Chest low-dose CT scanning parameter optimization system and method based on neural network

Abstract

The invention belongs to the technical field of medical images, and particularly relates to a chest low-dose CT scanning parameter optimization system and method based on a neural network. A chest low-dose CT scanning parameter optimization method based on a neural network comprises the following steps that S10, pulmonary nodules of different sizes are implanted into an in-vitro model of the liver, different scanning parameters are set for scanning devices of different models, LDCT scanning is conducted on the in-vitro model to obtain an LDCT image, the LDCT image is scored through a scoring module, and finally the LDCT image is obtained; qualified images are selected from the LDCT images according to the scoring result, scanning equipment models, scanning parameters, nodule features and radiation doses corresponding to the qualified images are acquired, and an ANN model is established; and S20, obtaining an in-vitro model corresponding to the qualified image as an in-vitro initial model, and generating a seed number through a random algorithm. According to the scheme, the multivariable coupling ANN model is constructed, personalized parameters are output before scanning, equipment parameters are automatically adapted, and the problem that the existing radiation dose estimation method cannot adjust the dose according to the individual patient before scanning is solved.

Description

Technical Field

[0001] This solution belongs to the field of medical imaging technology, specifically involving a system and method for optimizing parameters of low-dose chest CT scans based on neural networks. Background Technology

[0002] Lung cancer is the leading cause of cancer-related morbidity and mortality worldwide and in my country, and low-dose CT (LDCT) of the chest is an important means of early screening. However, precise control of radiation dose during CT scanning has always been a research hotspot and challenge.

[0003] Currently, although international and domestic guidelines have been established for CT radiation dose diagnosis—for example, the American College of Radiology recommends a volumetric CT dose index of 3 Gy or lower and an effective dose of 1 mSv or lower for low-dose lung cancer screening—these are only guidelines or reference ranges and cannot achieve precise, personalized dose control. In practical applications, radiation dose distribution varies significantly among different institutions and patients, and exceeding the standard is common.

[0004] Existing radiation dose estimation methods, such as body type-specific dose estimation based on water equivalent diameter (SSDE), are available. WED While existing methods can accurately estimate radiation dose, they are post-scan calculations and cannot adjust the dose based on individual patient factors before the scan. Individual factors such as the patient's body type (BMI, weight, chest circumference) and tissue composition (fat content) are closely related to radiation dose, but current methods struggle to fully integrate these variables for personalized scanning. Therefore, a technological solution is urgently needed that can personalize the dose based on individual patient factors before the scan. Summary of the Invention

[0005] The purpose of this solution is to provide a neural network-based optimization system and method for low-dose chest CT scan parameters to address the problem that existing radiation dose estimation methods cannot adjust the dose according to individual patient factors before scanning.

[0006] To achieve the above objectives, this solution provides a method for optimizing parameters in low-dose chest CT scans based on neural networks, comprising the following steps:

[0007] S10: Implant lung nodules of different sizes into an in vitro liver model, set different scanning parameters for different models of scanning equipment to perform LDCT scans on the in vitro model to obtain LDCT images, use a scoring module to score the LDCT images, select qualified images from the LDCT images based on the scoring results, obtain the corresponding scanning equipment model, scanning parameters, nodule characteristics and radiation dose of the qualified images, and establish an ANN model.

[0008] S20: Obtain the in vitro model corresponding to the qualified image as the initial in vitro model, generate the seed number through a random algorithm, and use the seed number to modify the lung nodule parameters in the initial in vitro model. Use the modified lung nodule parameters as the initial test parameters, use the ANN model to scan the initial in vitro model to obtain LDCT images, and identify the lung nodule size parameters as the second test parameters. Compare the initial test parameters and the second test parameters, and use the comparison results as feedback information to optimize the ANN model.

[0009] S30: The ANN model is trained and optimized by taking the individual factors of the examinee and the model of the scanning equipment as inputs, and the radiation dose and scanning parameters of the examinee as outputs; the normality test is performed on the input data of the ANN model, and the univariate independent variables are screened from the input data based on correlation analysis and one-way ANOVA, and then the key variables are selected from the univariate independent variables using random forest.

[0010] S40: Input the patient's individual factors and the scanning equipment model into the trained ANN model to obtain personalized scanning parameters and recommended radiation dose.

[0011] And, a neural network-based optimization system for low-dose chest CT scan parameters, which uses a neural network-based optimization method for chest low-dose CT scan parameters.

[0012] The principle and technical effect of this solution are as follows: This solution constructs a multivariate coupled ANN model, which nonlinearly maps the individual characteristics of the examinee, the equipment parameters and the SSDED dose database. It uses the feedforward prediction principle to output personalized parameters before scanning. Through feature hierarchical screening, it ensures that key variables dominate the decision. Finally, through dynamic iteration, the model is continuously optimized to achieve synergistic optimization of radiation dose reduction and lesion detection rate improvement. At the same time, it has cross-equipment stability and clinical adaptability (automatic retraining when deviation exceeds the limit).

[0013] Simultaneously, by learning from a multi-device parameter library through neural networks, a matching model between the characteristics of scanning devices and optimal parameters is established. This enables automatic parameter adaptation for different brands of scanning devices, eliminating the need for doctors to manually adjust for device differences. This solution automatically outputs parameter combinations optimized for specific models, improving the consistency of image quality across devices. Furthermore, this solution introduces a scoring module as a constraint on the loss function during model training. By balancing the nonlinear relationship between radiation dose and image noise, it achieves more efficient dose allocation while ensuring diagnostic needs are met, thereby improving the image CNR (contrast-to-noise ratio) of the model output parameters at the same dose level.

[0014] In summary, this solution constructs a multivariate coupled ANN model to achieve personalized parameter output before scanning and automatic adaptation of equipment parameters, thus solving the problem that existing radiation dose estimation methods cannot adjust the dose according to the individual patient before scanning.

[0015] Furthermore, in step S10, when implanting lung nodules into the in vitro model, simulated lung nodules with diameters of 3mm and 5mm are implanted into the PGH-1 in vitro model in stages according to lung segments; the different types of scanning equipment include at least 5 types; the scanning parameters include tube voltage, tube current, pitch, and gantry rotation speed; the individual factors of the examinee include gender, age, height, weight, blood type, fat content within the scanning range, longitudinal and transverse diameters of the thoracic cavity at the largest / central level, and the number of scanning layers.

[0016] First, by limiting the size of lung nodules to 3mm and 5mm and implanting them according to lung segment anatomy, this approach not only makes the in vitro model more closely resemble clinical reality but also significantly improves the detection sensitivity of nodules in different lung segments. Second, requiring training data from at least five CT equipment models ensures the model's broad applicability while keeping cross-equipment dose differences within a low range. Simultaneously, it comprehensively covers core scanning parameters such as tube voltage / current, reducing radiation dose through multi-parameter synergy. Furthermore, the precisely defined eight categories of individual factors (especially the strong correlation (r=0.71) between thoracic longitudinal and transverse diameters and tube current) enable the model to establish accurate dose-body constitution relationships, ultimately reducing prediction errors. More importantly, the use of the PGH-1 phantom ensures seamless translation from experimental to clinical applications, improving the success rate of scanning parameter matching. These mutually reinforcing technical effects together constitute an optimized solution that possesses both clinical realism and technical versatility.

[0017] Furthermore, in step S10, the radiation dose is calculated using the SSDEWED method; during calculation, an elliptical region of interest is drawn in the middle layer of the original sequence, the region of interest encompassing the entire transverse section, and the average CT value of the region of interest is recorded. ROI and area A ROI Calculate WED and conversion factor f size·MED The specific calculation formulas are shown in formulas (1)-(5) below:

[0018]

[0019] f size·MED =a×e -b×WED (2),

[0020] SSD MED =f size·MED ×CTDI vol (3),

[0021] DLP SSDE·MED =f size·MED ×DLP (4),

[0022] ED SSDE·MED =k×DLPSSDE·MED (5), where the values ​​of a and b for adults are: a = 3.70469, b = 0.03671937; k = 0.014mSv·

[0023] mGy -1 ·cm -1 .

[0024] First, this solution involves drawing an elliptical Region of Interest (ROI) encompassing the entire cross-section at an intermediate level.

[0025] The WED calculation not only fully considers patient body type specificity (such as chest size and tissue density distribution), reducing dose calculation error compared to traditional methods, but also ensures consistency of dose calculation results across different institutions through standardized calculation formulas (1)-(5) and adult-specific coefficients (a = 3.70469, b = 0.03671937). Simultaneously, this scheme forms a closed-loop optimization with the ANN model, improving model iteration speed through real-time feedback of SSDE and effective dose (ED) data. More notably, the SSDEWED method has a built-in device difference correction function, reducing the standard deviation of dose prediction on five different CT devices.

[0026] Ultimately, the effective dose calculated by formula (5) not only achieved accurate carcinogenic risk assessment, but also guided the scanning protocol to reduce the dose to sensitive organs, thus constructing a complete radiation risk management system while ensuring image quality.

[0027] Furthermore, the scoring module in step S10 includes subjective evaluation and objective evaluation. The subjective evaluation uses a 5-point scoring standard to score the image quality.

[0028] In the objective evaluation, three planes were selected, including the sternal jugular notch, the aortic arch, and the tracheal bifurcation plane. Circular regions of interest were drawn in the lung parenchyma, erector spinae muscles, and 1 cm in front of the chest wall at each plane. The average CT value and standard deviation of the regions of interest were recorded, and the signal-to-noise ratio and contrast-to-noise ratio of the images were calculated. The calculation formulas are shown in formulas (6) and (7) below:

[0029] The signal-to-noise ratio of the image = mean CT value of lung parenchyma / CT value of air SD (6), the contrast-to-noise ratio = (mean CT value of lung parenchyma - mean CT value of erector spinae muscle) / SD of erector spinae muscle (7), and the scoring module calculates the scoring result of LDCT image by combining subjective evaluation results and objective evaluation results.

[0030] This approach first combines a 5-point subjective scoring system with quantitative SNR / CNR indicators to ensure compatibility with clinical diagnostic practices while reducing inter-assessor variability, thus providing a reliable data foundation for model training. Simultaneously, the rigorously standardized three-plane ROI measurement criteria not only keep measurement errors within a low range but also achieve a high postural bias compensation rate through bilateral measurement of the erector spinae muscles, while the dual ROI strategy for lung parenchyma improves alveolar visualization clarity. Furthermore, the design of characteristic planes such as the sternal jugular notch comprehensively covers key anatomical structures while reducing the CNR variation coefficient in each region. This anatomically representative assessment scheme directly improves the detection rate of sub-centimeter nodules and the visualization clarity of mediastinal lymph nodes. Particularly noteworthy is the establishment of a CNR-dose quantification model, which not only achieves high consistency in subjective scoring but also intelligently adjusts tube current. This dose-quality balance mechanism ultimately leads to a significant leap in clinical target achievement rates.

[0031] In addition to its core benefits, this solution offers multiple additional values: firstly, symmetry analysis based on bilateral ROIs can automatically detect and compensate for postural deviations; secondly, standardized noise feature analysis not only identifies typical artifact patterns but also guides iterative reconstruction algorithms to reduce noise in specific areas. More importantly, the automated quality control process integrated into PACS saves doctors time, while the constructed "visual-physical parameter" conversion model improves the diagnostic efficiency of 3-point images. Ultimately, these technological innovations collectively constitute an intelligent platform integrating parameter optimization, equipment monitoring, AI training, and clinical decision support. In practical applications, this platform shortens doctors' adjustment time, improves the success rate of novice operators, and fully demonstrates the comprehensive value of this solution in enhancing the efficiency of lung cancer screening.

[0032] Furthermore, in step S20, the vectorization processing of lung nodule parameters includes constructing the original parameter vector V from the lung nodule parameters without seed number modification, by taking the diameter of the lung nodule, its three-dimensional coordinates in the in vitro model, and its morphological features. init = [d,x,y,z,s], where d is the nodule diameter, x,y,z are the spatial coordinates, and s is the morphology coefficient; the seed number is generated by a random algorithm using a scaling factor σ and an offset Δ(x,y,z), which updates the original parameter vector to the initial test parameter vector V. adj = [d·σ,x+Δx,y+Δy,z+Δz,s]; the second test parameter vector V obtained by recognizing LDCT images. det Includes measured values ​​for the corresponding dimensions; during comparison, first compare V with σ. det Perform scaling calibration, then calculate the Euclidean distance D = ||V|| between the two vectors. adj -V det ||2 and cosine similarity The comparison results of D and C are used as the comparison results of the initial test parameter and the second test parameter.

[0033] This vectorization process constructs an original parameter vector encompassing multidimensional features including diameter d, spatial coordinates (x, y, z), and morphological coefficient s. It then uses a seed number to generate a scaling factor and offset modification parameter that are dynamically adjusted from 0.8 to 1.2. Finally, it achieves triple value through scaling calibration and dual comparison using Euclidean distance and cosine similarity. First, it precisely quantifies nodule changes, transforming the geometric and morphological features of lung nodules into vectors. This allows changes in nodule size and location to be calculated quantitatively rather than qualitatively. For example, in an in vitro model simulating a patient with pulmonary interstitial fibrosis, it can accurately capture subtle shifts in the three-dimensional coordinates of nodules caused by model deformation (e.g., a 0.3mm shift in the z-axis can still be identified), solving the problem of quantifying complex deformations that traditional methods struggle with. Second, it adapts to image scaling interference. The scaling calibration mechanism offsets feature misjudgments caused by parameter adjustments such as low-dose scanning (e.g., tube current 50mA). Even if the image shows a visual magnification of nodules due to noise, it can be reverse-calibrated to ensure the authenticity of parameter comparisons. Third, it improves model feedback accuracy. Dual-index comparison allows the ANN model to obtain accurate feedback. For example, when metal denture artifacts are mixed into the in vitro model, cosine similarity can distinguish the morphological differences between the artifact and the real nodule (artifact similarity <0.6, real nodule >0.8), helping the model to efficiently optimize and iterate. It connects the nodule feature quantification, image interference cancellation, and model feedback calibration links, adapting to complex scenarios and providing intelligent and precise support for low-dose CT parameter optimization.

[0034] Furthermore, in step S20, after obtaining the seed number, the LDCT image obtained after modifying the lung nodule parameters with the seed number is used as the verification image, and the LDCT image obtained without modifying the lung nodule parameters with the seed number is used as the initial image; during the correlation analysis, the initial image score Q is first calculated. init With verification image score Q cal Difference ΔQ=|Q cal -Q init | and construct an influence factor matrix M = [σ,Δx,Δy,Δz,ΔQ] by combining the scaling factor σ of the seed number and the offset Δ(x,y,z), where (x,y,z) are the spatial coordinate offset values; calculate the correlation between each element in the matrix and the scoring deviation using the Pearson coefficient, and select seed number parameters with a correlation coefficient |r|>0.6 as key influence factors; dynamically adjust the weight coefficients of the scoring model based on the key influence factors, using the following adjustment formula: Where w i Let r be the weight of the i-th factor. i Its correlation coefficient.

[0035] By constructing a correlation analysis mechanism between the initial image and the verification image for scoring differences, the scaling factor of the seed number, spatial offset, and scoring differences are first coupled into an influence factor matrix. In an in vitro model scenario simulating a lung nodule near the chest wall artifact area (the seed number is modified to shift the nodule to the vicinity of the chest wall), the strong correlation between the scoring deviation and the seed number parameter (correlation coefficient |r|>0.7) when the offset Δy>2mm can be accurately identified, thus solving the problem that traditional methods struggle to quantify the interference of complex factors on scoring. Furthermore, key factors are screened based on the Pearson coefficient, and their weights are dynamically adjusted. For example, when low-dose scanning (tube voltage 80kV) leads to increased image noise, if the scaling factor σ=1.1 causes "visual blurring of the nodule,"... Misjudgment (ΔQ = 1.2) can be mitigated by adjusting weights to enhance the proportion of noise-resistant features such as the stability of erector spinae CT values, thereby improving the adaptability of the scoring model to low-dose scenarios. Simultaneously, accurate factor identification can filter out invalid feedback for the ANN model. For example, when respiratory motion artifacts are mixed in (easily misjudged as nodule abnormalities), the correlation coefficient can distinguish the influence of artifacts (r < 0.5) from real nodules (r > 0.6), preventing the model from falling into an erroneous optimization loop, improving iterative convergence speed, and ultimately establishing a closed-loop optimization link between seed number, image scoring, and model feedback. This is particularly suitable for clinical scenarios involving thick chest wall fat and numerous artifacts in elderly patients, promoting the optimization of low-dose CT parameters towards precision and personalization.

[0036] Furthermore, in step S10, an adaptive enhancement module is introduced when establishing the ANN model. The adaptive enhancement module automatically selects Gaussian filtering or nonlocal mean denoising algorithm to preprocess the LDCT image based on the hardware characteristics of the scanning device model and the nodule distribution density of the in vitro model. The hardware characteristics of the device model include detector resolution and X-ray tube power. When the device resolution is less than or equal to a preset resolution threshold, a multi-scale feature fusion method is enabled to enhance the feature extraction of lung nodules. The output of the adaptive enhancement module is used as the input layer data of the ANN model.

[0037] By introducing an adaptive enhancement module when building the ANN model, Gaussian filtering or nonlocal mean denoising preprocessing is automatically selected based on the hardware characteristics of the scanning equipment (detector resolution, X-ray tube power) and the distribution density of nodules in the in vitro model. For example, in scenarios with older equipment (detector resolution 0.6mm, below the preset threshold) and dense distribution of nodules in the in vitro model (≥3 simulated nodules per square centimeter), multi-scale feature fusion can enhance the extraction of edge features of small nodules (3mm in diameter), solving the problem that traditional fixed preprocessing is difficult to adapt to the differences between different equipment and models; the denoising algorithm dynamically selects and adapts to diverse noise. For example, when the X-ray tube power is low (80kW) and the image quantum noise is high, nonlocal mean denoising can preserve the gray-scale difference between nodules and lung parenchyma, avoiding excessive blurring of boundaries by Gaussian filtering; multi-scale feature fusion provides richer and more accurate nodule features for the input layer of the ANN model, helping the model improve the accuracy of lung nodule identification and parameter association learning under complex conditions (such as irregular nodule distribution due to chest wall deformation in elderly patients), allowing low-dose CT scan parameter optimization to move from coarse adaptation to precise customization, especially in the scenario of old equipment in primary hospitals, breaking through hardware limitations to explore data value and expand the boundaries of technology application.

[0038] Furthermore, in step S30, dynamic weight coefficients α are introduced during the training of the ANN model to construct a composite objective function L. total =α·L dose +(1-α)·L score L dose For radiation dose deviation loss, L score For image scoring loss; when the subject's age exceeds a preset age threshold or body mass index exceeds a preset body mass index, α is increased to preferentially reduce the radiation dose.

[0039] By introducing a dynamic weighting coefficient α (e.g., dynamically adjusted from 0.3 to 0.7) to construct a composite objective function, in scenarios involving elderly subjects (e.g., 72 years old, exceeding the preset threshold of 65 years old) or obese patients (BMI 32, exceeding the preset threshold of 28), increasing α to 0.6-0.7 prioritizes reducing radiation dose. For example, in elderly patients with chronic obstructive pulmonary disease, this approach can ensure image score L... score Under the premise of a score of ≥3 (i.e., meeting the diagnostic criteria), the radiation dose is reduced to address the problem of neglecting individual differences in traditional fixed-weighting; at the same time, the radiation dose bias (L) is balanced. dose Image rating (L) doseFor example, in chest scans of children (8 years old, within the age limit but requiring dose control), an initial α=0.4 may lead to a high dose (e.g., a deviation of 15%). Adjusting α to 0.55 can strengthen dose constraints and also maintain image quality (score fluctuation <0.5 points) by optimizing scan parameters (pitch from 0.9 to 1.1). Furthermore, it allows the ANN model to adapt to diverse clinical needs, distinguishing between radical follow-up examinations for cancer patients (if high image quality is required, α=0.3 to strengthen the score) and health check-up screenings (if low dose is prioritized, α=0.7). In the health check-up population (low risk, requiring wide coverage), the radiation dose for most examinees can be controlled below 3mSv, and the rate of missed diagnoses of positive nodules is low. This promotes the advancement of low-dose CT from technically feasible to clinically friendly, and establishes a synergistic link between individual differences, dose constraints, and image quality. It not only protects the radiation safety of high-risk groups but also adapts to diverse scenarios, empowering precision medical imaging. Especially in scenarios with mixed examinations of multiple diseases and differentiated needs, it fully demonstrates the value of precise control.

[0040] Furthermore, in step S40, a multi-device compatibility adaptation mechanism is added before outputting personalized scanning parameters. By calling historical data from multiple reference devices that are associated with the core performance of the target device, a parameter compatibility evaluation system is constructed to quantitatively analyze the adaptability of scanning parameters across different devices. Based on the adaptability grading results, the parameter adjustment process is dynamically triggered. A correction coefficient model is established based on the differences in the core performance of the devices to optimize the cross-device adaptability of the scanning parameters. The differences in the core performance of the devices include hardware characteristics and algorithm types. The optimized parameters are then verified in a closed loop using preset diagnostic standards to ensure that the actual application of the parameters on the target device meets the dual requirements of image quality and radiation dose.

[0041] This multi-device compatibility adaptation mechanism breaks through the limitations of a single model adapting to a single device. By linking core performance characteristics, it covers hardware features such as the number of detector rows and X-ray tube power, as well as algorithm types such as the differences between statistical iterative and deep learning noise reduction algorithms. It calls upon historical data from multiple brands of reference equipment. In county-level medical consortium scenarios (such as new CT scanners used in central hospitals and older equipment used in township health centers), it achieves cross-device parameter adaptation by leveraging the commonalities of iterative algorithms, solving the pain point of difficulty in unifying standards due to the complexity of equipment at the grassroots level. When constructing the parameter compatibility evaluation system, the quantitative correction of differences breaks through the blindness of traditional experience-based adjustments. It not only links hardware power and detector differences, but also incorporates the impact of algorithm types (such as the equivalent conversion between filtered back projection and iterative algorithms). For low-dose dedicated CT scanners equipped with deep learning noise reduction, it can specifically reduce tube current while maintaining image quality, breaking away from the conventional approach of only focusing on hardware and ignoring algorithm collaboration. In the closed-loop diagnostic verification process, preset image quality standards (such as a 5-point scoring system) and radiation safety thresholds ensure the optimized parameters perform well on the target equipment. In multi-center lung cancer screening scenarios, this keeps image quality deviations and radiation dose fluctuations between different brands of equipment within a controllable range, ensuring consistency of diagnostic data across institutions and avoiding the risks of missed diagnoses or excessive radiation. In long-term application, parameter adjustment data feeds back into the system optimization, enabling the rapid generation of adaptation rules for new equipment (such as photon-counting detector CT), shortening the clinical deployment cycle of the technology.

[0042] This solution replaces model binding with core performance correlation and replaces experience-based debugging with quantitative correction. It not only solves the equipment adaptation problem in scenarios with uneven distribution of medical resources, but also provides parameter solutions for multi-center research and cross-institutional collaboration. Especially in primary healthcare scenarios, it allows low-dose CT to break through hardware limitations and release its value. The innovative logic of combining software and hardware and data iteration far exceeds the single approach of traditional equipment adaptation, and promotes the extension of technology towards a more universal and intelligent direction. Attached Figure Description

[0043] Figure 1 This is a flowchart of a method for optimizing chest low-dose CT scan parameters based on neural networks, as described in an embodiment of the present invention. Detailed Implementation

[0044] The following will describe the concept and technical effects of the present invention clearly and completely with reference to embodiments, so as to fully understand the purpose, features and effects of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the scope of protection of the present invention.

[0045] like Figure 1 As shown, the method for optimizing parameters of low-dose chest CT scans based on neural networks includes the following steps:

[0046] S10: Implant lung nodules of different sizes into an in vitro liver model, set different scanning parameters for different models of scanning equipment to perform LDCT scans on the in vitro model to obtain LDCT images, use a scoring module to score the LDCT images, select qualified images from the LDCT images based on the scoring results, obtain the corresponding scanning equipment model, scanning parameters, nodule characteristics and radiation dose of the qualified images, and establish an ANN model.

[0047] S20: Obtain the in vitro model corresponding to the qualified image as the initial in vitro model, generate the seed number through a random algorithm, and use the seed number to modify the lung nodule parameters in the initial in vitro model. Use the modified lung nodule parameters as the initial test parameters, use the ANN model to scan the initial in vitro model to obtain LDCT images, and identify the lung nodule size parameters as the second test parameters. Compare the initial test parameters and the second test parameters, and use the comparison results as feedback information to optimize the ANN model.

[0048] S30: The ANN model is trained and optimized by taking the individual factors of the examinee and the model of the scanning equipment as inputs, and the radiation dose and scanning parameters of the examinee as outputs; the normality test is performed on the input data of the ANN model, and the univariate independent variables are screened from the input data based on correlation analysis and one-way ANOVA, and then the key variables are selected from the univariate independent variables using random forest.

[0049] S40: Input the patient's individual factors and the scanning equipment model into the trained ANN model to obtain personalized scanning parameters and recommended radiation dose.

[0050] In step S10, when implanting lung nodules into the in vitro model, simulated lung nodules with diameters of 3mm and 5mm are implanted into the PGH-1 in vitro model in stages according to lung segments; the different models of scanning equipment include at least 5 types; the scanning parameters include tube voltage, tube current, pitch and gantry speed; the individual factors of the examinee include gender, age, height, weight, blood type, fat content within the scanning range, longitudinal and transverse diameters of the thoracic cavity at the largest / central level and the number of scanning layers.

[0051] Specifically, the in vitro model includes 18 lung segments.

[0052] In step S10, the radiation dose is calculated using the SSDEWED method. During the calculation, an elliptical region of interest is drawn in the middle layer of the original sequence, encompassing the entire transverse section, and the average CT value of the region of interest is recorded. ROI and area A ROI Calculate WED and conversion factor f size·MED The specific calculation formulas are shown in formulas (1)-(5) below:

[0053]

[0054] f size·MED =a×e -b×WED (2),

[0055] SSD MED =f size·MED ×CTDI vol (3),

[0056] DLP SSDE·MED =f size·MED ×DLP (4),

[0057] ED SSDE·MED =k×DLP SSDE·MED (5), where the values ​​of a and b for adults are: a = 3.70469, b = 0.03671937; k = 0.014mSv·

[0058] mGy -1 ·cm -1 .

[0059] The scoring module in step S10 includes subjective evaluation and objective evaluation. The subjective evaluation uses a 5-point scoring system to score the image quality. The specific scoring rules are shown in Table 1 below:

[0060]

[0061] Table 1

[0062] Table 1 shows the subjective image quality evaluation table. For objective evaluation, three planes were selected: the sternal notch, the aortic arch, and the tracheal bifurcation plane. Circular regions of interest (ROIs) were drawn 1 cm anterior to the lung parenchyma, erector spinae muscles, and chest wall at each plane. The ROI area was approximately 200 mm². 2 To avoid artifacts such as bone, calcification, and clothing, when delineating the erector spinae muscles, avoid intermuscular spaces and ray sclerosis artifacts. The lung parenchyma and erector spinae muscles should be taken bilaterally. Record the average CT value and standard deviation of the region of interest, and calculate the signal-to-noise ratio and contrast-to-noise ratio of the image. The calculation formulas are shown in the following formulas (6) and (7):

[0063] The signal-to-noise ratio of the image = mean CT value of lung parenchyma / CT value of air SD (6), the contrast-to-noise ratio = (mean CT value of lung parenchyma - mean CT value of erector spinae muscle) / SD of erector spinae muscle (7), and the scoring module calculates the scoring result of LDCT image by combining subjective evaluation results and objective evaluation results.

[0064] In step S20, the vectorization of lung nodule parameters includes converting the diameter of the lung nodule, its three-dimensional coordinates in the in vitro model, and its morphological features into an original parameter vector V, which is the lung nodule parameter without seed number modification. init = [d,x,y,z,s], where d is the nodule diameter, x,y,z are the spatial coordinates, and s is the morphological coefficient. The vectorization of lung nodule parameters includes the following process:

[0065] Scaling factors σ0∈[0.8,1.2] and three-dimensional offsets Δ0(x,y,z)∈[-2mm,2mm] are randomly generated in a uniform distribution and used as initial values ​​for the first iteration.

[0066] Low-dose scanning experiments were conducted on 120 PGH-1 in vitro models containing known nodules using five different CT scanners (GE Optima, GE VCT, United Imaging uCT860, GE Revolution, and Philips iCT). Eight typical combinations of scanning parameters (80kV / 50mA, 100kV / 100mA, and 120kV / 150mA, with 40 cases for each combination) were used. Using measured lung nodule parameters as the gold standard, the quantitative relationships between σ, Δ, and tube voltage (kV) and tube current (mA) were established.

[0067] σ=1.00+0.10·(100-kV) / 20+0.05·(100-mA) / 50

[0068] Δx,y,z=0.5·(100-kV) / 20+0.3·(100-mA) / 50

[0069] The above formula was validated by multiple linear regression (F = 38.7, P < 0.001, R < 0.001). 2 =0.62), which can explain 62% of the factor variation.

[0070] In subsequent iterations, σ and Δ are no longer randomly generated. Instead, the precise σ and Δ are calculated directly using the aforementioned quantization formula based on the real-time scanning parameters (kV, mA), and the initial test parameter vector is updated: V adj =[d·σ,x+Δx,y+Δy,z+Δz,s].

[0071] The second test parameter vector V obtained from recognizing LDCT images det First, scale and calibrate it using σ, then calculate the Euclidean distance D = ||V. adj -V det ||2 and cosine similarity (D,C) is used as feedback information for ANN model optimization.

[0072] Specifically, in constructing the original parameter vector V initAt the same time, in conjunction with the lung segmentation and nodule implantation rules of the in vitro model in step S10 (i.e., implanting nodules in segments and stages according to lung segmentation), an additional lung segment correction coefficient k is added for nodules of different lung segments. seg Based on retrospective data from 200 cases, a lung segment correction factor k was set. seg The details are shown in Table 2 below:

[0073]

[0074]

[0075] Table 2

[0076] Using data from 200 chest CT scans (112 males, 88 females), the following parameters were used to determine the composition of the chest CT scan: Fat% and chest diameter (T-scan). ratio (Anterior-posterior diameter / Left-right diameter) establish a multiple regression:

[0077] k seg =1.00+0.12×(Fat%-20%)-0.15×(T ratio -0.75)

[0078] When Fat% ≥ 25% or T ratio A value ≥0.85 triggers a 0.90 / 0.88 correction for the lower leaf segment;

[0079] When Fat% ≤ 15% and T ratio A correction of 1.10 / 1.08 is triggered in the upper leaf segment when the value is ≤0.70;

[0080] Regression model R 2 =0.64, P<0.001, validation set MAE=0.02.

[0081] Update vector V init =[d,x,y,z,s,k seg ], where k seg The (lung segment correction factor) is calculated in real time based on the individual factors of the examinee. A baseline value k is first assigned according to the lung segment where the nodule is located. seg 0: Upper lobe segment (apical segment, posterior segment, anterior segment) k seg 0 = 1.10; k of the middle segment (outer segment, inner segment) seg 0 = 1.05; lower lobe segment (dorsal segment, basal segment) k seg 0 = 0.90. Then, perform linear correction on the individual factors collected in step S30: k seg =k seg 0+0.12×(Fat%–20%)–0.15×(T ratio –0.75), Fat% is the fat content (%) within the scanned area, based on the DICOM automatic segmentation results; T ratioThis is the ratio of the maximum anteroposterior diameter to the lateral diameter of the thorax, automatically measured from the midline. This k... seg After incorporating vectors, the ANN model can distinguish the differences in nodule parameters between individuals with "upper lobe lung - tall and thin" and "lower lobe lung - obese" characteristics during the training phase, thereby improving the model's adaptability to individual clinical differences.

[0082] When generating the scaling factor σ and offset Δ(x,y,z) using the seed number, the historical constraints of the scanning parameters recorded in step S10 must be satisfied simultaneously. Based on retrospective low-dose scanning data from 180 cases of the same type of equipment (tube voltage 80–120kV, tube current 50–200mA), the quantum noise index QNI was found to have a quadratic relationship with the scanning parameters: QNI=0.2+0.015×(100-kV)+0.008×(100-mA); to avoid the scaling factor σ and offset Δ amplifying noise, a constraint function is defined: σ range =1±0.1·QNI,Δ range = 2mm·QNI; For example, when the tube voltage is 80kV and the tube current is 50mA, QNI = 0.2 + 0.015 × 20 + 0.008 × 50 = 0.9 → σ range = [0.91, 1.09] (normally [0.8, 1.2] for unconstrained values), Δ range = ±1.8mm (normal ±2mm). In practical applications, first read the current scanning protocol (kV, mA) from step S10, then calculate QNI in real time according to the above formula, and narrow the value range of σ and Δ. Finally, within the narrowed range, values ​​are still randomly selected with a uniform distribution to generate the seed number. The narrowed σ and Δ are used as new features input into the ANN model; after normality testing and random forest key variable screening, their contribution is verified (feature importance ≥ 0.05) to ensure that the seed number generation always matches the actual scanning noise level and reduces invalid iterations.

[0083] When calculating the Euclidean distance D and cosine similarity C, quantitative weight adjustments were made for the individual factors of the examinee in step S20 (including fat percentage and ABO genotype). Analysis of data from 300 adult CT studies revealed the following: the average BMI for type O was 25.5 ± 10.2, and for type AB it was 19.53 ± 2.35; the regression coefficient for type O BMI was β = +2.53 (P = 0.025). β was linearly mapped to [0.90, 1.10], with a weight of 1.00 + 0.04 × β. Fat interference was calculated as z-axis coordinate error of 0.5 mm × (Fat% - 20%) and morphological coefficient error of 0.05 × (Fat% - 20%). Dynamic weight adjustments were made: z-axis was [1 + 0.02 × (Fat% - 20%)] × weight, with an upper limit of 1.4; morphological coefficient was [0.8 - 0.01 × (Fat% - 20%)] / weight, with a lower limit of 0.6. The dynamic weights for simulating obese subjects with a fat percentage of 35% and type A are: z-axis weight [1 + 0.02 × 15] × 0.97 = 1.30 × 0.97 ≈ 1.26, and morphological coefficient weight [0.8 - 0.01 × 15] / 0.97 = 0.65 / 0.97 ≈ 0.67. These adjusted dynamic weights of 1.26 and 0.67 are input as additional features into the ANN, enabling personalized scanning parameters to automatically increase tube current compensation and z-axis reconstruction slice thickness weight during low-dose obese scanning, effectively reducing positioning errors caused by fat artifacts.

[0084] This vectorization process, through adjustments using lung segment correction coefficients and body fat weights, adapts to differences in lung segment anatomy and individual patient constitutions, addressing the problem of traditional methods neglecting individual clinical characteristics. In multi-center clinical validation, it improves the accuracy of nodule parameter identification and reduces model misjudgments due to individual differences for different subjects, such as obese and tall-slender individuals. Based on the LDCT historical logs in this protocol, a model (Ri) of QNI = 0.15 + 0.012·(100-kV) + 0.006·(mA-100) / 50 is established. 2 =0.72). Seed number σ, Δ according to σ range =1±0.10·QNI、Δ range=2mm(1-QNI) real-time narrowing; at 80kV / 50mA, σ∈[0.961,1.039], Δ∈±1.22mm. The narrowed σ and Δ are used as physical constraint features input to the ANN, which works in conjunction with the normality test and random forest screening in step S30. The number of iterations is reduced from 180 to 120, and the validation set MAE is reduced from 1.20 to 0.93mm, ensuring that the training data fits the real scanning noise level, improving the model iteration efficiency, and reducing invalid parameter feedback. For common clinical interferences such as fat artifacts, the above-mentioned dynamic adjustment of vector comparison weights reduces the nodule misdiagnosis rate in low-dose CT applications for obese patients, ensures radiation dose safety, and achieves a dual improvement in accurate identification and low-dose safety. It breaks through the limitations of the traditional vector comparison method in terms of scenario uniformity, promotes the optimization of low-dose CT parameters towards a more intelligent and accurate direction, and the multi-stage collaboration makes the technology adaptable to diverse clinical scenarios, laying a solid foundation for model training and practical application.

[0085] In step S20, after obtaining the seed number, the LDCT image obtained after modifying the lung nodule parameters with the seed number is used as the verification image, and the LDCT image obtained without modifying the lung nodule parameters with the seed number is used as the initial image; during the correlation analysis, the initial image score Q is first calculated. init With verification image score Q cal Difference ΔQ=|Q cal -Q init | and construct an influence factor matrix M = [σ,Δx,Δy,Δz,ΔQ] by combining the seed number scaling factor σ and the offset Δ(x,y,z). Calculate the correlation between each element in the matrix and the scoring deviation using the Pearson coefficient, and select seed number parameters with a correlation coefficient |r|>0.6 as key influence factors. Dynamically adjust the weight coefficients of the scoring model based on these key influence factors, using the following adjustment formula: Where w i Let r be the weight of the i-th factor. i Its correlation coefficient.

[0086] Specifically, when acquiring initial and verification images, the rules for implanting nodules in segments according to lung size are combined with the correlation analysis of images corresponding to nodules in different lung segments (such as the upper lobe, middle lobe, and lower lobe). Because different lung segments differ in anatomical structure and physiological movement—for example, the lower lobe is more affected by heart and diaphragm movement—the calculation of the difference in scores between the initial and verification images and the construction of the influence factor matrix need to incorporate a lung segment correction coefficient. This correction coefficient can be coordinated with the features of different layers (sternal notch, aortic arch, etc.) when calculating radiation dose using the SSDEWED method. Specifically, the sternal notch layer is used for tall, thin individuals in the upper lobe, and the diaphragmatic layer is used for obese individuals in the lower lobe, synchronizing the QNI-dose mapping with the lung segment movement amplitude; the ANN simultaneously receives k... segThe system incorporates WED (Weight, Surface, and Edge) labels, and gradient-directed segment-specific weights to achieve three-dimensional collaborative learning of "dose-segment-lung segment." This enables the ANN model to more accurately learn the relationship between nodule parameters and image scores across different lung segments and scanning slices.

[0087] The scaling factor σ and offset Δ(x,y,z) for the seed number are generated by referencing historical data from different scanning equipment models and different scanning parameters (such as tube voltage and tube current). When scanning with a low tube voltage (e.g., 80kV), the fluctuation range of σ is appropriately reduced to avoid excessive interference from noise in the nodule parameter changes after seed number modification due to high image noise at low doses. This makes the generated calibration image more consistent with the actual low-dose scanning scenario, providing a more reliable data foundation for subsequent correlation analysis. This also aligns with the goal of focusing on radiation dose calculation and pursuing high-quality low-dose imaging. Appropriate reduction refers to calculating the quantum noise level under the current scanning conditions based on the noise-parameter model (QNI = 0.15 + 0.012·(100-kV) + 0.006·(mA-100) / 50) obtained from historical data regression, and narrowing the fluctuation range of the seed number scaling factor σ from the usual ±20% to ±(10%·QNI). When QNI≈0.39 at 80kV / 50mA, the σ range is reduced to ±3.9%. Similarly, the upper limit of the offset Δ is reduced from ±2mm to ±1.22mm. This prevents excessive perturbation while retaining necessary variation, hence the term "appropriate".

[0088] When using the Pearson coefficient to screen key influencing factors, five seed parameters—σ, Δx, Δy, Δz, and Δs (morphological shift)—are extracted for each score level, combining the subjective and objective evaluation results of the scoring module. The Pearson coefficient r between these parameters and the scoring deviation is then calculated. For images corresponding to different score levels in the subjective evaluation, the seed numbers σ, Δx, Δy, Δz, and Δs are extracted, and their r values ​​with the scoring deviation are calculated. The correlation analysis between these seed parameters and the scoring deviation is treated differently. For images with a subjective evaluation of 5 points (i.e., excellent image quality), the Pearson coefficient r between σ, Δx, Δy, Δz, and Δs and the scoring deviation is calculated, retaining only parameters with r ≥ 0.35 and P < 0.05. For images with a score of 3 points (i.e., image quality barely meets diagnostic requirements), if r ≤ -0.30 for σ, Δz, or Δs and P < 0.05, the corresponding parameter is marked as a defect factor.

[0089] Using five parameters σ, Δx, Δy, Δz, and Δs as initial independent variables and the scoring deviation as the dependent variable, a stepwise regression method was employed (entering the standard P-value). enter =0.05, removing standard P removeVariable selection is performed using β = 0.10. For example: σ first enters the model (β = -0.42, P = 0.002); Δz enters the model (β = -0.38, P = 0.007); Δs enters the model (β = -0.33, P = 0.012); at this time, the P values ​​of Δx and Δy are 0.18 and 0.23 respectively, both > 0.10, and are therefore removed. The final model contains three terms: σ, Δz, and Δs. After adjustment, R... 2 _adj=0.68, F=38.7, P<0.001, jointly explained 68% of the rating bias, and was therefore identified as a key factor.

[0090] When dynamically adjusting the weight coefficients of the scoring model, the results of vectorization processing and vector comparison of lung nodule parameters are considered. If the vector comparison reveals that the nodule morphology coefficient s changes specifically due to the modification of the seed number (e.g., fluctuations in the s value cause the nodule shape to change from a near-circular shape to an irregular shape in the image) and is strongly correlated with the scoring deviation, the weight of the morphology coefficient s in the scoring model is adjusted simultaneously when adjusting the weight of the corresponding seed number parameter. This achieves the linkage optimization between the scoring model and the nodule parameter vector analysis, allowing the scoring model to more comprehensively and accurately assess image quality. It also enables different stages (nodule parameter modification, image scoring, and model optimization) in the entire low-dose CT scan parameter optimization process to coordinate and promote each other.

[0091] The selected key influencing factors and adjusted scoring model weights are used as feedback information in the ANN model training and optimization process. These key influencing factors serve as important features in the ANN model's input data, participating in normality tests, correlation analyses, and key variable selection processes. This allows the ANN model to focus more on variables that play a crucial role in image scoring and nodule parameter optimization, improving training efficiency and effectiveness. Simultaneously, the adjusted scoring model weights guide the ANN model during training to more rationally balance radiation dose bias loss and image scoring loss (e.g., constructing a composite objective function). This allows the personalized scanning parameters output by the ANN model to better meet the image quality requirements for nodule diagnosis while pursuing low radiation dose, driving the ANN model towards greater intelligence and greater alignment with clinical practice.

[0092] By integrating multiple components such as lung segmentation, radiation dose calculation, ANN model, scoring module, and nodule parameter vectorization, this approach breaks down the isolation between steps, constructing a complete collaborative chain from nodule parameter modification, image acquisition and scoring, to key factor selection, model weight adjustment, and then feedback to ANN model training. This makes the entire method for optimizing low-dose chest CT scan parameters an organic whole, with each step supporting and optimizing the others, enhancing the systematicity and scientific rigor of the approach and solving the problems of loose connections and poor coordination between steps in traditional optimization processes.

[0093] In this specific implementation, on the one hand, more precise screening of key influencing factors and adjustment of scoring model weights can effectively improve the accuracy and comprehensiveness of image scoring, allowing the scoring model to more sensitively capture the impact of seed number modification and nodule parameter changes on image quality, thereby providing more reliable image quality assessment results for subsequent scanning parameter optimization. On the other hand, synergy with ANN model training enables the ANN model to learn more critical and valuable features and patterns, improving the accuracy of the model's output of personalized scanning parameters, achieving a dual improvement in image quality optimization and model optimization. This helps to obtain higher-quality CT images that better meet clinical diagnostic needs under low-dose conditions, while making the optimized scanning parameters more closely aligned with the individual patient and the actual situation of the scanning equipment. Specifically, adjusting the scoring model weights refers to adjusting the correlation coefficient r of the above three factors. i Substitute into the dynamic weight formula Real-time updates of subjective and objective scoring weights make the scoring model more sensitive to nodule edge blurring, z-axis artifacts, or shape distortion, thereby outputting more accurate image quality scores and providing reliable labels for subsequent ANN parameter optimization.

[0094] Meanwhile, this solution generates verification images by combining historical data from different lung segments and scanning parameters, then filters key factors based on subjective and objective evaluation results of the associated images, and finally adjusts the model weights through vector analysis of nodule parameters (this process involves "creating verification images from real historical data of lung segment-equipment-dose, then using subjective and objective scoring to filter out the few factors that truly affect quality, and finally incorporating these factors into the vector weights for real-time adjustment, so that the model can accurately locate nodules 'at a glance' on whom, in which segment, and at what dose."). The entire implementation details fully consider the diversity (including different lung segments, different equipment, and different doses) and complexity (including image noise and nodule morphology changes) of actual clinical scanning, enabling the optimized method to better adapt to actual clinical application scenarios. Under the premise of ensuring low radiation dose, it improves the ability to display and diagnose nodules in different patients and under different scanning conditions, providing stronger technical support for accurate clinical diagnosis. It breaks through the limitations of traditional optimization methods in clinical application and demonstrates stronger clinical practical value.

[0095] In step S10, an adaptive enhancement module is introduced when establishing the ANN model. The adaptive enhancement module automatically selects Gaussian filtering or nonlocal mean denoising algorithm to preprocess the LDCT image based on the hardware characteristics of the scanning device model and the nodule distribution density of the in vitro model. The hardware characteristics of the device model include detector resolution and X-ray tube power. When the device resolution is less than or equal to a preset resolution threshold, a multi-scale feature fusion method is enabled to enhance the feature extraction of lung nodules. The output of the adaptive enhancement module is used as the input layer data of the ANN model.

[0096] Specifically, the adaptive enhancement module determines the filtering algorithm in one go based on the three-dimensional decision table of "detector resolution - X-ray tube power - lung lobe", which includes the following steps:

[0097] A1: First, classify according to detector resolution:

[0098] A1-a: When the detector resolution is ≤0.5mm (high resolution), proceed to step A2;

[0099] A1-b: When the detector resolution is >0.5mm (low resolution), proceed directly to step A3.

[0100] In the case of A2 high resolution, a second judgment is made using the X-ray tube power. If:

[0101] A2-a: If the X-ray tube power is ≥100kW, then nonlocal mean denoising should be selected (e.g., h=0.8·σ). noise );

[0102] A2-b: If the X-ray tube power is between 80–100kW, then nonlocal mean denoising should be selected (e.g., h = 0.8·σ). noise );

[0103] A2-c: If the X-ray tube power is <80kW, then select a Gaussian filter (e.g., σ=1.2pixel).

[0104] A3: In low-resolution cases, regardless of the power of the X-ray tube, a Gaussian filter (e.g., σ = 1.2 pixels) should be selected.

[0105] A4: Fine-tune according to the lung lobe location; for the upper lobe (high z-coordinate, small nodules and low density), if Gaussian filtering has been selected, add a multi-scale feature fusion step to enhance the 3mm nodule edge; if non-local mean has been selected, keep the original algorithm unchanged. For the lower lobe (more overlapping nodules), maintain the algorithm selected in steps A1–3, and do not add any further processing.

[0106] When the device resolution is less than or equal to a preset threshold, the multi-scale feature fusion module is activated, incorporating objective evaluation metrics (including signal-to-noise ratio and contrast-to-noise ratio) from the scoring module. The noise ratio at different scales is calculated within the lung parenchyma region of interest (i.e., avoiding artifacts). If low-dose scanning (tube current ≤ 50mA) results in high image noise, the weights of the multi-scale fusion are automatically adjusted (e.g., increasing the proportion of high-frequency detail subbands to 40%). When the device resolution is ≤ 0.5mm, the signal-to-noise ratio (SNR) at four scales (s = 1, 2, 3, 4) is calculated within the lung parenchyma ROI. s _Contrast Noise Ratio (CNR) s The weight is calculated as exp(-|SNR) s -SNR refThe weights (| / τ) are adjusted in real time, where τ is a scale parameter that adjusts the sensitivity of the weights to differences in signal-to-noise ratio. The four scales need to be divided according to frequency: s=1 (finest scale, 25% weight allocation) or s=2 (second-finest scale, 15% weight allocation) for high-frequency sub-bands, s=3 (medium scale, 20% weight allocation) for mid-frequency sub-bands, and s=4 (coarsest scale, 40% weight allocation) for low-frequency sub-bands. The original weight base is 25% (equivalent allocation).

[0107] If the tube current ≤50mA results in high noise, the weight is increased from 25% to 40% to improve the gradient at the 3mm nodule edge. The fused features are output through a device compatibility adaptation layer (unifying the 0.6mm equivalent resolution), ensuring that the same algorithm can run stably on older GE, United Imaging, and Philips devices. This achieves low-dose, high-resolution imaging without the need for high-end hardware, compensating for the loss of nodule details in low-dose CT scans. Simultaneously, the fusion results are linked with the device compatibility adaptation mechanism. For older devices in primary hospitals (i.e., devices with low resolution but requiring common parameters), multi-scale feature transformation allows the same enhancement algorithm to adapt to the hardware differences of different brands of equipment, overcoming the limitation of new algorithms relying on high-end hardware.

[0108] The output data of the adaptive enhancement module (i.e., preprocessed LDCT images) is hierarchically labeled with individual subject factors (including age and body mass index) before being input into the ANN model. For elderly subjects (with chest wall fat thickening features added to the simulated in vitro model), the feature extraction of the erector spinae muscles and the region of interest in front of the chest wall is enhanced and associated with individual subject factors (including fat content and blood type), enabling the ANN model to learn the association rules of "fat layer interference, nodule features, and scanning parameter optimization". For models with a high body mass index (≥28), multi-scale fusion is used to highlight hilar nodules (i.e., spatial coordinates x, y close to the mediastinum), compensating for misjudgments of nodule displacement caused by chest wall deformation in obese subjects, making the ANN model training data closer to the real clinical scenario and improving the personalization level of parameter optimization.

[0109] This specific implementation breaks through the limitations of single preprocessing adapted to the entire lung lobe. By combining lung lobe segmentation, equipment characteristics, and nodule distribution, the feature extraction of small nodules and overlapping nodules is more in line with clinical anatomical patterns, solving the pain points of missed diagnosis of the upper lobe and misdiagnosis of the lower lobe in low-dose CT.

[0110] Multi-scale fusion, combined with noise ratio and individual factors, not only compensates for image defects in low-dose scanning but also breaks through equipment hardware barriers, enabling even outdated equipment at the grassroots level to run advanced enhancement algorithms, thus promoting the downward spread of technology.

[0111] Preprocessed data with stratified labeling of clinical factors allows the ANN model to learn more granular individual differences, image features, and parameter optimization rules. This upgrades personalized scan parameter recommendations from device adaptation to precise individual adaptation, providing higher-quality input data for image scoring models. Ultimately, this achieves end-to-end collaboration between hardware, algorithms, models, and clinical scenarios. This closed-loop synergy far surpasses the traditional design approach where preprocessing is independent of model training.

[0112] In step S30, dynamic weight coefficients α are introduced during the training of the ANN model to construct a composite objective function L. total =α·L dose +(1-α)·L score L dose For radiation dose deviation loss, L score For image scoring loss; when the subject's age exceeds a preset age threshold or body mass index exceeds a preset body mass index, α is increased to preferentially reduce the radiation dose.

[0113] When setting the dynamic weighting coefficient α, BMI is first calculated based on height and weight. Then, a baseline value is calculated using the formula 0.5 + 0.1 * [(Age - 65) / 5 + (BMI - 28) / 5 + (Fat% - 20%) / 10], combined with the fat percentage (Fat%) and age within the scanning range. This baseline value is then finely adjusted based on real-time subjective scores. If the real-time subjective score is ≥4 points, α = baseline value - 0.2 (not less than 0.3); if the real-time subjective score is ≤3 points, α remains unchanged. This ensures a dual constraint of "safety first, quality second." Here, Age is the subject's age (years), and BMI is the body mass index (kg / m²). 2 Fat% represents the fat content (%) within the scanned area.

[0114] When setting the dynamic weight coefficient α, a grading rule is constructed by associating individual factors of the examinee (including height, weight, and fat content within the scanning range). For elderly examinees, the lung tissue elasticity reduction feature is added to the simulated in vitro model. If their fat content exceeds 30% (i.e., a simulated fat layer is added around the nodules in the in vitro model), α is increased. At the same time, the weight of the loss of the nodule morphology coefficient s (which is easily deformed due to fat compression) is increased by 20% in the vectorization of lung nodule parameters. This ensures that the model prioritizes the accurate identification of nodule morphological features when reducing radiation dose. For examinees with a normal body mass index but whose age exceeds the threshold, the subjective evaluation of the scoring module is combined. If the subjective score is ≥4 points (i.e., excellent image quality), the weight of α on radiation dose is appropriately reduced to retain the scanning parameter features of high-quality images and avoid image quality fluctuations caused by excessive dose control. This achieves dynamic weight adjustment under the dual constraints of individual physical condition and image quality.

[0115] The in vitro model was stratified according to Fat%; when Fat% ≥ 30%, a 2mm fat layer was filled around the nodules to simulate decreased lung tissue elasticity. The weight of the morphological coefficient s was increased by 20% so that the ANN could preferentially maintain the accuracy of s even when the dose was reduced. For example, the subject was 75 years old with a BMI of 22 kg / m². 2 Fat% 25%, real-time subjective score = 5 points, base value 0.5 + 0.1·(10 / 5) = 0.7 → after adjustment α = 0.5, the corresponding tube current is only reduced by 8%, DLP remains at 0.9mSv, avoiding excessive dose reduction.

[0116] Constructing a composite objective function L total At that time, the radiation dose calculated using the SSDEWED method was incorporated as L. dose Real-time feedback. When the equipment model is a low-dose dedicated CT (such as a photon counting detector), it automatically identifies its hardware characteristics (including low tube power and high detector efficiency) and adjusts L... dose When adjusting the threshold calculation, first read the device model and protocol tag in the DICOM header file and match it with the built-in hardware database: if it is identified as a "photon counting detector" → mark gain. flag =1 (low-dose dedicated hardware such as photon counting detectors), other traditional detectors gain flag = 0 (traditional detector (no additional gain)). gain flag This is a Boolean (0 / 1) flag used to quickly identify whether the current device has "low-dose gain" capability during software runtime. According to the measured calibration table L... dose_threshold = 0.8mSv × (1 + 0.2·gain) flag This means that the threshold for photon counting devices is relaxed to 0.96 mSv, while the threshold for traditional devices remains at 0.8 mSv.

[0117] By correlating the radiation dose deviation loss with the low-dose gain coefficient of the equipment hardware (e.g., allowing a 10% increase in dose deviation when the gain coefficient is ≥1.2), and compensating for dose misjudgments caused by algorithm differences in new equipment, the quantum efficiency of the photon counting detector is improved by approximately 1.2 times, and the tube current can be reduced by 15–20% at the same noise level. Therefore, the low-dose gain coefficient G is defined as 1.2, and the dose deviation loss is corrected to L' dose =|SSDE pred -SSDE target | / (SSDE target •G), when G=1.2, a 20% increase in dose bias is allowed (the denominator becomes larger), preventing excessive penalty due to hardware advantages and ensuring that the model prioritizes image quality over over-compressed dose. Meanwhile, L scoreThe access device compatibility adaptation mechanism adjusts the image scoring loss for different brand devices (such as brand A and brand B) through an algorithm equivalent transformation coefficient (such as a coefficient of 0.8 when the iterative algorithm is converted to the filtering back projection algorithm). This allows the same objective function to adapt to the scoring standard differences of multiple brand devices, breaking through the limitation of composite functions relying on a single device.

[0118] The model iteration triggered by dynamic weight adjustment is linked to the comparison results of lung nodule parameter vectors. If the vector comparison finds that the correlation between the nodule diameter d (due to the modification of the seed number) and the scoring deviation exceeds 0.7 (i.e., strong correlation), after updating α, the objective evaluation threshold of the scoring model is simultaneously corrected (e.g., the signal-to-noise ratio pass threshold is adjusted from 20 to 20·(1+0.05·Δd), where Δd is the change in nodule diameter), so that the scoring model and ANN model training form a closed loop of parameter adjustment, scoring correction and model iteration. For pediatric subjects (i.e., the simulated in vitro model with the nodule size reduced to 2mm), the radiation dose loss weight is reduced by dynamic weighting, and the image scoring loss weight is strengthened to compensate for the defect that small nodules are easily missed in low-dose scans of children, making the ANN model training more in line with the clinical needs of special populations (such as children and the elderly).

[0119] In step S40, a multi-device compatibility adaptation mechanism is added before outputting personalized scanning parameters. This involves calling historical data from multiple reference devices with core performance correlations to the target device to construct a parameter compatibility evaluation system. This system quantifies and analyzes the adaptability of scanning parameters across different devices. The adaptability is determined by calling the scanning results of at least three historical devices with the same core parameters (detector row count, X-ray tube power, reconstruction algorithm) as the target device, calculating the Mahalanobis distance of their "image quality-dose" scatter plots, and normalizing the distance to a 0–1 score, i.e., the adaptability score. A score <0.7 triggers parameter correction, while a score ≥0.7 allows direct pass-through. Based on the adaptability grading results, a parameter adjustment process is dynamically triggered. A correction coefficient model is established based on the differences in core device performance to optimize the cross-device adaptability of the scanning parameters. These core performance differences include hardware characteristics and algorithm types. The optimized parameters are then validated using preset diagnostic standards to ensure that the actual application of the parameters on the target device meets both image quality and radiation dose requirements.

[0120] When constructing a parameter compatibility assessment system using historical data from reference devices, data is extracted according to lung segmentation and nodule implantation rules, categorized by upper and lower lobes. For upper lobe nodules, spatial coordinates and morphological coefficients of lung nodule parameters are associated to construct a dedicated subsystem. When the target device is a low-tube voltage device, reference device data with the same low-tube voltage and complete lung segment data is prioritized to fit the actual scanning scenario and address the issue of ignoring differences in lung anatomical segmentation.

[0121] When optimizing parameters based on device performance differences, the results of the correlation scoring module are considered. If the target device image has a low subjective score but a decent objective score, the parameters related to the morphology coefficient are adjusted specifically by comparing the nodule parameter vectors to compensate for the impact of device differences. When acquiring the impact of device differences, a triplet of "subjective score - objective index - morphology vector error" is saved for each device, and the difference is quantified using Mahalanobis distance; a difference <0.7 is considered significant. All cases of significant differences undergo a closed-loop iteration of corrected coefficient model → ANN training → re-validation → re-correction, automatically generating a "device-parameter-effect" mapping table, forming a reusable multi-device adaptation rule library. Simultaneously, the adapted parameters are input into the ANN model for training, forming a closed loop of parameter adjustment, model optimization, and re-adaptation, providing data for the ANN model to learn multi-device adaptation rules.

[0122] When optimizing parameters during closed-loop validation, individual subject factors are considered, and validation is performed separately for in vitro models such as those for the elderly and obese individuals. The elderly model is used to validate the visualization of small nodules in the upper lobe of the lungs under low-dose administration, while the obese model is used to validate the interference of fat artifacts on nodule visualization. The validation results are fed back to the device compatibility mechanism, enriching the diagnostic standard library and improving the reliability of the fit.

[0123] In one specific embodiment of this scheme, 400 subjects with diagnosed pulmonary nodules ≤5mm in diameter recruited by the hospital underwent low-dose scanning experiments. The model was iterated and optimized using multi-dimensional individual data from the subjects to obtain a personalized adult chest LDCT scan model. This personalized adult chest LDCT scan model was then used in a double-blind controlled experiment with 200 recruited physical examination subjects (conventional chest dose vs. personalized low-dose model dose) to verify the stability and effectiveness of the personalized low-dose model. The data from this experiment was then used to further iterate and optimize the model, resulting in a stable and personalized LDCT scan model.

[0124] The hospital incorporates the examinee's gender, age, height, weight, blood type, fat content within the scan range, longitudinal and transverse diameters of the thoracic cavity at the largest / central level, number of scan layers, and TCM constitution classification into a multi-dimensional category of individual factors.

[0125] The hospital used five different brands / models of CT scanners (GE Optima, GE VCT, United Imaging uCT860, GE Revolution, and Philips iCT), covering mainstream clinical models to ensure cross-device compatibility of the model. The scanning parameters for acquiring qualified images (low-dose images) using the five scanning devices are shown in Table 3 below. Table 3 shows the scanning parameters for the low-dose experiment of the in vitro model:

[0126]

[0127]

[0128] Table 3

[0129] The control group used standard dose scanning parameters for each device (e.g., tube voltage 120kV, tube current 200-300mA, pitch 1.0), while the experimental group used optimized personalized low-dose parameters (output via an ANN model, tube voltage 80-100kV, intelligent tube current adjustment to 50-150mA, pitch 0.6-1.375). Both groups used the same iterative algorithm (e.g., SAFIRE, MBIR). The scanning range was from the lung apex to the costophrenic angle, with a slice thickness of 5mm, a reconstruction slice interval of 5mm, and a gantry rotation speed of 0.5-0.8s to ensure standardized scanning protocols.

[0130] All subjects underwent dual-phase scanning (plain scan + contrast-enhanced scan). Two senior radiologists (with over 10 years of experience) evaluated the LDCT images using a double-blind method. The scoring module combined a subjective 5-point scale (Table 2) with objective indicators (signal-to-noise ratio, contrast-to-noise ratio), and the final mean score was used as the image quality score (Q). final Simultaneously record radiation dose parameters: volumetric CT dose index, dose-length product (DLP), and transmit via SSDE. WED The effective dose (ED) is calculated using formulas (1)-(5) to ensure the accuracy of dose assessment.

[0131] Based on the original eight categories of individual factors (gender, age, height, weight, blood type, fat content within the scanning range, maximum thoracic diameter at the largest level, and number of scanning layers), nine new TCM constitution types (balanced constitution, qi deficiency constitution, yang deficiency constitution, yin deficiency constitution, phlegm-dampness constitution, damp-heat constitution, blood stasis constitution, qi stagnation constitution, and special constitution) were added as auxiliary variables. Constitutional characteristics were quantified through questionnaires combined with tongue and pulse data. One-Hot coding was incorporated into an ANN model to explore the impact of constitution differences on radiation sensitivity. Respiratory gating signals were introduced as dynamic parameters. Respiratory frequency and amplitude data were collected using a chest-strap respiratory sensor, adding a time dimension feature to the model to optimize the impact of respiratory motion artifacts on the reconstructed images.

[0132] In the retrospective data of this study (400 adult chest CT data), the core body shape indicators and dose-related parameters of 9 TCM constitutions were measured, including the average chest wall fat thickness, anteroposterior diameter of the chest, and dose-body shape regression coefficient (all statistically verified, P<0.05): Phlegm-dampness constitution: average chest wall fat thickness 28mm, dose-body shape regression coefficient +0.28 (P<0.01) → an additional 20% radiation dose is required for the same image quality; Yang deficiency constitution: average chest wall fat thickness 16mm, dose-body shape regression coefficient -0.15 (P=0.03) → a 10% reduction in radiation dose is tolerable; Balanced constitution: average chest wall fat thickness 20mm, dose-body shape regression coefficient approximately 0 → maintain the clinical reference dose (no adjustment required); the measured parameters of the remaining 6 constitutions (Qi deficiency, Yin deficiency, damp-heat, blood stasis, Qi stagnation, and special constitution) were recorded according to the same logic to ensure that the correlation between constitution and dose is quantifiable.

[0133] To transform physical differences into parameters identifiable by the model, the dose-body size regression coefficients were first linearly mapped to the clinical safety range of [0.85, 1.20]. Then, the radiation sensitivity coefficient was calculated using the formula: w rad = 1 + 0.35 × regression coefficient, calculate the final value (0.35 is the clinical calibration factor to ensure w rad (Adapting to the dose control requirements of low-dose CT), ultimately yielding w for 9 body types. rad (e.g., phlegm-dampness constitution 1.20, yang deficiency constitution 0.85, balanced constitution 1.00). In ANN model training, the TCM constitution classification is transformed into a 9-dimensional vector using One-Hot encoding (each dimension corresponds to one constitution type, "1" indicates belonging to that constitution, and "0" indicates not belonging to it), and is then compared with w. rad These features, together serving as input features for dimensions 9–17 of the model, synergize with the original eight categories of individual factors (gender, age, etc.). These features directly participate in the calculation of the composite objective function: L total =α·L dose ·W score +(1-α)·L score L total α is the composite objective function value (the smaller the value, the better the balance between dose and image quality); α is the dynamic weighting coefficient (ranging from 0.2 to 0.7, calculated in real-time based on the subject's age and BMI); L dose The radiation dose deviation loss is calculated using the formula |SSDE. pred -SSDE target | / (SSDE target (SSDE pred To predict body type-specific doses, SSDE target (Target dose); L scoreImage rating loss is the squared error between the predicted image quality score on a 5-point scale and the artificial gold standard score.

[0134] Using the above methods, the model can automatically adjust parameters according to differences in constitution: phlegm-dampness constitution due to w rad =1.20 (high sensitivity), which will trigger alpha upregulation to preferentially reduce radiation dose; Yang deficiency constitution due to w rad =0.85 (low sensitivity) will trigger α downregulation to focus on ensuring image quality, ultimately achieving individualized dose allocation based on individual differences in physical condition.

[0135] 9 types of body types w rad The numerical values ​​and their basis are shown in Table 4 below:

[0136]

[0137]

[0138] Table 4

[0139] For the training data, a GAN-based bidirectional feature alignment model was used to perform multi-scale feature alignment between the original LDCT images and the reconstructed images. The specific process is as follows: The image was decomposed into four scales (j=1-4) using wavelet transform, where j=1-2 is the high-frequency sub-band (focusing on nodule edges), and j=3-4 is the low-frequency sub-band (focusing on lung structure contours). Hessian matrix analysis was used to analyze nodule features at each scale. Cosine similarity and Wasserstein distance were used to evaluate the matching degree between the reconstructed features and reference features (a predefined nodule feature library from an in vitro model), dynamically adjusting the weights of the scoring module to enhance the feature extraction capability for small nodules (e.g., less than 3mm). A contrastive learning mechanism was introduced, performing proportional scaling (σ=0.5, 1.5) and rotation transformations (θ=0°, 90°, 180°, 270°) on the input images to generate multi-view training samples. This forced the model to learn the scale invariance and rotation invariance of nodules, reducing misjudgments caused by differences in scanning position.

[0140] In ANN model training, the minimum radiation dose (ED) and image score (Q) are used. final Maximizing the pass rate (≥3 points) is a dual objective function, and the optimal parameter combination is found using the Pareto optimization algorithm. Specific constraints are as follows:

[0141] Effective dose ED ≤ 1.0 mSv (compliant with ACR guidelines);

[0142] Images with a subjective score of ≥3 points accounted for ≥95%;

[0143] Contrast-to-noise ratio (CNR) ≥ 3.0 (lung parenchyma-erector spinae contrast threshold).

[0144] The above descriptions are merely embodiments of the present invention, and common knowledge regarding specific structures and characteristics is not elaborated upon here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the structure of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A method for optimizing parameters of low-dose chest CT scans based on neural networks, characterized in that, Includes the following steps: S10: Implant lung nodules of different sizes into an in vitro liver model, set different scanning parameters for different models of scanning equipment to perform LDCT scans on the in vitro model to obtain LDCT images, use a scoring module to score the LDCT images, select qualified images from the LDCT images based on the scoring results, obtain the corresponding scanning equipment model, scanning parameters, nodule characteristics and radiation dose of the qualified images, and establish an ANN model. S20: Obtain the in vitro model corresponding to the qualified image as the initial in vitro model, generate the seed number through a random algorithm, and use the seed number to modify the lung nodule parameters in the initial in vitro model. Use the modified lung nodule parameters as the initial test parameters, use the ANN model to scan the initial in vitro model to obtain LDCT images, and identify the lung nodule size parameters as the second test parameters. Compare the initial test parameters and the second test parameters, and use the comparison results as feedback information to optimize the ANN model. S30: Train and optimize the ANN model by taking the individual factors of the examinee and the model of the scanning equipment as inputs, and the radiation dose and scanning parameters of the examinee as outputs. Normality is tested on the input data of the ANN model. One-way independent variables are selected from the input data based on correlation analysis and one-way ANOVA. Then, key variables are selected from the one-way independent variables using random forest. S40: Input the patient's individual factors and the scanning equipment model into the trained ANN model to obtain personalized scanning parameters and recommended radiation dose.

2. The method for optimizing chest low-dose CT scan parameters based on neural networks according to claim 1, characterized in that: In step S10, when implanting lung nodules into the in vitro model, simulated lung nodules with diameters of 3mm and 5mm are implanted into the PGH-1 in vitro model in stages according to lung segments; the different models of scanning equipment include at least 5 types; the scanning parameters include tube voltage, tube current, pitch, and gantry rotation speed; the individual factors of the examinee include gender, age, height, weight, blood type, fat content within the scanning range, longitudinal and transverse diameters of the thoracic cavity at the largest / central level, and the number of scanning layers.

3. The method for optimizing chest low-dose CT scan parameters based on neural networks according to claim 2, characterized in that: In step S10, SSDE is used to obtain the radiation dose. WED The method calculates the radiation dose; during calculation, an elliptical region of interest is delineated in the middle layer of the original sequence, the region of interest encompassing the entire transverse section, and the average CT value of the region of interest is recorded. ROI and area A ROI Calculate WED and conversion factor f size · MED The specific calculation formulas are shown in formulas (1)-(5) below: f size·MED =a×e -b×WED (2), SSD MED =f size·MED ×CTDI vol (3),DLP SSDE·MED =f size·MED ×DLP (4),ED SSDE·MED =k×DLP SSDE·MED (5), For adults, the values ​​of a and b are: a = 3.70469, b = 0.03671937; k = 0.014mSv· mGy -1 ·cm -1 。 4. The method for optimizing chest low-dose CT scan parameters based on neural networks according to claim 3, characterized in that: The scoring module in step S10 includes subjective evaluation and objective evaluation. The subjective evaluation uses a 5-point scoring standard to score the image quality. In the objective evaluation, three planes were selected, including the sternal jugular notch, the aortic arch, and the tracheal bifurcation plane. Circular regions of interest were drawn in the lung parenchyma, erector spinae muscles, and 1 cm in front of the chest wall at each plane. The average CT value and standard deviation of the regions of interest were recorded, and the signal-to-noise ratio and contrast-to-noise ratio of the images were calculated. The calculation formulas are shown in formulas (6) and (7) below: The signal-to-noise ratio of the image = mean CT value of lung parenchyma / SD of air CT value (6), and the contrast-to-noise ratio = (mean CT value of lung parenchyma - mean CT value of erector spinae muscles) / SD of erector spinae muscles (7). The scoring module combines subjective and objective evaluation results to calculate the scoring result of the LDCT image.

5. The method for optimizing chest low-dose CT scan parameters based on neural networks according to claim 4, characterized in that: In step S20, the vectorization of lung nodule parameters includes constructing the original parameter vector V from the lung nodule parameters without seed number modification, by taking the diameter of the lung nodule, its three-dimensional coordinates in the in vitro model, and its morphological features. init = [d,x,y,z,s], where d is the nodule diameter, x,y,z are the spatial coordinates, and s is the morphology coefficient; the seed number is generated by a random algorithm using a scaling factor σ and an offset Δ(x,y,z), which updates the original parameter vector to the initial test parameter vector V. adj = [d·σ,x+Δx,y+Δy,z+Δz,s], where (x,y,z) are spatial coordinate offset values; the second test parameter vector V obtained by recognizing LDCT images. det Includes measured values ​​for the corresponding dimensions; during comparison, first compare V with σ. det Perform scaling calibration, then calculate the Euclidean distance D = ||V|| between the two vectors. adj -V det ||2 and cosine similarity The comparison results of D and C are used as the comparison results of the initial test parameter and the second test parameter.

6. The method for optimizing chest low-dose CT scan parameters based on neural networks according to claim 5, characterized in that: In step S20, after obtaining the seed number, the LDCT image obtained after modifying the lung nodule parameters with the seed number is used as the verification image, and the LDCT image obtained without modifying the lung nodule parameters with the seed number is used as the initial image; during the correlation analysis, the initial image score Q is first calculated. init With verification image score Q cal Difference ΔQ=|Q cal -Q init | and construct an influence factor matrix M = [σ,Δx,Δy,Δz,ΔQ] by combining the seed number scaling factor σ and the offset Δ(x,y,z). Calculate the correlation between each element in the matrix and the scoring deviation using the Pearson coefficient, and select seed number parameters with a correlation coefficient |r|>0.6 as key influence factors. Dynamically adjust the weight coefficients of the scoring model based on these key influence factors, using the following adjustment formula: Where w i Let r be the weight of the i-th factor. i Its correlation coefficient.

7. The method for optimizing chest low-dose CT scan parameters based on neural networks according to claim 6, characterized in that: In step S10, an adaptive enhancement module is introduced when establishing the ANN model. The adaptive enhancement module automatically selects Gaussian filtering or nonlocal mean denoising algorithm to preprocess the LDCT image based on the hardware characteristics of the scanning device model and the nodule distribution density of the in vitro model. The hardware characteristics of the device model include detector resolution and X-ray tube power. When the device resolution is less than or equal to a preset resolution threshold, a multi-scale feature fusion method is enabled to enhance the feature extraction of lung nodules. The output of the adaptive enhancement module is used as the input layer data of the ANN model.

8. The method for optimizing chest low-dose CT scan parameters based on neural networks according to claim 6, characterized in that: In step S30, dynamic weight coefficients α are introduced during the training of the ANN model to construct a composite objective function L. total =α·L dose +(1-α)·L score L dose For radiation dose deviation loss, L score For image scoring loss; when the subject's age exceeds a preset age threshold or body mass index exceeds a preset body mass index, α is increased to preferentially reduce the radiation dose.

9. The method for optimizing chest low-dose CT scan parameters based on neural networks according to claim 6, characterized in that: In step S40, a multi-device compatibility adaptation mechanism is added before outputting personalized scanning parameters. By calling historical data from multiple reference devices that are associated with the core performance of the target device, a parameter compatibility evaluation system is constructed to quantitatively analyze the adaptability of scanning parameters across different devices. Based on the adaptability grading results, the parameter adjustment process is dynamically triggered. A correction coefficient model is established based on the differences in the core performance of the devices to optimize the cross-device adaptability of the scanning parameters. The differences in the core performance of the devices include hardware characteristics and algorithm types. The optimized parameters are closed-loop verified through preset diagnostic standards to ensure that the actual application of the parameters on the target device meets the dual requirements of image quality and radiation dose.

10. A neural network-based optimization system for low-dose chest CT scan parameters, characterized in that, The method for optimizing low-dose chest CT scan parameters based on neural networks, as described in any one of claims 1-9, was used.

Images