System for the automated detection of lung anomalies using deep learning-based CT image reconstruction

Abstract

A system for the automated detection of lung anomalies using deep learning-based computed tomography image reconstruction, comprising: a computed tomography unit configured to generate raw projection data corresponding to a thoracic region of a subject; a preprocessing unit operationally coupled to the computed tomography acquisition unit and configured to convert the raw projection data into normalized projection representations through logarithmic transformation, scatter correction, and geometric calibration;a reconstruction processor that is communicatively linked to the preprocessing unit and configured to reconstruct volumetric image data from the normalized projection representations using a trained deep neural network architecture consisting of a multitude of convolutional layers arranged for feature extraction in projection space, domain transformation, and image space refinement; a segmentation processor that is operationally coupled to the reconstruction processor and configured to segment the reconstructed volumetric image data into lung regions and subregions based on learned spatial features; a feature extraction processor configured to derive spatial, morphological, and textual features at various scales from the segmented lung regions;a classification processor that is operationally coupled to the feature extraction processor and configured to identify and classify lung anomalies based on the extracted features using a trained neural network; and a visualization unit configured to display detected anomalies and generate diagnostic outputs, wherein the reconstruction processor and the classification processor are further configured to work together so that the classification results influence the refinement of the reconstruction.

Description

Technical field of the invention The present invention relates generally to the field of medical imaging and diagnostic systems, in particular a system and device for the automated detection of lung anomalies using advanced computed tomography (CT) image reconstruction and deep learning methods. The invention integrates image acquisition, reconstruction, enhancement, and diagnostic evaluation into a unified machine structure designed for real-time or near-real-time lung function testing. Background of the invention Computed tomography (CT) is frequently used to diagnose lung diseases such as pulmonary nodules, fibrosis, emphysema, pneumonia, and malignant tumors. Conventional CT reconstruction techniques, including filtered backprojection and iterative reconstruction, often exhibit a trade-off between radiation dose, image noise, and spatial resolution. Furthermore, manual interpretation of CT images by radiologists is time-consuming and subject to inconsistency between examiners. Existing computerized detection systems lack sufficient integration between reconstruction and diagnosis, resulting in suboptimal detection accuracy. Therefore, there is a need for a unified system capable of reconstructing high-quality CT images from low-dose scans while simultaneously detecting lung abnormalities with high sensitivity and specificity. The detection and diagnosis of lung abnormalities using computed tomography (CT) has become a cornerstone of modern clinical practice, particularly in diseases such as lung cancer, interstitial lung disease, chronic obstructive pulmonary disease (COPD), and infectious diseases including pneumonia. CT imaging provides high-resolution, volumetric representations of thoracic anatomy, enabling the visualization of the finest structural details that are not visible in conventional radiography. However, the increasing use of CT imaging has also brought challenges, including radiation exposure, limitations in image quality, and the growing burden on radiologists to provide accurate and timely interpretations.These challenges have driven the development of various computer-aided and technical approaches aimed at improving CT image reconstruction and automating diagnostic processes. Conventional CT image reconstruction techniques, such as filtered backprojection (FBP), are widely used due to their computational efficiency and simplicity. FBP reconstructs images by applying an inverse Radon transform to the projection data, typically combined with filtering to reduce blurring effects. While FBP provides clinically acceptable images at standard radiation doses, its performance deteriorates significantly at low doses. Noise amplification, fringe artifacts, and contrast loss are common problems that arise with reduced projection or photon counts. With the increasing focus on minimizing radiation exposure for patients, low-dose CT protocols have gained importance, exacerbating the limitations of FBP-based reconstruction. To address these weaknesses, iterative reconstruction (IR) methods have been introduced as an alternative to analytical methods. IR approaches, including algebraic reconstruction (ART), simultaneous iterative reconstruction (SIRT), and model-based iterative reconstruction (MBIR), attempt to solve the inverse imaging problem by iteratively refining an image estimate based on forward and backward projection models. These methods integrate statistical models for noise and system geometry, resulting in improved image quality and reduced artifacts compared to FBP, particularly at low radiation doses. However, IR methods are computationally intensive and often require significant processing time, limiting their applicability in real-time clinical workflows.Furthermore, the convergence behavior of IR methods can be sensitive to parameter selection, and overregularization can lead to excessively smooth images that obscure subtle pathological features such as small nodules or incipient fibrosis. Parallel to advances in reconstruction techniques, computer-aided detection (CAD) systems were developed to assist radiologists in identifying lung abnormalities. Early CAD systems used manually created features and rule-based procedures to detect suspicious areas in CT images. These systems typically segmented lung regions and then extracted features such as intensity, shape, and texture, which were subsequently classified using traditional machine learning techniques like support vector machines or random forests. While these approaches improved detection sensitivity, they were limited by their reliance on manually created features that often failed to adequately represent the complex and heterogeneous nature of lung pathologies. Consequently, these systems exhibited high false-positive rates and limited generalizability across different patient populations and imaging conditions. The advent of deep learning has fundamentally changed medical image analysis, particularly the detection of lung anomalies. Convolutional neural networks (CNNs) are widely used for tasks such as image classification, segmentation, and object recognition due to their ability to automatically learn hierarchical feature representations from data. In CT imaging, deep learning-based CAD systems have demonstrated improved performance compared to traditional methods, especially in the detection of lung nodules and the classification of disease patterns. Architectures such as U-Net, ResNet, and DenseNet have been adapted for medical imaging applications, enabling end-to-end learning of raw or minimally processed images.Despite these advances, most deep learning-based CAD systems work with reconstructed images generated using conventional reconstruction methods and therefore inherit their limitations. More recently, deep learning has also been applied to the CT reconstruction process itself, giving rise to data-driven reconstruction approaches. These methods aim to learn to map raw projection data or sinograms directly onto high-quality image representations. Techniques such as iterative reconstruction, deep image prior models, and generative adversarial networks (GANs) have been investigated for this purpose. These approaches have shown promise in reducing noise and artifacts while preserving anatomical detail, particularly in low-dose imaging. However, existing deep learning-based reconstruction methods are often developed independently of downstream diagnostic tasks, resulting in a lack of integration between image reconstruction and anomaly detection.Consequently, the reconstructed images may be optimized for visual quality rather than diagnostic relevance, which may limit their effectiveness in clinical decision-making. Another limitation of current solutions lies in the lack of end-to-end optimization of the entire imaging chain. In conventional workflows, CT acquisition, image reconstruction, and diagnostic analysis are treated as separate steps, each with its own optimization criteria. This fragmented approach overlooks the interactions between these steps, which could otherwise be leveraged to improve the overall system performance. For example, reconstruction techniques can suppress noise, resulting in the loss of subtle features crucial for the early detection of diseases, while detection techniques may struggle to compensate for reconstruction artifacts. The absence of a unified framework for the combined optimization of reconstruction and detection processes remains a significant drawback of existing systems. Furthermore, many deep learning-based solutions require large amounts of annotated training data, which can be difficult and expensive to obtain in the medical field. The variability of imaging protocols, scanner types, and patient populations further complicates the development of robust models that transfer well across different clinical settings. Domain adaptation and transfer learning techniques have been proposed to address these issues, but they introduce additional complexity and may not fully resolve data heterogeneity. Moreover, the black-box nature of deep learning models raises concerns about interpretability and clinical confidence, as clinicians may be hesitant to rely on systems that do not transparently explain their predictions. Another important aspect is the computing and hardware requirements associated with advanced reconstruction and detection techniques. High-performance computers, including GPUs and specialized accelerators, are often necessary to achieve acceptable processing times. This requirement can limit the use of such systems in resource-constrained environments, particularly in developing countries with limited access to modern medical infrastructure. Furthermore, integrating these computing components into existing CT systems presents technical challenges regarding data throughput, system latency, and thermal management. In addition to technical limitations, regulatory and clinical validation challenges also hinder the widespread adoption of automated lung function monitoring systems. Compliance with medical device regulations requires extensive testing, validation, and documentation, which can be time-consuming and costly. Furthermore, clinical acceptance depends on demonstrating not only accuracy but also reliability, robustness, and compatibility with existing workflows. Many existing solutions have been primarily evaluated in controlled research environments, resulting in limited real-world performance data. In light of the preceding discussion, it is clear that despite significant advances in CT image reconstruction and the automated detection of lung anomalies, existing solutions have several drawbacks. These include suboptimal performance in low-dose imaging, a lack of integration between reconstruction and detection processes, high computational complexity, reliance on large annotated datasets, limited interpretability, and challenges in clinical application. These limitations underscore the need for an integrated system that combines deep-learning-based reconstruction with automated detection within a unified framework, thereby enabling improved image quality, higher diagnostic accuracy, and efficient integration into the clinical workflow. Summary of the invention The present invention relates to a system and an associated device for the automated detection of lung anomalies using CT image reconstruction based on deep learning. The system comprises an integrated hardware structure with a CT acquisition unit, a data conversion and preprocessing unit, a reconstruction processor, a feature extraction processor, a classification processor, and a visualization and reporting unit. The reconstruction processor utilizes deep convolutional neural networks and hybrid encoder-decoder architectures to reconstruct high-resolution volume images from the raw projection data. The classification processor detects and categorizes lung anomalies using trained neural networks that operate on the reconstructed images and intermediate feature representations.The system also integrates a feedback control between the reconstruction and detection phases to iteratively improve diagnostic accuracy. The present invention aims to provide an integrated system and device for the automated detection of lung anomalies. This system combines deep learning-based CT image reconstruction with advanced diagnostic analysis, thereby enabling a simultaneous improvement in image quality and detection accuracy. The invention overcomes the limitations of conventional reconstruction methods by generating high-resolution volume data from low-dose CT data, while preserving important anatomical and pathological details for reliable clinical interpretation. A further objective of the invention is the development of a unified computing architecture in which the processes of image reconstruction, segmentation, feature extraction, and anomaly classification are closely linked, so that intermediate results from one stage dynamically inform and optimize subsequent stages. This integrated approach is intended to reduce error propagation, improve robustness, and achieve end-to-end optimization of the pulmonary diagnostic workflow. Another objective of the invention is to enable the accurate detection and classification of a wide range of lung anomalies, including but not limited to lung nodules, ground-glass opacities, consolidations, emphysematous changes and fibrotic patterns, by utilizing multi-scale, multi-branch deep learning architectures capable of capturing both global contextual information and fine-grained structural variations within lung tissue. Another objective of the invention is the integration of an adaptive feedback mechanism between the reconstruction processor and the classification processor, whereby diagnostic outputs are used to iteratively refine the reconstruction parameters and thus improve both the visual fidelity and the diagnostic relevance of the reconstructed CT images in a closed control loop. Another objective of the invention is to provide a system capable of operating in real time or near real time in clinical environments by utilizing optimized processing units, parallel computations and efficient memory management, thereby reducing diagnostic latency and enabling timely clinical decision-making. A further aim of the invention is to minimize radiation exposure for patients by enabling effective reconstruction and analysis of low-dose CT data without compromising diagnostic accuracy. This supports safer imaging procedures while maintaining high clinical utility. Another objective of the invention is to improve interpretability and clinical applicability by generating explainable outputs, including localized areas of interest, probability maps, and quantitative measurements of detected anomalies, which help clinicians understand and validate the system's predictions. Another objective of the invention is to provide a structurally integrated machine device that combines CT acquisition components with embedded computing units, data processing circuits and visualization interfaces in a compact and efficient hardware configuration, thereby enabling seamless use in the clinical environment. Another objective of the invention is to ensure compatibility with existing medical imaging infrastructure, including hospital information systems and image archiving and communication systems, through standardized communication interfaces and data formats, thereby enabling seamless integration into established clinical workflows. Another objective of the invention is the development of a scalable and adaptable system that enables continuous learning and performance improvements through the incorporation of new clinical data, thereby improving generalizability across different patient populations, imaging protocols and scanner configurations, while ensuring consistent diagnostic performance. BRIEF DESCRIPTION OF THE IMAGE 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. 1 shows a block diagram of a system for the automated detection of lung anomalies by means of reconstruction of computed tomography images based on deep learning. 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 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. 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 thereof. 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. 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. 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. Embodiments of the present disclosure are described in detail below with reference to the attached drawing. Fig. 1 shows a block diagram of a system for the automated detection of lung anomalies using deep learning-based computed tomography image reconstruction. The system 100 comprises: a computed tomography acquisition unit (102) for generating raw projection data of a thoracic region; a preprocessing unit (104) connected to the computed tomography acquisition unit that converts the raw projection data into normalized projection representations by logarithmic transformation, scatter correction, and geometric calibration; and a reconstruction processor (106) connected to the preprocessing unit that reconstructs volumetric image data from the normalized projection representations using a trained deep neural network architecture.This architecture includes several convolutional layers arranged for feature extraction in projection space, domain transformation, and refinement in image space; a segmentation processor (108) connected to the reconstruction processor, which segments the reconstructed volumetric image data into lung regions and subregions based on learned spatial features.A feature extraction processor (110) configured to derive spatial, morphological, and textual features at various scales from the segmented lung regions; a classification processor (112) operationally coupled to the feature extraction processor and configured to identify and classify pulmonary anomalies based on the extracted features using a trained neural network; and a visualization unit (114) configured to display detected anomalies and generate diagnostic outputs, wherein the reconstruction processor and the classification processor are further configured to operate in an interconnected manner such that the classification results influence the reconstruction refinement. In one embodiment, the preprocessing unit (104) further comprises a correction processor configured to perform beam hardening correction, detector response normalization and noise distribution compensation before generating the normalized projection representations, thereby ensuring the consistency of the input data for the reconstruction processor. In one embodiment, the reconstruction processor (106) comprises a sinogram feature processor configured to extract features from the projection area using a variety of convolution filters, a transformation processor configured to map features from the projection area into representations from the image area by learned inverse transformations, and a refinement processor configured to improve the quality of the reconstructed image using residual connections and attention-based weighting of feature maps. In one embodiment, the refinement processor is further configured to suppress fringe artifacts and quantum noise by applying learned filter operations that selectively preserve high-frequency anatomical structures while attenuating noise components. In one embodiment, the segmentation processor (108) comprises a multi-scale folding structure configured to perform hierarchical feature aggregation, combining low-level spatial features and high-level contextual features through skip connections to achieve accurate delineation of lung boundaries and internal anatomical structures. In one embodiment, the segmentation processor (108) is further configured to use dilated convolution operations to expand receptive fields without loss of spatial resolution, thus enabling the detection of diffuse lung patterns. In one embodiment, the feature extraction processor (110) is configured to compute three-dimensional volumetric descriptors, including voxel intensity distributions, gradient-based texture representations, and shape-based descriptors derived from segmented regions. In one embodiment, the feature extraction processor (110) further comprises a graph construction processor configured to represent anatomical structures as interconnected nodes and edges, thus enabling relational feature analysis between adjacent lung regions. In one embodiment, the classification processor (112) comprises a plurality of parallel processing branches, each branch being trained to detect a specific category of lung abnormalities, including nodules, ground-glass opacities, consolidations and fibrotic patterns, and the outputs of the plurality of branches being combined using a probabilistic aggregation processor. In one embodiment, the classification processor (112) is further configured to generate localization maps that show spatial areas associated with detected anomalies, thereby ensuring the interpretability of the classification results. The system is implemented entirely through dedicated physical components arranged as integrated hardware modules for imaging and processing, rather than through abstract or software-defined functions. The computed tomography unit is implemented as a radiation generation and detector unit, consisting of X-ray source hardware, rotating gantry mechanics, and solid-state detector arrays. These are configured to physically acquire projection signals from a patient's thoracic region and output corresponding electrical projection data. The preprocessing unit consists of solid-function electronics with signal conditioning circuits, digital filter hardware, and calibration logic blocks. These perform logarithmic conversion, scatter compensation, geometric alignment, beam hardening correction, detector normalization, and noise reduction directly on the incoming projection signals.The reconstruction processor utilizes high-performance computing devices such as application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or parallel matrix computational circuits. These perform convolution operations, feature extraction in the projection domain, learned inverse transformations, and image space refinement through dedicated hardware-accelerated convolution engines, residual processing blocks, and attention weighting circuits. The segmentation processor is implemented as specialized image processing hardware and includes multiscale convolution modules, dilated filter banks, and skip-connection routing circuits for the physical partitioning of volumetric data into lung regions and substructures.The feature extraction processor is implemented as hardware for volumetric analysis and includes voxel computation arrays, gradient computation circuits, texture coding logic, and graph mapping hardware for establishing node-edge relationships between anatomical regions. The classification processor is implemented through parallel inference hardware branches consisting of dedicated neural circuits that simultaneously operate on different paths for anomaly detection and combine their outputs using hardware-based probabilistic fusion logic. The visualization unit is implemented as image rendering hardware and generates display-ready diagnostic outputs and spatial maps. The present invention relates to a system and an associated device for the automated detection of lung anomalies by means of an integrated pipeline that combines deep learning-based computed tomography image reconstruction with advanced diagnostic inference. The system begins with a computed tomography unit that generates raw projection data from various angled views of a patient's thoracic region. This projection data is transferred to a preprocessing unit that performs a series of normalization operations. These include logarithmic intensity transformation to convert the measured photon intensities into attenuation coefficients, correction of beam hardening effects of polychromatic X-ray sources, compensation for detector non-uniformity, and geometric calibration to align the projection data with the system geometry.The preprocessing unit also performs noise distribution adjustment to standardize the statistical properties of the projection data and thus generate normalized projection representations suitable for input into the reconstruction processor. The reconstruction processor implements a deep learning-based reconstruction method that operates directly on the normalized projection representations. It includes a feature extraction stage in the projection domain, where multiple convolutional layers process the projection data to learn spatial correlations, noise distributions, and structural patterns of the acquisition geometry. These convolutional layers are configured with kernels of varying sizes to capture both local and global dependencies between detector channels and projection angles. The output of this stage is a set of high-dimensional feature maps representing coded information about the projection domain. Following feature extraction in the projection domain, the reconstruction processor performs a learned domain transformation, mapping the encoded features into the image space. This transformation is implemented using a neural operator that approximates the inverse mapping between projection and image space, effectively replacing or extending the analytical inverse Radon transform. The transformation is trained to maintain structural consistency while compensating for incomplete or noisy projection data. The resulting intermediate image domain representation is then processed in an image refinement phase using residual learning structures. These structures include skip connections that allow the passing of lower-level spatial information along with higher-level features, thus preventing the loss of fine anatomical details.The refinement phase also includes attention-based weighting mechanisms that assign adaptive meaning to feature channels. This enables the selective highlighting of diagnostically relevant regions while simultaneously suppressing noise and reconstruction artifacts such as banding. The volumetric image generated by the reconstruction processor is then fed to a segmentation processor configured to delineate lung regions and internal anatomical structures. The segmentation processor employs a multiscale convolutional architecture, processing feature maps at different resolutions. Low-resolution representations capture the global anatomical context, while high-resolution representations preserve contour accuracy. These representations are combined through hierarchical aggregation processes, enabling precise segmentation of the lung parenchyma, airways, and surrounding tissue. The segmentation processor utilizes advanced convolution operations to enlarge the receptive field without reducing spatial resolution. This allows for the detection of diffuse pathological patterns, such as interstitial changes. Following segmentation, the segmented lung regions are processed by a feature extraction processor, which computes a comprehensive set of descriptors. This processor analyzes the intensity distributions of the voxels to capture density variations associated with different lung diseases. Gradient-based operators are applied to quantify edge information and texture heterogeneity, while a morphological analysis characterizes the shapes and spatial distributions of anatomical structures. In addition to the voxel-based features, the feature extraction processor creates a graph-based representation of the lung anatomy. In this representation, the nodes correspond to localized regions or segmented structures, and the edges represent spatial or functional relationships.This graph representation allows the system to capture relational dependencies between neighboring regions, which are crucial for identifying patterns such as clustered nodules or diffuse disease spread. The extracted features are fed into a classification processor configured for the detection and categorization of lung anomalies. This processor consists of multiple parallel processing branches, each trained to detect a specific anomaly class. Each branch processes the input features using a sequence of convolutional and fully connected layers to generate class-specific predictions. The outputs of the parallel branches are aggregated using a probabilistic fusion mechanism that calculates confidence scores for each anomaly class. The classification processor also generates spatial localization maps by projecting the class activation information onto the reconstructed image, thus identifying the regions that contribute most to each classification decision.This allows for interpretation by highlighting areas of interest that correspond to the detected anomalies. A key aspect of the invention is the integration of a feedback processor that establishes a closed-loop interaction between the classification and reconstruction processors. The feedback processor evaluates classification results, including confidence scores and localization maps, to identify areas where the reconstruction quality may be insufficient for an accurate diagnosis. Based on this evaluation, the feedback processor generates adjustment signals that are forwarded to the reconstruction processor. Using these signals, the reconstruction parameters are modified through gradient-based optimization, thereby selectively controlling the reconstruction process to improve features in diagnostically relevant areas. This iterative optimization process continues until the convergence criteria are met.The result is reconstructed images that are optimized in terms of both visual quality and diagnostic accuracy. The system also includes a visualization unit that displays the reconstructed volume data along with segmentation overlays and anomaly indicators. The visualization unit utilizes a rendering processor that generates three-dimensional representations, allowing clinicians to view anatomical structures from various perspectives. A reporting processor integrated into the visualization unit creates structured diagnostic reports with quantitative measurements such as lesion size, volume extent, spatial coordinates, and associated confidence levels. These reports are formatted for integration with external medical information systems. The entire system is supported by a memory unit that stores trained neural network parameters, intermediate results, and historical patient data. This memory unit enables incremental learning by allowing the system to update its parameters based on newly acquired, annotated datasets, thereby improving performance over time. The reconstruction and classification processors are implemented on parallel processing hardware to ensure efficient computation and thus real-time or near-real-time operation. The integration of data acquisition, reconstruction, segmentation, feature extraction, classification, and feedback processes within a unified system architecture ensures the optimization of each individual phase with regard to the overarching diagnostic goal. By combining deep-learning-based reconstruction with automated detection and iterative feedback, the invention achieves improved image quality, higher detection accuracy, and efficient integration into the clinical workflow, thereby overcoming the limitations of existing systems. The invention describes a machine-implemented system in the form of a structural device, consisting of a gantry-mounted CT acquisition unit and an integrated computing unit with multiple processing units and memory architectures. The CT acquisition unit comprises an X-ray source and a detector array for generating projection data of the thoracic region of a subject. The acquired projection data are transferred to a data conversion unit, which performs a logarithmic transformation, beam hardening correction, and normalization to generate sinogram representations. The system comprises a reconstruction processor with one or more graphics processing units (GPUs) and tensor processing units (TPUs) connected to high-speed memory. The reconstruction processor is configured to execute a deep learning-based reconstruction algorithm that replaces or complements traditional analytical reconstruction. The reconstruction algorithm incorporates a trained neural network architecture with a sinogram-area feature extractor, a domain transformation unit, and an image-area refinement network. The sinogram-area feature extractor uses convolutional layers to capture projection inconsistencies and noise characteristics. The domain transformation unit maps sinogram features to image space using learned inverse operators.The image area refinement network uses residual learning and attention mechanisms to improve structural accuracy and suppress artifacts. This results in a reconstructed volumetric CT image with an improved signal-to-noise ratio. The system further includes a segmentation and feature extraction processor that receives the reconstructed volume image and subdivides it into anatomically relevant regions, including lung lobes and airway structures. The segmentation processor utilizes a multiscale convolutional neural network (CNN) with dilated folds and skip connections to obtain fine details. Following segmentation, a feature extraction processor computes spatial, textual, and morphological features using convolutional kernels and graph-based representations of lung structures. A classification processor is operationally coupled to the feature extraction processor and is used to identify lung anomalies using a deep neural network trained on annotated datasets. The classification processor has a multi-stage architecture, with each branch specializing in the detection of specific anomalies such as nodules, ground-glass opacities, consolidations, and fibrotic patterns. The outputs of the branches are aggregated using a probabilistic fusion unit, which generates a confidence score for each detected anomaly. The classification processor is also configured to employ attention-driven mechanisms to locate relevant areas and deliver traceable outputs. The system also includes a feedback optimization processor that iteratively refines the reconstruction process based on the classification results. This processor analyzes discrepancies between predicted anomalies and expected anatomical patterns and adjusts the reconstruction parameters using backpropagation-based optimization. This closed-loop interaction between reconstruction and detection improves both image quality and diagnostic performance. The device comprises a visualization and reporting unit with a display interface and a structured report processor. The visualization unit displays reconstructed images, segmentation overlays, and detected anomalies in three dimensions. The report processor generates automated diagnostic reports with quantitative metrics such as lesion size, volume, and probability values. The device also features a communication interface for transmitting results to hospital information systems and cloud-based databases. The device's design integrates all units into a compact housing that is mounted next to or integrated into the CT gantry, thus enabling real-time processing. The device features a cooling system, an energy management circuit, and high-speed data buses to ensure efficient operation under continuous imaging conditions. During operation, the system acquires raw data from a patient's CT scans, preprocesses it to generate normalized sinograms, and reconstructs high-resolution volumetric data using a deep learning-based reconstruction processor. The reconstructed images are segmented and analyzed to extract relevant features, which are then processed by a classification processor to detect lung abnormalities. The system iteratively refines the reconstruction based on detection feedback and generates a comprehensive diagnostic report for clinical use. The main objective of the invention is to provide an integrated system for reconstructing high-quality CT images from low-dose data while simultaneously detecting lung anomalies with high precision. A further objective is to reduce diagnostic time and variability by automating the detection process using deep learning methods. Furthermore, the invention aims to provide a device that integrates data acquisition, reconstruction, and diagnostic processing in a unified, real-time capable structure. Finally, the invention seeks to improve interpretability through visualization and comprehensible results to support clinicians in their decision-making. The present invention relates to the field of medical imaging and computer-aided diagnostic systems, in particular a system and a device for the automated detection of lung anomalies by means of deep learning-based computed tomography image reconstruction. The invention relates in particular to integrated computing architectures that combine projection data processing, image reconstruction, segmentation, feature extraction, and anomaly classification in a unified system to enable improved diagnostic accuracy in low-dose computed tomography scans. 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. 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 the automated detection of lung anomalies using deep learning-based image reconstruction in computed tomography. 102 Computed tomography acquisition unit 104 Preprocessing unit 106 Reconstruction processor 108 Segmentation processor 110 Feature extraction processor 112 Classification processor 114 Visualization unit

Claims

A system for the automated detection of lung anomalies using deep learning-based computed tomography image reconstruction, comprising: a computed tomography unit configured to generate raw projection data corresponding to a thoracic region of a subject; a preprocessing unit operationally coupled to the computed tomography acquisition unit and configured to convert the raw projection data into normalized projection representations through logarithmic transformation, scatter correction, and geometric calibration;a reconstruction processor that is communicatively linked to the preprocessing unit and configured to reconstruct volumetric image data from the normalized projection representations using a trained deep neural network architecture consisting of a multitude of convolutional layers arranged for feature extraction in projection space, domain transformation, and image space refinement; a segmentation processor that is operationally coupled to the reconstruction processor and configured to segment the reconstructed volumetric image data into lung regions and subregions based on learned spatial features; a feature extraction processor configured to derive spatial, morphological, and textual features at various scales from the segmented lung regions;a classification processor that is operationally coupled to the feature extraction processor and configured to identify and classify lung anomalies based on the extracted features using a trained neural network; and a visualization unit configured to display detected anomalies and generate diagnostic outputs, wherein the reconstruction processor and the classification processor are further configured to work together so that the classification results influence the refinement of the reconstruction. System according to claim 1, wherein the preprocessing unit further comprises a correction processor configured to perform beam hardening correction, detector response normalization and noise distribution compensation prior to generating the normalized projection representations, thereby ensuring the consistency of the input data for the reconstruction processor. System according to claim 1, wherein the reconstruction processor comprises a sinogram feature processor for extracting projection domain features by means of a plurality of convolution filters, a transformation processor for mapping projection domain features into image domain representations by means of learned inverse transformation, and a refinement processor for improving the quality of the reconstructed image by means of residual connections and attention-based weighting of feature maps. System according to claim 3, wherein the refinement processor is further configured to suppress fringe artifacts and quantum noise by applying learned filter operations that selectively preserve high-frequency anatomical structures while attenuating noise components. System according to claim 1, wherein the segmentation processor comprises a multiscale folding structure configured for hierarchical feature aggregation, combining low-level spatial features and high-level contextual features by skip connections to achieve accurate delineation of lung boundaries and internal anatomical structures. System according to claim 5, wherein the segmentation processor is further configured to utilize dilated folding operations to expand the receptive fields without loss of spatial resolution, thereby enabling the detection of diffuse lung patterns. System according to claim 1, wherein the feature extraction processor is configured to compute three-dimensional volumetric descriptors, including voxel intensity distributions, gradient-based texture representations, and shape-based descriptors derived from segmented regions. System according to claim 7, wherein the feature extraction processor further comprises a graph construction processor configured to represent anatomical structures as interconnected nodes and edges, thus enabling relational feature analysis between adjacent lung regions. System according to claim 1, wherein the classification processor comprises a plurality of parallel processing branches, each of which is trained to detect a specific category of lung abnormalities, including nodules, ground-glass opacities, consolidations and fibrotic patterns, and wherein the outputs of the plurality of branches are combined by means of a probabilistic aggregation processor. System according to claim 9, wherein the classification processor is further configured to generate localization maps that show spatial areas associated with detected anomalies, thereby ensuring the interpretability of the classification results.

Images