Medical early-stage image feature intelligent identification system and method based on CT

Abstract

The invention discloses a CT-based medical early-stage image feature intelligent identification system and method, and relates to the technical field of image feature identification. The medical early-stage image feature intelligent identification system based on the CT comprises a model memory occupation monitoring module, a CT image first-stage monitoring module and a CT image second-stage monitoring module, and a model memory occupation monitoring result is obtained by executing model memory occupation monitoring; the first-stage monitoring result of the CT image is obtained by executing the first-stage monitoring of the CT image, and the second-stage monitoring result of the CT image is obtained by executing the second-stage monitoring of the CT image, so that the intelligent recognition accuracy of the CT image features is improved, and the problem of inaccurate intelligent recognition of the CT image features in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of image feature recognition technology, and in particular to an intelligent recognition system and method for early medical image features based on CT. Background Technology

[0002] In the early stages of intelligent recognition of medical imaging features, the process begins with acquiring neurological images of the brain, including CT (Computed Tomography) and MRI (Magnetic Resonance Imaging) scans. Next, feature extraction is performed on the CT images. The CT medical image data is input into UNet, which uses downsampling to separate and detect vascular tissue in the vascular region, acquiring corresponding high-density signals. These high-density signals are then filtered to obtain density filtering signals, identifying the corresponding cerebral arteries and acquiring high-density features of the cerebral arteries. The CT medical image data represents the data acquired from the CT images. Then, feature parameters from the high-density features of the cerebral arteries are extracted, including the number of vessels, vessel diameter, and vessel length. Based on these feature parameters, the vascular load value corresponding to the CT scan image is calculated. For the MRI scan images, feature extraction is performed to obtain the thrombus shape coefficient. Next, the obtained vascular load value and thrombus shape coefficient are compared with the corresponding preset vascular load value and preset thrombus shape coefficient. Brain tissue regions with vascular load values ​​greater than the preset vascular load value and thrombus shape coefficients are marked as abnormal brain tissue regions in the brain neural image, and corresponding feature indicators are obtained. Based on the feature indicators, baseline labels are obtained for the brain tissue regions, and an abnormality index is obtained. Then, the abnormal brain tissue regions are located based on the abnormality index. Specifically, ischemic regions are identified in the abnormal brain tissue regions to obtain ischemic brain regions. Finally, the damaged neural tissue is classified into lesions based on the preset etiology identification network to determine the etiology: the image data corresponding to the damaged neural tissue is input into the input layer of the etiology identification network to obtain input early medical image data. The key features in the input early medical image data are extracted by the feature extraction layer of the network. The image data includes the data corresponding to the key features in the early medical image data. Then, the lesion classification layer classifies the selected features into lesions to achieve intelligent recognition of early medical image features.

[0003] For example, Chinese invention patent application CN114240935B discloses a method and apparatus for medical image feature recognition by spatial-frequency domain feature fusion, comprising: acquiring a medical image to be diagnosed; performing spatial image preprocessing on the medical image to be diagnosed to obtain a spatial standard medical image; performing frequency image preprocessing on the medical image to be diagnosed to obtain a frequency standard medical image; fusing the spatial standard medical image and the frequency standard medical image to obtain standard medical image features; inputting the standard medical image features into a trained convolutional neural network model to obtain a predicted value of the corresponding medical image to be diagnosed; if the predicted value is greater than a set diagnostic threshold, the diagnosis result is determined to be a positive example; otherwise, the diagnosis result is determined to be a negative example.

[0004] For example, Chinese invention patent application CN113240659B discloses a deep learning-based image feature extraction method, comprising: receiving first input data representing an unannotated image of a cardiac magnetic resonance imaging system from a cardiac magnetic resonance imaging system by a deep learning network; preprocessing the unannotated image by the deep learning network to generate second input data representing a salient image of the cardiac magnetic resonance imaging, and corresponding annotation data representing salient regions of the salient image; processing the first and second input data to perform training of the deep learning network by detecting target features in the unannotated image within the salient regions identified in the salient image; and using the trained deep learning network to process third input data to identify lesion structures in new unannotated images.

[0005] The above-mentioned technology has at least the following technical problems: In existing technologies, when extracting features from CT images using deep learning models, as the depth of feature extraction increases, in order to reduce the memory footprint of CT images in the deep learning model, CT images may be repeatedly subjected to convolution and pooling operations in the max pooling layer of the deep learning model in a downsampling manner. This reduces the spatial dimension of the CT images, leading to a decrease in CT image resolution. As a result, the fine structure of CT images, such as the sub-millimeter-level vessel wall texture of the middle cerebral artery, is aggregated into low-frequency, blurry blocks, causing the topological relationships between different vessels to become blurred. Consequently, the boundaries of the high-density signs of the middle cerebral artery cannot be accurately identified due to insufficient resolution, resulting in inaccurate identification of the high-density signs of the middle cerebral artery. This leads to the problem of inaccurate intelligent recognition of CT image features. Summary of the Invention

[0006] To address the technical problem of inaccurate intelligent recognition of CT image features in existing technologies, this invention provides a CT-based intelligent recognition system and method for early medical image features. The technical solution is as follows: On one hand, a CT-based intelligent recognition system for early medical image features is provided, comprising the following modules: a model memory usage monitoring module, a first-stage CT image monitoring module, and a second-stage CT image monitoring module. The model memory usage monitoring module is used to determine whether to perform model memory usage optimization based on the model memory usage monitoring results obtained during the feature extraction process of CT images using a deep learning model. Model memory usage optimization includes: CT image multi-threading operations to reduce the memory usage of CT images and CT image resolution evaluation to assess the pass / failability of model memory usage optimization. The first-stage CT image monitoring module is used to, after the model memory usage monitoring is deemed satisfactory, based on... The first-stage CT image monitoring module determines whether to perform first-stage CT image optimization based on the first-stage monitoring results. First-stage optimization includes spatial filtering to achieve a smooth transition in the number of pixels in CT image sub-blocks and temporal filtering to improve the structural integrity of CT image sub-blocks. The second-stage CT image monitoring module, after the first-stage monitoring is deemed satisfactory, determines whether the second-stage monitoring results meet the qualification criteria. If they do, a CT feature recognition accuracy assessment is performed; otherwise, an abnormal second-stage monitoring alert is sent to designated personnel.

[0007] On the other hand, a method for intelligent recognition of early medical image features based on CT is provided, including the following steps: during the feature extraction process of CT images, it is determined whether to optimize model memory usage based on the model memory usage monitoring results obtained by performing model memory usage monitoring; after the model memory usage monitoring is qualified, it is determined whether to optimize CT images in the first stage based on the CT image first stage monitoring results obtained by performing CT image first stage monitoring; after the CT image first stage monitoring is qualified, it is determined whether the CT image second stage monitoring qualification conditions are met based on the CT image second stage monitoring results obtained by performing CT image second stage monitoring. If they are met, a CT feature recognition accuracy assessment is performed; otherwise, an abnormal CT image second stage monitoring prompt is sent to a preset personnel.

[0008] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: 1. By performing model memory usage monitoring to obtain the monitoring results and determine whether model memory usage optimization is needed, it helps to accurately measure the model memory usage, reduce incomplete CT image processing caused by model memory overload, improve the integrity and reliability of CT images, and thus improve the stability of CT image feature recognition. By performing CT image first-stage monitoring to obtain the monitoring results and determine whether CT image first-stage optimization is needed, it helps to accurately measure the passability of CT image processing, reduce inaccurate CT images caused by noise and artifact interference, improve the accuracy and passability of CT images, and thus improve the reliability of CT image feature recognition. By performing CT image second-stage monitoring to obtain the monitoring results and determine whether the CT image second-stage monitoring passability conditions are met, it helps to accurately measure the passability of CT image processing in the first-stage monitoring, improve the coherence and accuracy of CT image processing, and thus improve the accuracy of CT image feature recognition.

[0009] 2. The first-stage monitoring results of CT images are obtained by harmonizing and averaging the contrast and resolution results of CT images and the model memory usage monitoring indicators. Compared with the existing technology, the spatial dimension of CT images is reduced, which leads to a decrease in CT image resolution and inaccurate fine structure of CT images. This solution helps to improve the balance between the accuracy and efficiency of CT image recognition, ensure the stability and reliability of CT image processing, and thus improve the qualification of CT image feature recognition.

[0010] 3. By harmonizing and averaging the first-stage monitoring indicators and the resolution density results of CT images, the second-stage monitoring results of CT images are obtained. Compared with existing technologies, which cannot accurately identify the middle cerebral artery high-density sign due to insufficient resolution, this solution helps to preserve and enhance the details of CT images, improve the accuracy and reliability of CT image recognition results, reduce the inaccuracy of CT image feature recognition due to CT image quality problems, and thus improve the accuracy of CT image feature recognition. Attached Figure Description

[0011] Figure 1 A schematic diagram of the structure of a CT-based intelligent medical early image feature recognition system provided in an embodiment of this application; Figure 2 A schematic diagram of the architecture of a CT-based intelligent recognition system for early medical image features is provided for an embodiment of this application. Figure 3 A schematic diagram illustrating the framework for optimizing model memory usage in a CT-based intelligent recognition system for early medical image features, provided in an embodiment of this application. Figure 4A flowchart illustrating an intelligent recognition method for early medical image features based on CT, provided for embodiments of this application; Figure 5 The present application provides an embodiment of a CT-based intelligent recognition system for early medical imaging features, which includes CT images of the brain's internal structure. Figure 6 This application provides an embodiment of a CT-based intelligent recognition system for early medical imaging features, which identifies the internal structure of brain CT images. Detailed Implementation

[0012] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0013] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0014] In the embodiments of this invention, the terms "image" and "picture" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning. Similarly, the terms "of," "corresponding (relevant)," and "corresponding" may sometimes be used interchangeably. It should be noted that, without emphasizing the distinction between them, they convey the same meaning.

[0015] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0016] This invention provides a CT-based intelligent recognition system and method for early medical image features, such as... Figure 1 The schematic diagram shown is of a CT-based intelligent medical early image feature recognition system, which includes the following modules: Model memory usage monitoring module: During the feature extraction process of CT images using the deep learning model, model memory usage monitoring is performed to obtain the monitoring results and determine whether the deep learning model meets the acceptable memory usage conditions. If it meets the conditions, the first stage of CT image monitoring is performed; otherwise, model memory usage optimization is performed. Model memory usage monitoring helps to accurately assess the memory usage when CT images are input into the model, ensuring the complete execution of the CT image processing workflow, thereby enhancing the integrity of CT image data and the reliability of CT image feature recognition, and providing a stable foundation for CT image processing.

[0017] The CT image first-stage monitoring module: After the model memory usage monitoring is qualified, the first-stage CT image monitoring is performed to obtain the first-stage CT image monitoring results and determine whether they meet the qualification conditions of the first-stage CT image monitoring. If they meet the conditions, the second-stage CT image monitoring is performed; otherwise, the first-stage CT image optimization is performed. The first-stage CT image monitoring helps to accurately evaluate the qualification of CT image processing, effectively suppress the interference of noise and artifacts on CT images, and improve the accuracy of CT image processing.

[0018] The second-stage CT image monitoring module: After the first-stage CT image monitoring is qualified, the second-stage CT image monitoring is performed to obtain the results of the second-stage CT image monitoring and to determine whether it meets the qualification conditions of the second-stage CT image monitoring. If it meets the conditions, the accuracy of CT feature recognition is evaluated; otherwise, an abnormal CT image monitoring prompt is sent to the preset personnel. The second-stage CT image monitoring helps to accurately evaluate the qualification of CT image processing in the first-stage CT image monitoring, which can effectively optimize the coherence and accuracy of CT image processing, thereby improving the accuracy of image feature extraction and recognition.

[0019] like Figure 2 The diagram shown is an architectural schematic of a CT-based intelligent recognition system for early medical image features provided in this application embodiment. First, model memory usage monitoring is performed to determine if the acquired model memory usage monitoring results meet the qualification conditions for deep learning model memory usage. If they do, the first stage of CT image monitoring is performed; otherwise, model memory usage optimization is performed. Then, the re-acquired model memory usage monitoring results after optimization are determined to meet the qualification conditions for deep learning model memory usage. If they do, the first stage of CT image monitoring is performed; otherwise, an abnormal model memory usage optimization notification is sent to a preset personnel. Next, the first stage of CT image monitoring is performed, and the acquired first stage CT image monitoring results are determined to meet the qualification conditions for the first stage of CT image monitoring. If the conditions are met, the second stage of CT image monitoring is performed; otherwise, the first stage of CT image optimization is performed to obtain qualified results for CT image sub-blocks. It is then determined whether the first stage monitoring results of the CT images re-obtained after the first stage optimization meet the qualified conditions for the first stage of CT image monitoring and whether the qualified results of the CT image sub-blocks are greater than the preset qualified results of the CT image sub-blocks. If so, the second stage of CT image monitoring is performed; otherwise, an abnormal prompt for the first stage of CT image optimization is sent to the preset personnel. Finally, the second stage of CT image monitoring is performed to determine whether the obtained second stage monitoring results meet the qualified conditions for the second stage of CT image monitoring. If so, the CT feature recognition accuracy assessment is performed; otherwise, an abnormal prompt for the second stage of CT image monitoring is sent to the preset personnel.

[0020] It should be added that, prior to the design of the CT-based intelligent recognition system for early medical image features, a database storing various preset data is established in this application. The database includes, but is not limited to, the capacity occupied by the preset deep learning model, the preset CT image resolution evaluation results, and the preset CT image first-stage monitoring results, etc., and the various values ​​are directly set by the technicians.

[0021] In this embodiment, the model memory usage monitoring module is used to obtain model memory usage monitoring results when performing model memory usage monitoring, in order to measure the model memory usage and reduce the level of model memory usage when CT images are input into the model. Next, the CT image first-stage monitoring module is used to perform CT image first-stage monitoring after the model memory usage monitoring is qualified, in order to obtain CT image first-stage monitoring results, in order to measure the overall clarity of the CT image and improve the resolution of the CT image, thereby improving the clarity of the CT image. Finally, the CT image second-stage monitoring module is used to perform CT image second-stage monitoring after the CT image first-stage monitoring is qualified, in order to obtain CT image second-stage monitoring results, in order to measure the fineness of the CT image features and improve the accuracy of CT image feature recognition.

[0022] Furthermore, the specific process for determining whether to perform model memory usage optimization based on the model memory usage monitoring results obtained from the execution model memory usage monitoring is as follows: After inputting a deep learning model, such as a deep neural network model, a convolutional neural network model, or 3D ResNet-50, into the CT images, it is determined whether the obtained model memory usage monitoring results meet the qualified conditions for deep learning model memory usage. The deep learning model's memory usage capacity during the preset model memory usage monitoring time period is monitored using the nvidia-smi tool, and its proportion is quantified by comparing it with the preset deep learning model's memory usage capacity, i.e., a ratio calculation, to obtain the model memory usage monitoring results, which reflect the deep learning model's memory usage status. The preset deep learning model's memory usage capacity is represented by the average value of the deep learning model's memory usage capacity over historical time periods. The preset model memory usage monitoring time period refers to the preset time period corresponding to the execution of model memory usage monitoring set in advance by the personnel. The qualified condition for deep learning model memory usage means that the model memory usage monitoring result is less than the preset model memory usage monitoring result, which is represented by the average value of the model memory usage monitoring results over historical time periods. If the model memory usage monitoring result meets the qualified condition for deep learning model memory usage, the first stage of CT image monitoring is performed; otherwise, model memory usage optimization is performed.

[0023] In this embodiment, performing model memory usage monitoring helps to accurately assess the actual memory usage of the model, enabling precise analysis of memory usage during model operation, reducing computational inefficiency or performance degradation caused by deep learning model memory overload, fundamentally improving the stability and execution efficiency of deep learning models, and thus improving the overall computational efficiency and utilization of CT images.

[0024] Furthermore, the specific process for optimizing model memory usage is as follows: First, perform multi-threaded CT image processing. This involves using a multi-threaded programming method to progressively increase the number of CT image sub-blocks with a preset ratio as the adjustment step size. This progressive increase helps balance processing speed and system stability, thereby reducing model memory usage while improving CT image processing speed. The number of CT image sub-blocks is less than the preset maximum number, which is directly set by the operators. The preset ratio of CT image sub-blocks is obtained by inputting model memory usage monitoring results and CT image volume into an image sub-block mapping set. This mapping set reflects the mapping relationship between model memory usage monitoring results, CT image memory usage capacity, and the corresponding preset ratio of CT image sub-blocks. This mapping set is obtained through repeated verification by operators based on historical experience and represents the acceptable level of the number of CT image sub-blocks. Multi-threaded CT image processing is performed to improve CT image processing efficiency and reduce CT image memory usage. The first step is to assess the acceptable level of memory usage. The second step is to evaluate the CT image resolution. The specific process for evaluating the CT image resolution is as follows: First, determine whether the acquired CT image resolution evaluation result is greater than the preset CT image resolution evaluation result. The preset CT image resolution evaluation result is represented by the average value of CT image resolution evaluation results over a historical time period. Second, determine whether the acquired model memory usage monitoring result meets the acceptable conditions for deep learning model memory usage. Third, monitor the CT image resolution over the preset CT image resolution evaluation time period using an image measuring instrument, and quantify its proportion with the preset CT image resolution to obtain the CT image resolution evaluation result, which reflects the acceptable level of CT image resolution. The preset CT image resolution is represented by the average value of CT image resolution over a historical time period. Fourth, if the re-acquired CT image resolution evaluation result after model memory usage optimization is not less than the preset CT image resolution evaluation result, and the model memory usage monitoring result meets the acceptable conditions for deep learning model memory usage, then the first stage of CT image monitoring for evaluating the degree of interference in CT image processing is executed. Otherwise, an abnormal prompt for model memory usage optimization is sent to the designated personnel.

[0025] It should be added that, such as Figure 3The diagram shows a framework for optimizing model memory usage in a CT-based intelligent recognition system for early medical image features, as provided in this application embodiment. First, model memory usage monitoring is performed. It is determined whether the acquired model memory usage monitoring results meet the qualification conditions for deep learning model memory usage. If they do, the first stage of CT image monitoring is performed; otherwise, model memory usage optimization is performed. This optimization includes performing multi-threaded CT image operations and CT image resolution evaluation. It is determined that the re-acquired CT image resolution evaluation result after model memory usage optimization is not less than the preset CT image resolution evaluation result, and the model memory usage monitoring results meet the qualification conditions for deep learning model memory usage. If they do, the first stage of CT image monitoring is performed; otherwise, a model memory usage optimization anomaly alert is sent to a preset personnel.

[0026] In this embodiment, optimizing model memory usage helps improve the integrity of CT image processing, enabling fine-grained control over CT image memory usage. This balances the relationship between computational resources and memory consumption, improves the efficiency of parallel CT image processing, and ensures that key image quality features are not sacrificed during memory optimization, thereby enhancing the reliability and effectiveness of the CT image input model. Simultaneously, this enhances the stability of model operation and the overall efficiency of CT image processing.

[0027] Furthermore, the specific process of performing the first stage of CT image monitoring is as follows: It is determined whether the acquired first-stage CT image monitoring results meet the qualification criteria for the first stage of CT image monitoring; if the first-stage CT image monitoring results meet the qualification criteria, then the second stage of CT image monitoring, used to verify the qualification level of CT image processing during the first stage of CT image monitoring, is performed; otherwise, the first stage of CT image optimization is performed; the first-stage CT image monitoring results are represented by the harmonic average of the CT image contrast and resolution results and the model memory usage monitoring index, used to reflect the overall interference level of the CT image. When the noise in the input CT image increases, in order to maintain or improve the reconstruction quality, the model... Memory consumption grows non-linearly, meaning that the model memory usage monitoring index increases, leading to a decrease in model memory space. The smaller the memory usage of the transmitted CT images, the more noise appears in the CT images, thus increasing the CT image contrast and resolution results. The qualified condition for the first stage of CT image monitoring is that the first stage monitoring result is lower than the preset first stage monitoring result, which is represented by the average of the first stage monitoring results over a historical period. The CT image contrast and resolution results are obtained by weighting the CT image contrast and resolution index with the resolution control rate; that is, assigning different weights to multiple variables according to the priorities set by the preset personnel. The result is expressed through CT... The performance testing phantom monitors the contrast resolution of CT images during the first phase of monitoring, and quantifies it by comparing it with the preset CT image contrast resolution. This ratio is used to obtain the CT image contrast resolution index, which reflects the degree of interference in CT image contrast resolution. The preset first phase of CT image monitoring time period refers to the preset time period corresponding to the first phase of CT image monitoring set in advance by the personnel. The model memory usage monitoring index is represented by the result of weighted processing of qualified model memory usage monitoring results and memory regulation rate, which reflects the impact of deep learning model memory usage. Qualified model memory usage monitoring results are represented by the model memory usage monitoring results corresponding to the qualified deep learning model memory usage conditions. The memory regulation rate is used to reflect the degree of influence of the model memory usage monitoring index on the first phase of CT image monitoring results, and the resolution regulation rate is used to reflect the degree of influence of CT image contrast resolution results on the first phase of CT image monitoring results.

[0028] It should be added that the CT image contrast resolution results and model memory usage monitoring indicators are input into a mapping set. Specifically, the CT image contrast resolution results and model memory usage monitoring indicators are normalized using a data normalization algorithm. The CT image signal-to-noise ratio (SNR) is then input into a linear regression model, and the corresponding resolution control rate and memory control rate are obtained through fitting using a linear regression algorithm. Finally, the ranges of the CT image SNR, CT image contrast resolution results, and model memory usage monitoring indicators are mapped to preset intervals, which are pre-set by designated personnel, thus establishing the mapping set. In this embodiment, the resolution control rate and memory control rate both range from 0 to 1. Since more noise in the CT image means a higher CT image contrast resolution result, the greater the power of the CT image noise, resulting in a lower CT image SNR, fewer photons received by the CT image pixel unit, and a weakened signal, leading to a decrease in CT image resolution and consequently a decrease in the model memory usage monitoring indicators.

[0029] In this embodiment, performing the first-stage monitoring of CT images helps to accurately assess the qualification of the first-stage monitoring of CT images, realize a multi-dimensional assessment of the stability of the initial stage of CT image processing, reduce the impact of noise interference on the clarity of CT images, improve the qualification of CT images, and provide a stable CT image foundation for the subsequent processing stages of CT images, thereby ensuring the continuity, robustness and reliability of the CT image processing process as a whole.

[0030] Furthermore, the first stage of CT image optimization involves both spatial and temporal filtering operations. The specific process for spatial filtering is as follows: a pre-defined CT image sub-block corresponding to a pre-set CT image center point is used as the central CT image sub-block. Based on the pre-set distance between the pre-set CT image center point and the center point of a secondary CT image sub-block, a pre-set weight is assigned to the secondary CT image sub-block. The number of pixels in the secondary CT image sub-block, monitored by the image sensor, is then weighted with the pre-set weights to obtain the image sub-block pixel result. The secondary CT image sub-block represents any CT image sub-block other than the central CT image sub-block. CT image sub-blocks with pixel results not less than a pre-set pixel result are marked as first qualified CT image sub-blocks; otherwise, they are marked as first abnormal CT image sub-blocks. The pre-set pixel result is represented by the average pixel result of image sub-blocks over a historical time period. Spatial filtering is performed to suppress noise interference in the CT images, thereby improving their qualification.

[0031] In this embodiment, performing first-stage optimization of CT images helps improve the quality of CT images, reduce the incoherence of CT images caused by noise and artifacts, suppress the interference of random noise and local abnormalities in CT images on the overall image quality, thereby improving the signal-to-noise ratio and visual consistency of CT images, enhancing the usability and quality of CT images, improving the qualification and reliability of CT image processing, and thus providing a stable medical image foundation for subsequent CT image processing stages.

[0032] Furthermore, the specific process of performing temporal filtering on CT images is as follows: The obtained preset CT image sub-block pixel intensity characterization value and the CT image sub-block pixel intensity characterization value are proportionally quantified to obtain the image pixel intensity result used to reflect the degree of change of the CT image sub-block. The preset CT image sub-block pixel intensity characterization value is represented by the average value of the CT image sub-block pixel intensity characterization values ​​over a historical time period. The contrast of each CT image sub-block at the end of the CT image temporal filtering operation is monitored by the high contrast resolution test module, and its average value is represented as the CT image sub-block pixel intensity characterization value. For example, for ischemic stroke in neurology, the powerful feature extraction capability of 3D ResNet-50 can be used in combination with the "bone window" display technology commonly used in CT images. The deep learning model first preprocesses the original CT images to simulate the bone window view commonly used in clinical practice, highlighting the contrast of high-density structures such as blood vessels. To ensure smooth information transmission from input to output, the system accurately captures subtle high-density signals from the middle cerebral artery. For example, in an emergency cerebral hemorrhage screening scenario, the system can quickly locate the responsible vessel by analyzing abnormally bright areas in bone window mode. CT image sub-blocks with pixel intensity results lower than the preset pixel intensity result are marked as the second qualified CT image sub-blocks; conversely, those with higher pixel intensity results are marked as the second abnormal CT image sub-blocks. CT image temporal filtering is performed to smooth noise and improve the clarity of CT image sub-blocks. CT image sub-blocks that are both the first and second qualified CT image sub-blocks are marked as qualified CT image sub-blocks and undergo CT image equalization adjustment. CT image sub-blocks other than qualified CT image sub-blocks are marked as unqualified CT image sub-blocks and undergo CT image filtering. Both CT image equalization adjustment and CT image filtering are marked as CT image distribution adjustment. The preset pixel intensity result is represented by the average pixel intensity result over a historical time period.

[0033] In this embodiment, performing temporal filtering on CT images helps to distinguish between the real tissue structure and transient noise interference in CT images, identify and suppress abnormal pixel changes in CT images caused by random noise, improve the temporal consistency and visual coherence within CT image sub-blocks, thereby enhancing the stability constraints in the temporal domain, providing a high-quality CT image foundation for subsequent CT image processing, and thus improving the temporal stability and diagnostic usability of CT images.

[0034] Furthermore, the specific process of CT image sub-block adjustment is as follows: CT image filtering involves denoising sub-blocks of sub-blocks using bilateral filtering to preserve edge information, thereby improving the clarity of the sub-blocks; CT image equalization involves enhancing the resolution of sub-blocks of sub-blocks of sub-blocks based on adaptive histogram equalization. For example, when processing sub-blocks of sub-blocks of small blood vessels or soft tissue boundaries, adaptive histogram equalization adaptively remaps grayscale values ​​according to pixel intensity distribution characteristics to enhance local contrast. If a sub-block of sub-blocks is dark and has compressed dynamic range due to equipment noise interference, adaptive histogram equalization will prioritize increasing the gradient change rate in the central region of the sub-block, making the originally blurred blood vessel wall contours sharp and discernible, ensuring that the enhanced CT image can highlight the resolution. While maintaining structural details, natural texture features can also be preserved; CT image equalization adjustments are performed to enhance the grayscale differences in local areas within CT image sub-blocks to reduce background noise; the proportion of qualified CT image sub-blocks obtained after the first stage of CT image optimization is quantified to reflect the qualification level of CT image sub-blocks; if the first stage monitoring results of CT images re-obtained after the first stage of CT image optimization meet the qualification conditions of the first stage of CT image monitoring and the qualified CT image sub-block results are greater than the preset qualified CT image sub-block results, then the second stage of CT image monitoring is executed; otherwise, an abnormal CT image optimization prompt is sent to the preset personnel. For example, for ischemic stroke in neurology, segmentation parameters can be optimized in real time based on the statistical characteristics of CT image sub-blocks (such as histogram distribution and gradient changes). By combining multi-scale contextual information extracted from 3D ResNet-50, the deep learning model can intelligently distinguish the boundary regions between gray and white matter in normal brain tissue, while suppressing artifacts caused by field inhomogeneity. Because these boundary regions have complex anatomical structures and small volumes, the deep learning model adds refined branch networks targeting the cerebellar dentate nucleus and brainstem reticular formation. These sub-networks inherit the pre-trained weights of the main network but focus on morphological features of specific regions through transfer learning, such as the cerebellar interlobular sulcus pattern and signal intensity differences in brainstem nuclei, providing early biomarkers for neurodegenerative diseases.

[0035] In this embodiment, performing CT image filtering and CT image equalization helps to implement differentiated processing for CT image sub-blocks in different states, thereby achieving a fine balance between noise suppression and detail enhancement, preserving the edge structure features of CT image sub-blocks, reducing the edge blurring problem of CT image sub-blocks caused by traditional filtering methods, reducing the increase in background noise caused by global CT image processing, ensuring the quality and stability of CT images, and thus improving the reliability of CT image processing and the qualification of output results.

[0036] Furthermore, the specific process of performing the second-stage CT image monitoring is as follows: First, determine whether the acquired second-stage CT image monitoring results meet the qualification criteria for second-stage CT image monitoring. The second-stage CT image monitoring results are represented by a harmonic average of the first-stage CT image monitoring indicators and the CT image resolution density results. This is used to verify the qualification level of the CT images after the first-stage monitoring. Because the fixed total number of photons in the CT image is distributed to more pixels, the signal-to-noise ratio of each pixel decreases, and noise becomes more obvious. This indicates an increase in the first-stage CT image monitoring indicators, meaning a smaller number of CT image pixels, resulting in lower CT image resolution and thus lower CT image resolution density. The qualification criteria for second-stage CT image monitoring indicate that the second-stage CT image monitoring results are greater than the preset second-stage CT image monitoring results. The first-stage CT image monitoring indicators are represented by a weighted processing result combining the qualified first-stage CT image monitoring results with the first monitoring result adjustment rate. This reflects the impact of the first-stage CT image monitoring on the second-stage CT image monitoring. The first monitoring result adjustment rate reflects the degree of influence of the first-stage CT image monitoring indicators on the second-stage CT image monitoring results. The qualified first-stage CT image monitoring... The results are represented by the first-stage monitoring results of CT images that meet the qualification conditions for the first stage of CT image monitoring. The resolution density of CT images during the preset second-stage monitoring period is monitored using a CT experimental detector. This resolution density is then quantified as a proportion of the preset CT image resolution density, and weighted using a resolution density adjustment rate to obtain the final CT image resolution density result. The resolution density adjustment rate reflects the degree of influence of the CT image resolution density result on the second-stage monitoring results. If the second-stage monitoring results meet the qualification conditions, the corresponding CT images are input into a deep learning model to perform a CT feature recognition accuracy assessment, reflecting the accuracy of the output CT image feature recognition. For example, CT images are randomly divided into training and validation sets. The training set is input into a random forest model for training to obtain a training model. Newly acquired CT images are input into the training model to output the number of features recognized by the first model. Multivariate logistic regression analysis is used to screen statistically significant evaluation indicators. Based on the screened evaluation indicators, a preset specific AI training optimization model is constructed and validated. For example, the regression equation is established as: logit(P)=β0+β1×X1+β2×X2+…+β n ×X n Where P is the dynamic scene feature coverage, logit(P) represents the number of features recognized by the first model, and X1, X2…X n The selected indicators are β0, which is a constant term, and β1, β2, ..., β... nThe regression coefficient is denoted by n, which represents the number of selected evaluation indicators. The area under the curve corresponding to the dynamic data in the calculation scenario is analyzed using ROC (Receiver Operating Characteristic Curve). The closer the area under the curve is to 1, the higher the training value of the model. In the validation set, the model is trained and optimized using a preset specific AI and validated. The stability and accuracy of the deep learning model are verified again by ROC curve analysis. If the validation result is good, the deep learning model is considered qualified. Otherwise, an abnormality alert for the second stage of CT image monitoring is sent to the preset personnel.

[0037] It should be added that the first-stage monitoring indicators of CT images and the resolution density results of CT images are respectively input into the mapping set. The mapping set is a result representation constructed by preset personnel through preset mapping relationships to map the first-stage monitoring indicators of CT images and the resolution density results of CT images to the corresponding first monitoring result control rate and resolution density control rate. In this embodiment, the values ​​of the first monitoring result control rate and resolution density control rate are both in the range of 0 to 1.

[0038] In this embodiment, performing the second-stage monitoring of CT images helps to accurately assess the qualification of the second-stage monitoring of CT images, reduce misjudgment of CT image features due to the lack of CT image details, reduce the loss of CT image feature details due to noise suppression, filtering or compression, ensure that the images have sufficient spatial resolution and feature expression capabilities, and improve the high reliability and trustworthiness of CT images.

[0039] Furthermore, the specific process of performing CT feature recognition accuracy assessment is as follows: It is determined whether the obtained CT feature recognition accuracy assessment result conforms to the CT feature recognition accuracy assessment result; the CT feature recognition accuracy assessment result is obtained by quantifying the ratio between the spatial resolution of CT images during the preset CT feature recognition accuracy assessment time period and the preset CT image spatial resolution using a CT experimental detector. This ratio reflects the accuracy of CT image feature recognition. The preset CT image spatial resolution is represented by the average value of CT image spatial resolution over historical time periods. The preset CT feature recognition accuracy assessment time period refers to the preset time period corresponding to the CT feature recognition accuracy assessment set in advance by preset personnel; the CT image spatial resolution represents the minimum line pair spacing of CT images that the deep learning model can resolve; if the CT feature recognition accuracy assessment result conforms to the CT feature recognition accuracy assessment result, a CT image feature recognition qualified prompt is sent to the preset personnel; otherwise, a CT image feature recognition abnormality prompt is sent to the preset personnel.

[0040] In this embodiment, performing a CT feature recognition accuracy assessment helps to accurately evaluate the precision of CT image feature recognition, ensure CT image quality, thereby ensuring the reliability of subsequent CT image analysis, avoiding misjudgment or missed detection of CT image features due to insufficient CT image resolution, and thus improving the quality control and reliability of the entire CT image processing.

[0041] like Figure 4 The diagram shown is a flowchart of an intelligent recognition method for early medical image features based on CT, provided in an embodiment of this application. Model memory usage monitoring: During the feature extraction process of CT images, the model memory usage monitoring results obtained are used to determine whether to optimize model memory usage.

[0042] CT image first-stage monitoring: After the model memory usage monitoring is qualified, the first-stage monitoring results of CT images obtained by performing the first-stage monitoring of CT images determine whether to perform the first-stage optimization of CT images.

[0043] Second-stage CT image monitoring: After the first-stage CT image monitoring is qualified, the second-stage CT image monitoring results are used to determine whether the second-stage CT image monitoring is qualified. If it is qualified, the CT feature recognition accuracy assessment is performed; otherwise, an abnormal CT image monitoring prompt is sent to the preset personnel.

[0044] like Figure 5 The image shown is a CT image of the internal brain structure provided by an intelligent recognition system for early medical imaging features based on CT, according to an embodiment of this application. Figure 6 The image shown is a CT image for identifying the internal brain structure using a CT-based intelligent recognition system for early medical imaging features, provided in an embodiment of this application. The CT image includes ventricles, sulci, and gyri. The central bright area is part of a ventricle. Figure 5 compared to, Figure 6 Low-density shadow areas have been identified in CT images and marked with a red box indicating hypoattenuation. This indicates that the absorption of X-rays or other radiation from a plain head CT scan is lower than the normal range, which may suggest insufficient local blood perfusion.

[0045] In summary, the embodiments of this application, by performing model memory usage monitoring to obtain the monitoring results and determining whether to optimize model memory usage, help to accurately measure the model memory usage, reduce incomplete CT image processing caused by model memory overload, improve the integrity and reliability of CT images, and thus improve the stability of CT image feature recognition. By performing first-stage CT image monitoring to obtain the first-stage monitoring results and determining whether to optimize CT images, it helps to accurately measure the passability of CT image processing, reduce inaccurate CT images caused by noise and artifact interference, improve the accuracy and passability of CT images, and thus improve the reliability of CT image feature recognition. By performing second-stage CT image monitoring to obtain the second-stage monitoring results and determining whether they meet the passability conditions of second-stage CT image monitoring, it helps to accurately measure the passability of CT image processing in the first-stage monitoring, improve the coherence and accuracy of CT image processing, and thus improve the accuracy of CT image feature recognition.

[0046] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A CT-based medical early image feature intelligent identification system, characterized in that, Comprise the following modules: model memory occupation monitoring module, CT image first stage monitoring module and CT image second stage monitoring module; The model memory occupation monitoring module is used to judge whether to perform model memory occupation optimization based on the model memory occupation monitoring result obtained by executing model memory occupation monitoring in the process of feature extraction of CT image by deep learning model, and the model memory occupation optimization includes CT image multi-thread operation for reducing the memory usage of CT image and CT image resolution evaluation for evaluating the qualification of model memory occupation optimization; The CT image first stage monitoring module is used to judge whether to perform CT image first stage optimization based on the CT image first stage monitoring result obtained by executing CT image first stage monitoring after the model memory occupation monitoring is qualified, and the CT image first stage optimization includes CT image spatial filtering operation for realizing smooth transition of pixel point number of CT image sub-block and CT image time filtering operation for improving the structural integrity of CT image sub-block; The CT image second stage monitoring module is used to judge whether to meet the CT image second stage monitoring qualified condition based on the CT image second stage monitoring result obtained by executing CT image second stage monitoring after the CT image first stage monitoring is qualified, and if so, CT feature recognition precision evaluation is performed, otherwise, CT image second stage monitoring exception prompt is sent to the preset personnel. 2.The CT-based medical early image feature intelligent recognition system according to claim 1, characterized in that, The specific process of judging whether to perform model memory occupation optimization based on the model memory occupation monitoring result obtained by executing model memory occupation monitoring is as follows: After the CT image is input into the deep learning model, it is judged whether the obtained model memory occupation monitoring result meets the deep learning model memory occupation qualified condition; The model memory occupation monitoring result is represented by the result of proportioning quantization of the deep learning model occupation capacity and the preset deep learning model occupation capacity, and is used to reflect the deep learning model memory occupation situation; The deep learning model memory occupation qualified condition indicates that the model memory occupation monitoring result is less than the preset model memory occupation monitoring result; If the model memory occupation monitoring result meets the deep learning model memory occupation qualified condition, CT image first stage monitoring is performed, otherwise, model memory occupation optimization is performed. 3.The CT-based medical early image feature intelligent identification system according to claim 2, characterized in that, The specific process of performing model memory occupation optimization is as follows: First step, CT image multi-thread operation is performed; The CT image multi-thread operation indicates that the CT image sub-block number is increased step by step with the amplitude corresponding to the preset CT image sub-block number ratio as the adjustment step; The preset CT image sub-block number ratio is obtained by inputting the model memory occupation monitoring result and the CT image volume into the image sub-block mapping set; The CT image multi-thread operation is used to improve the CT image processing efficiency to improve the qualification of reducing the memory occupation of CT image; Second step, CT image resolution evaluation is performed; The specific process of performing CT image resolution evaluation is as follows: The acquired CT image resolution evaluation result is greater than the preset CT image resolution evaluation result, and the acquired model memory occupation monitoring result meets the deep learning model memory occupation qualified condition; The CT image resolution evaluation result is used to reflect the qualified degree of the CT image resolution; If the CT image resolution evaluation result obtained after the model memory occupation optimization is not less than the preset CT image resolution evaluation result, and the model memory occupation monitoring result meets the deep learning model memory occupation qualified condition, the CT image first stage monitoring for evaluating the CT image processing interference degree is performed, otherwise, the model memory occupation optimization exception prompt is sent to the preset personnel. 4.The CT-based medical early image feature intelligent identification system according to claim 2, characterized in that, The specific process of performing the CT image first stage monitoring is as follows: The acquired CT image first stage monitoring result is greater than the CT image first stage monitoring qualified condition; If the CT image first stage monitoring result meets the CT image first stage monitoring qualified condition, the CT image second stage monitoring for verifying the CT image processing qualified degree in the CT image first stage monitoring process is performed, otherwise, the CT image first stage optimization is performed; The CT image first stage monitoring result is represented by the result of the harmonic average processing of the CT image contrast resolution result and the model memory occupation monitoring index, and is used to reflect the interference degree of the overall CT image; The CT image contrast resolution result is used to reflect the interference degree of the CT image contrast resolution; The model memory occupation monitoring index is used to reflect the influence of the deep learning model memory occupation. 5.The CT-based medical early image feature intelligent identification system according to claim 4, characterized in that, The CT image first stage optimization is performed, which means that the CT image spatial filtering operation and the CT image time filtering operation are performed at the same time. The specific process of performing the CT image spatial filtering operation is as follows: The image sub-block pixel result is obtained; The image sub-block pixel result is greater than the preset image sub-block pixel result, and the corresponding CT image sub-block is marked as the first qualified CT image sub-block, otherwise, it is marked as the first abnormal CT image sub-block; The CT image spatial filtering operation is used to suppress the noise interference of the CT image to improve the qualification of the CT image. 6.The CT-based medical early image feature intelligent identification system according to claim 5, characterized in that, The specific process of performing the CT image time filtering operation is as follows: The acquired preset CT image sub-block pixel intensity representation value is quantified by the proportion with the CT image sub-block pixel intensity representation value to obtain the image pixel intensity result for reflecting the change degree of the CT image sub-block; The image pixel intensity result is less than the preset image pixel intensity result corresponding to the CT image sub-block, and the corresponding CT image sub-block is marked as the second qualified CT image sub-block, otherwise, it is marked as the second abnormal CT image sub-block; The CT image time filtering operation is used to smooth the noise to improve the clarity of the CT image sub-block; The CT image sub-block which is both the first qualified CT image sub-block and the second qualified CT image sub-block is marked as the qualified CT image sub-block and is subjected to the CT image equalization adjustment; The CT image sub-block corresponding to the qualified CT image sub-block is marked as the unqualified CT image sub-block and is subjected to the CT image filtering operation. The CT image equalization adjustment and the CT image filtering operation are marked as CT image sub-block adjustment. 7.The CT-based medical early image feature intelligent recognition system according to claim 6, characterized in that, The specific process of the CT image sub-block adjustment is as follows: The CT image filtering operation means that the unqualified CT image sub-block is denoised to retain the edge information of the CT image sub-block, thereby improving the definition of the CT image sub-block. The CT image equalization adjustment means that the qualified CT image sub-block is subjected to CT image resolution enhancement. The CT image equalization adjustment is used to enhance the gray difference of the local area in the CT image sub-block to reduce the background noise of the CT image sub-block. The number of qualified CT image sub-blocks obtained after the first-stage CT image optimization is quantified by the proportion of the number of unqualified CT image sub-blocks to obtain a CT image sub-block qualification result reflecting the qualification degree of the CT image sub-block. If the first-stage CT image monitoring result re-acquired after the first-stage CT image optimization meets the first-stage CT image monitoring qualification condition and the CT image sub-block qualification result is greater than the preset CT image sub-block qualification result, the second-stage CT image monitoring is performed, otherwise, a first-stage CT image optimization exception prompt is sent to the preset personnel. 8.The CT-based medical early image feature intelligent identification system according to claim 7, characterized in that, The specific process of the second-stage CT image monitoring is as follows: Determine whether the acquired second-stage CT image monitoring result meets the second-stage CT image monitoring qualification condition; The second-stage CT image monitoring result is a result of harmonic average processing of the first-stage CT image monitoring index and the CT image resolution density result, and is used to verify the qualification degree of the CT image after the first-stage CT image monitoring; The second-stage CT image monitoring qualification condition means that the second-stage CT image monitoring result is greater than the preset second-stage CT image monitoring result; The first-stage CT image monitoring index is a result of empowerment processing of the qualified first-stage CT image monitoring result combined with the first monitoring result control rate, and is used to reflect the influence of the first-stage CT image monitoring on the second-stage CT image monitoring; The first monitoring result control rate is used to reflect the influence degree of the first-stage CT image monitoring index on the second-stage CT image monitoring result; The CT image resolution density result is a result of empowerment processing of the CT image resolution density combined with the resolution density control rate after proportion quantization of the CT image resolution density and the preset CT image resolution density; The resolution density control rate is used to reflect the influence degree of the CT image resolution density result on the second-stage CT image monitoring result; If the second-stage CT image monitoring result meets the second-stage CT image monitoring qualification condition, the corresponding CT image is input into the deep learning model to perform CT feature recognition accuracy evaluation reflecting the feature recognition accuracy of the output CT image, otherwise, a second-stage CT image monitoring exception prompt is sent to the preset personnel. 9.The CT-based medical early image feature intelligent identification system according to claim 8, characterized in that, The specific process of the CT feature recognition accuracy evaluation is as follows: Determine whether the acquired CT feature recognition accuracy evaluation result meets the CT feature recognition accuracy evaluation result; The CT feature recognition accuracy evaluation result is used to reflect the accuracy of the CT image feature recognition; If the CT feature recognition accuracy evaluation result is consistent with the CT feature recognition accuracy evaluation result, a CT image feature recognition qualified prompt is sent to a preset person, otherwise, a CT image feature recognition exception is sent to the preset person.

10. A method for dynamic specificity AI training optimization based on scene characteristics, characterized in that, The application is applied to the CT-based medical early image feature intelligent recognition system in any one of claims 1-9, and is characterized by comprising the following steps: In the feature extraction process of the CT image, it is judged whether to perform model memory occupation optimization based on the model memory occupation monitoring result obtained based on the model memory occupation monitoring; After the model memory occupation monitoring is qualified, it is judged whether to perform CT image first stage optimization based on the CT image first stage monitoring result obtained based on the CT image first stage monitoring; After the CT image first stage monitoring is qualified, it is judged whether to meet the CT image second stage monitoring qualified condition based on the CT image second stage monitoring result obtained based on the CT image second stage monitoring, if yes, the CT feature recognition accuracy evaluation is performed, otherwise, a CT image second stage monitoring exception prompt is sent to a preset person.

Images