identified and excluded from further analysis using adaptive thresholding and morphological filtering. The preprocessed image then undergoes tissue region recognition to ensure that only biologically relevant areas are forwarded for feature extraction.

[0046] The patch generation unit segments the normalized whole-specimen image into a structured set of overlapping patches at different magnification levels. This segmentation is achieved by scanning image areas at predefined spatial scales corresponding to high, medium, and low magnifications. Each patch is associated with spatial coordinates that specify its position in the original image. The overlap of adjacent patches is maintained to ensure the continuity of contextual information and to avoid artifacts at the edges during feature extraction. The multi-scale patch set is then passed to the hierarchical feature extraction unit, which is the central analytical component of the system.

[0047] The hierarchical feature extraction unit processes the multiscale image sections using parallel convolution processes. Each process is specifically configured to capture features at a corresponding spatial resolution. Within each process, the input images undergo a sequence of convolution operations in which learned filters are applied to extract localized patterns such as edges, textures, and structural motifs. These convolution operations are followed by normalization layers that stabilize the distribution of feature activations, and nonlinear activation functions that increase representation complexity. Pooling operations are integrated into the sequence to progressively reduce spatial dimensionality while preserving important information.In the high-resolution process, the convolution filters are optimized to detect fine cellular details such as nuclear contours, chromatin density variations, and mitotic figures. In the medium-resolution process, the filters are adapted to capture spatial relationships between cells, including clustering and glandular formation. In the low-resolution process, the filters are configured to represent global tissue organization, including architectural distortions and regional heterogeneity.

[0048] The outputs of the parallel convolution paths are transformed into feature vectors that encode information at different scales. These feature vectors are fed to the feature aggregation unit, where a cross-scale fusion technique is implemented. The aggregation process uses an attention-based weighting mechanism that dynamically assigns importance values ​​to the features derived from each scale. The attention weights are calculated by a learned transformation that evaluates the contextual relevance of each feature vector to the overall image representation. The weighted feature vectors are then concatenated and passed through a series of transformation layers that project the combined representation into a unified feature space.This fusion process ensures that both fine-grained and coarse-grained information contribute proportionally to the final representation, thus improving the system's ability to distinguish between differentiating features.

[0049] The unified feature representation is then provided to the classification unit, which performs predictive inference using a sequence of fully connected layers. Each layer applies a linear transformation followed by nonlinear activation, progressively refining the representation to a set of class-specific values. The final layer generates a probability output corresponding to predefined diagnostic categories. The probabilities are calculated using a normalization function that ensures their sum equals one. To improve the reliability of the predicted probabilities, a calibration procedure is applied, adjusting the output distribution based on a scaling parameter learned during the training phase. This calibration step reduces overconfidence in the predictions and improves the interpretability of the results.

[0050] In addition to classification, the system optionally performs segmentation to locate pathological areas in the image. This is achieved using an encoder-decoder architecture based on aggregated feature representation. The encoder compresses the spatial information into a compact representation, while the decoder incrementally reconstructs a spatial map corresponding to the pixel-level class assignments. Skip connections are inserted between the corresponding encoder and decoder layers to preserve fine spatial details that might otherwise be lost during compression. The resulting segmentation map is aligned to the original image coordinates using the positional metadata of the image segments, enabling precise visualization of the relevant areas.

[0051] The system is trained using annotated datasets of labeled histopathological images. The training unit iteratively optimizes the parameters of the convolutional paths, aggregation mechanism, and classification layers based on a composite objective function. This objective function includes a component that penalizes discrepancies between predicted and actual class labels, and a component that measures the overlap between predicted and actual segmentation maps. Optimization is performed using gradient-based methods, where the gradients of the objective function with respect to the model parameters are calculated and used to update the parameters toward minimizing overall loss. To improve generalization, data augmentation techniques are employed during the training process to vary the input data, including geometric transformations and intensity perturbations.

[0052] During operation, the system runs in streaming mode, with image acquisition, preprocessing, and feature extraction occurring in parallel. The processing unit, equipped with parallel computing capabilities, efficiently performs convolution operations, enabling near real-time analysis of incoming image data. The final results, including classification results and segmented regions, are displayed by the visualization unit, which overlays diagnostic information onto the original image. This integrated technology pipeline ensures precise, scalable, and interpretable analysis of histopathological images, thereby supporting clinical decision-making and optimizing diagnostic workflows.

[0053] In one embodiment, the system comprises a slide handling unit that receives histopathological samples on slides and digitizes them into high-resolution whole-slide images using an optical scanning mechanism. The optical scanning mechanism consists of a motorized stage, a high numerical aperture lens, and a CMOS-based image sensor that captures images at various magnification levels. The digitized image data is transferred to a preprocessing unit via a dedicated processor.

[0054] The preprocessing unit is configured to perform staining normalization via color deconvolution, which separates the hematoxylin and eosin components. This is followed by histogram matching against a reference template. The preprocessing unit also includes a noise reduction unit with Gaussian filtering and an artifact removal unit that eliminates background areas and tissue folds using morphological operations and threshold-based segmentation.

[0055] The system further includes a hierarchical feature extraction unit implemented via a deep convolutional architecture with multiple parallel processing branches. Each branch is configured to process image sections with different spatial resolutions. A first branch processes high-resolution sections to capture detailed cell structures such as nuclear shape, chromatin texture, and mitotic figures. A second branch processes medium-resolution sections to capture glandular organization and cell cluster patterns. A third branch processes low-resolution sections to capture global tissue architecture and the spatial distribution of pathological areas. Each branch consists of a sequence of convolutional layers, batch normalization layers, and nonlinear activation functions, followed by pooling layers to reduce spatial dimensionality.

[0056] The outputs of the parallel branches are fed to a feature aggregation unit configured for cross-scale fusion of the extracted features. This feature aggregation unit incorporates an attention-based weighting mechanism that assigns importance values ​​to features of each scale based on their contextual relevance. The aggregated feature representation is then passed to a fully connected classification unit consisting of one or more dense layers and a softmax activation function. This unit generates probability values ​​corresponding to predefined diagnostic classes.

[0057] In one embodiment, the classification unit is additionally configured to integrate a probabilistic calibration mechanism with temperature scaling to improve prediction reliability. Furthermore, it can include a segmentation unit that generates pixel-accurate annotations using an encoder-decoder architecture with skip connections, thus enabling the localization of pathological areas.

[0058] The system also includes a training unit that trains the hierarchical feature extraction model using annotated histopathological datasets. The training unit employs a composite loss function that incorporates a categorical cross-entropy component to determine classification accuracy and a Dice coefficient-based component to determine segmentation accuracy. The training process also includes data augmentation techniques such as rotation, scaling, and color variation to improve generalization.

[0059] The invention further discloses a special device for histopathological analysis, consisting of a housing with an integrated slide scanning unit, a processing unit, and a user interface module. The slide scanning unit comprises a precision-controlled motorized table for moving the slide along orthogonal axes, a light source for uniform light distribution, and an optical lens system for focusing on tissue sections at different depths.

[0060] The computing unit consists of a high-performance processing unit with a graphics processing unit (GPU) and a memory that stores the trained model parameters and intermediate results of the feature maps. The computing unit is coupled with the slide scanning unit to enable real-time image acquisition and processing.

[0061] The user interface module includes a high-resolution display unit for showing analyzed images, classification results, and highlighted areas. The device also includes input elements for user interaction, a data storage unit for archiving processed images, and a communication interface for transferring results to external systems such as hospital information systems.

[0062] The device's structural configuration ensures the compact integration of imaging, processing, and visualization components, thus enabling its use in clinical laboratories. The device is also equipped with thermal management components, including heat sinks and fans, to guarantee optimal operating conditions for consistently high computing performance.

[0063] The present invention relates to digital pathology and medical image processing, in particular a system and a device for automated histopathological image analysis using deep learning-based, multiscale cellular feature extraction. The invention lies at the interface of biomedical imaging, artificial intelligence, and diagnostic decision support systems and specifically aims at the automated interpretation of digitized tissue samples through hierarchical analysis of cellular and tissue structures across various spatial resolutions. The described system integrates advanced image acquisition, preprocessing, feature calculation, and predictive classification techniques to enable precise, reproducible, and scalable histopathological assessment in clinical and research settings.

[0064] The drawing and the preceding description illustrate embodiments. Those skilled in the art will recognize that one or more of the described elements can be combined to form a single functional element. Alternatively, certain elements can be divided into several functional elements. Elements of one embodiment can be added to another. For example, the process flows described here can be modified and are not limited to the manner described herein. Furthermore, the actions of a flowchart need not be performed in the sequence shown; nor do all actions necessarily need to be carried out. Actions that do not depend on other actions can be performed in parallel with the other actions. The scope of protection of the embodiments is in no way limited by these specific examples. Numerous variations, whether explicitly stated in the description or not, such as...Differences in structure, dimensions, and materials are possible. The scope of protection of the embodiments is at least as comprehensive as described by the following claims.

[0065] The advantages, other benefits, and problem solutions have been described above with reference to specific embodiments. However, the advantages, benefits, problem solutions, and any components that can effect or enhance an advantage, benefit, or solution are not to be construed as critical, necessary, or essential features or components of the claims. REFERENCES 100 A system for automated histopathological image analysis using in-depth, multi-stage extraction of cellular features. 102 Slide digitization unit 104 Pre-processing unit 106 Patch Generation Unit 108 hierarchical feature extraction unit 110 Feature Aggregation Unit 112 Classification unit

Claims

A system for automated histopathological image analysis using deep, multiscale cellular feature extraction, comprising: a slide digitizing unit configured to receive a glass histological slide and generate a high-resolution digital image using an optical imaging arrangement consisting of an illumination source, a lens assembly, and an image sensor; a preprocessing unit operationally coupled to the slide digitizing unit and configured to perform stain normalization, color separation, and artifact suppression on the digital image; and a patch generation unit configured to segment the digital image into a multitude of image patches with different magnification levels.a hierarchical feature extraction unit consisting of a plurality of convolutional processing layers arranged in parallel paths, each path configured to process image sections according to a specific spatial resolution to extract cellular and structural features; a feature aggregation unit configured to combine feature representations obtained from a plurality of parallel paths using an adaptive weighting mechanism; and a classification unit configured to receive the aggregated feature representation and generate a diagnostic output corresponding to one or more pathological classes. System according to claim 1, wherein the slide digitization unit further comprises a motorized table configured to move the histological slide along orthogonal axes with submicrometer-accurate positioning, and a focus control unit configured to dynamically adjust the focus depth to capture tissue sections of varying thickness and thereby ensure uniform image sharpness across the entire slide. System according to claim 1, wherein the preprocessing unit comprises a color deconvolution processor configured to split the color components into different channels representing different tissue components, followed by a normalization processor configured to adjust the statistical distribution of color intensities to a predefined reference distribution using histogram transformation techniques. System according to claim 1, wherein the preprocessing unit further comprises a noise reduction processor configured to apply spatial filtering with a Gaussian kernel, and an artifact removal processor configured to detect and eliminate non-tissue areas, folds and deposits using morphological operations and threshold-based segmentation. System according to claim 1, wherein the patch generation unit is configured to extract overlapping image sections at predefined scales corresponding to high, medium and low magnification, and wherein each section retains spatial coordinate information for subsequent mapping to the original digital image. System according to claim 1, wherein the hierarchical feature extraction unit comprises a first folding path for processing high-resolution image sections for the extraction of nuclear morphology features such as shape, interface irregularity and chromatin distribution, a second folding path for processing medium-resolution image sections for the extraction of cell arrangement features such as cluster formation and gland formation, and a third folding path for processing low-resolution image sections for the extraction of global tissue architecture features. System according to claim 6, wherein each convolution path comprises a sequence of convolution layers, normalization layers, nonlinear activation functions and pooling layers arranged such that the spatial dimensionality is progressively reduced while simultaneously preserving discriminative feature information. System according to claim 1, wherein the feature aggregation unit comprises an attention-based processor configured to assign adaptive importance weights to the feature vectors obtained from each spatial resolution, the weights being determined on the basis of the contextual relevance of the features within a given image area. System according to claim 8, wherein the feature aggregation unit further comprises a fusion processor configured to link and transform the weighted feature vectors into a unified representation using fully connected layers. System according to claim 1, wherein the classification unit comprises a plurality of fully interconnected layers followed by a probabilistic output layer configured to generate class probabilities according to predefined diagnostic categories, and wherein the classification unit further comprises a calibration processor configured to adjust the probability outputs to improve prediction reliability.

Images