Pulmonary nodule sub-voxel volume VDT method based on deep learning

Abstract

The invention discloses a pulmonary nodule sub-voxel volume VDT method based on deep learning, and aims to solve the problem of adjustable mapping of depth model probability output to physical volume and the problem of deviation introduced by anisotropic voxels and reconstruction kernels. According to the method, direction-conditioned spatial variable point spread function distribution self-supervised estimation, edge spread function and modulation transfer function consistency constraint, global mapping and local residual error calibration of two-segment conditional flow under monotonous bounded constraint, and occupancy rate quantile output are carried out; and in combination with subvoxel integration and quality control closed-loop trigger conservative expansion and artificial recheck, the technical effects of cross-device and acquisition protocol robust volume and VDT point estimation and interval, uncertainty quantization and error reduction are realized.

Description

Technical Field

[0001] This invention relates to the field of medical image data processing, and in particular to a deep learning-based VDT method for lung nodules. Background Technology

[0002] With the widespread use of chest CT in the follow-up of pulmonary nodules, the volume doubling time, as an important indicator for assessing the growth rate of nodules, depends on the calculation of the volume at two time points and the acquisition time interval.

[0003] Volume measurement technology has evolved from binary segmentation based on thresholding and region growing, to semi-automatic 3D volume measurement, and then to deep learning segmentation models represented by 3D convolution and U-shaped networks, which can output nodule probability maps and achieve good segmentation performance on multi-center data.

[0004] To reduce volumetric bias, some studies have attempted methods such as intensity models, fuzzy classification, deconvolution, and super-resolution, but these often assume that the point spread function space is invariant and isotropic, and few systems combine imaging physics with DICOM parameters.

[0005] The shortcomings of existing technologies include:

[0006] 1. The probability output of the deep model lacks a calibrable mapping to the physical volume, and does not explicitly guarantee monotonicity and boundedness. Under partial volume and voxel anisotropy conditions, it is prone to systematic deviations in volume and volume doubling time.

[0007] 2. The point spread function modeling is insufficient, failing to consider the spatial variability caused by the boundary direction, lacking consistency constraints based on the edge spread function and modulation transfer function, as well as uncertainty distribution and confidence output, and lacking robustness across devices and acquisition protocols.

[0008] 3. Uncertainty and quality control are weak. Usually, only point estimates of volume and capacity doubling time are given, lacking quantile intervals and triggering strategies, making it difficult to take timely conservative expansion and manual review to reduce risks.

[0009] Therefore, a VDT method for lung nodules that can overcome the shortcomings of the prior art is a problem that needs to be solved by those skilled in the art. Summary of the Invention

[0010] One objective of this invention is to propose a deep learning-based VDT method for lung nodule subvoxel volume. Addressing the problems in existing technologies, such as the difficulty in calibrating the probability output of deep models to physical volume, volume deviation caused by anisotropic voxels and reconstruction kernels, and insufficient uncertainty and quality control, this invention proposes a self-supervised estimation of the spatially variable point spread function conditioned on DICOM parameters and edge orientation fields. It employs consistency constraints between the edge spread function and the modulation transfer function to construct a two-stage conditional flow that completes the global mapping from probability to occupancy under monotonically bounded constraints, and performs residual calibration locally on the nodule. It directly outputs the occupancy quantile and performs subvoxel integration on anisotropic voxels. Combined with a quality control closed loop, it implements conservative expansion and manual verification upon triggering. This invention offers the technical advantages of being robust across devices and acquisition protocols, significantly reducing systematic errors in volume and volume doubling time, providing interpretable uncertainty quantification, and offering more stable interval estimation.

[0011] A deep learning-based VDT method for lung nodule subvoxel volume according to an embodiment of the present invention is characterized by comprising:

[0012] S1. Based on chest CT body data at at least two time points and the corresponding DICOM protocol parameters, generate nodule probability maps at the first and second time points, detect the edges of the blood vessel wall and airway wall and calculate the edge direction, generate the edge direction field at the first and second time points, and calculate the acquisition time interval according to the acquisition date and time in the DICOM protocol parameters.

[0013] S2. Using the edge direction field as a condition, self-supervised learning is implemented under the chest CT body data and DICOM protocol parameters. The consistency constraint between the edge spread function in the spatial domain and the modulation transfer function in the frequency domain is adopted to establish the spatial variable point spread function distribution in a directional conditional manner, and generate the spatial variable point spread function distribution map and point spread function confidence map at the first time point and the second time point.

[0014] S3. Using the nodule probability map and the spatial variable point diffusion function distribution map at the first and second time points as conditions, and under the DICOM protocol parameters, construct the first segment of the two-segment conditional flow. Under monotonically bounded constraints, map the nodule probability to the occupancy rate on a voxel-by-voxel basis to obtain the initial value map of the global occupancy rate at the first and second time points.

[0015] S4. Based on the initial global occupancy rate map, locate the local area of ​​the nodule in the chest CT volume data and crop the local intensity block of the nodule. With the spatial variable point diffusion function distribution map and DICOM protocol parameters as conditions, construct the second segment of the two-segment conditional flow and perform local residual calibration. Under monotonically bounded constraints, directly generate the quantile results of the occupancy rate of each voxel, which are the lower quantile occupancy rate map, median occupancy rate map and upper quantile occupancy rate map at the first time point and the second time point, respectively.

[0016] S5. Under the anisotropic voxel conditions limited by the DICOM protocol parameters, perform sub-voxel integration on the median occupancy map to generate volume point estimates for the first and second time points. Calculate the volume intervals for the first and second time points based on the lower and upper quantile occupancy maps. Simultaneously, based on the statistical characteristics of the dispersion between occupancy quantile maps and the point spread function confidence map, generate occupancy quantile dispersion index and point spread function confidence statistical index.

[0017] S6. Calculate VDT point estimates based on volume point estimates and acquisition time intervals, and calculate VDT intervals based on volume intervals and acquisition time intervals.

[0018] S7. When the quality control status meets the preset triggering conditions, expand the volume interval between two time points according to the conservative strategy and recalculate the conservative VDT interval, while generating a manual review prompt.

[0019] Optionally, step S1 specifically includes:

[0020] Based on chest CT data from at least two time points and the corresponding DICOM protocol parameters, a deep learning ensemble segmentation model is used to generate a nodule probability map at the first time point and a nodule probability map at the second time point.

[0021] Based on chest CT scan data, the edges of the blood vessel wall and airway wall are detected and the edge direction is calculated to generate the edge direction field at the first time point and the edge direction field at the second time point.

[0022] The acquisition time interval is generated based on the acquisition date and time in the DICOM protocol parameters;

[0023] Output the nodule probability map at the first time point, the nodule probability map at the second time point, the edge direction field at the first time point, the edge direction field at the second time point, and the acquisition time interval.

[0024] Terminology definition:

[0025] The chest CT volume data is three-dimensional voxel data reconstructed based on DICOM sequences and aligned in the patient coordinate system. Voxel intensity is represented by HU values, and anisotropic voxels and different reconstruction kernels are allowed, and are associated with corresponding DICOM protocol parameters.

[0026] The DICOM protocol parameters are a set of parameters in the DICOM metadata associated with chest CT body data, used to describe spatial geometry, imaging physics and temporal information, including but not limited to fields such as pixel pitch, slice thickness and slice spacing, voxel orientation and position, reconstruction / convolution kernel, tube voltage / current, pitch and rotation time, equipment manufacturer / model, and acquisition / series date and time;

[0027] The nodule probability map is a three-dimensional probability field defined on the same voxel grid as the chest CT volume data. Each voxel takes a value in the range [0,1], which is used to represent the probability that the voxel is occupied by lung nodule tissue.

[0028] The edges of the blood vessel wall and airway wall are intensity transition regions between the blood vessel lumen or airway lumen and the surrounding tissue in chest CT, and their spatial location is obtained by edge detection methods based on intensity gradient, structural tensor or learning.

[0029] The edge direction is the local normal direction defined at the aforementioned edge location, and is taken as a unit vector along the direction of maximum increase in gray intensity, used to characterize the directional information of the boundary.

[0030] The edge direction field is a three-dimensional vector field of the same size and coordinate system as the chest CT volume data. It provides a unit normal direction at the voxels that are determined to be edges, and non-edge voxels are marked with a mask or left blank.

[0031] The acquisition date and time are a set of fields in the DICOM protocol parameters used to identify the time of a CT sequence acquisition. Series Date / Time is preferred, and if missing, Acquisition Date / Time or similar available fields are used.

[0032] The acquisition time interval is the difference between the effective acquisition times of two CT sequences at different time points, expressed in hours or days, and calculated from the acquisition date and time. When a single sequence contains multiple slice timestamps, the representative time of the sequence (such as the start time, series time, or manufacturer-defined representative time) is used as the effective acquisition time for that time point.

[0033] The deep learning ensemble segmentation model is a set of models consisting of two or more deep neural network segmentation sub-models, which output a nodule probability map aligned with chest CT body data through fusion strategies (including but not limited to weighted average, voting, or stacking).

[0034] Optionally, step S2 specifically includes:

[0035] Based on the edge direction fields at the first and second time points, self-supervised learning is implemented under the conditions of chest CT volume data and DICOM protocol parameters. The distribution expression of the spatial variable point spread function is established by joint optimization with the consistency constraints of the edge spread function in the spatial domain and the consistency constraints of the modulation transfer function in the frequency domain, and the distribution map of the spatial variable point spread function at the first and second time points is generated. The point spread function distribution of each voxel is parameterized by the mean and covariance.

[0036] Simultaneously, based on the uncertainty assessment during the estimation process, confidence maps of the point spread function at the first and second time points are generated.

[0037] Output the spatial variable point diffusion function distribution map at the first time point, the spatial variable point diffusion function distribution map at the second time point, the point diffusion function confidence map at the first time point, and the point diffusion function confidence map at the second time point.

[0038] Terminology definition:

[0039] The self-supervised learning is a learning process that does not rely on manual annotation, uses the consistency between the spatial domain edge spread function and the frequency domain modulation transfer function as the supervision signal, and estimates the distribution of the spatial variable point spread function by minimizing the consistency loss.

[0040] The edge spread function is an intensity curve describing the response of the imaging system to an ideal step edge. It is sampled from chest CT volume data along the edge normal direction and is used to reflect the degree of local blur and provide spatial constraints for point spread function estimation.

[0041] The modulation transfer function is the spatial frequency response amplitude of the imaging system, which can be obtained by the Fourier transform of the point spread function or the line spread function. It is used to reflect the transfer capability at different spatial frequencies and to provide frequency domain constraints for the estimation of the point spread function.

[0042] The consistency constraint is a constraint that aligns the edge spread function and modulation transfer function derived from the estimated point spread function with the observed or prior ESF / MTF, respectively. It is implemented through a loss function or regularization term to ensure physical consistency between the spatial domain and the frequency domain.

[0043] The directional conditionalization uses directional information such as edge direction field as conditional input or conditional variable, so that the estimation of the point spread function changes with the local boundary direction, in order to characterize the directional anisotropic fuzziness.

[0044] The point spread function is the unit impulse response or equivalent blur kernel of the imaging system, used to characterize local spatial resolution and part of the volume effect.

[0045] The spatially variable point spread function is a point spread function that varies with voxel position and local direction in chest CT volume data, so that the degree of ambiguity at different positions or directions can be modeled differently.

[0046] The distribution of the point spread function is expressed as a modeling expression of the uncertainty distribution of the point spread function parameters of each voxel, so as to give the center value and dispersion of the parameter estimate at the same time.

[0047] The spatially variable point diffusion function distribution map is a three-dimensional parameter field aligned with the chest CT volume data. The parameters of the point diffusion function distribution are stored at each voxel for subsequent occupancy mapping and local calibration.

[0048] The point spread function distribution "parameterized by mean and covariance" is a way of characterizing the probability distribution of the point spread function parameters using a mean vector and a covariance matrix. The mean represents the expected value of the parameter, and the covariance represents the uncertainty and correlation of the parameter. The parameters may include, but are not limited to, the scale, orientation and shape of the anisotropic kernel.

[0049] The point spread function confidence plot is a three-dimensional scalar field aligned with the chest CT volume data, reflecting the confidence level of the spread function estimate for each voxel. It is generated based on statistics such as uncertainty assessment and consistency residuals, and a higher value indicates a higher confidence level.

[0050] The uncertainty assessment is the process of estimating the uncertainty of the point spread function, which can be implemented based on the model's probability output, parameter covariance, ensemble or randomization methods, and used to generate a point spread function confidence map.

[0051] Optionally, step S3 specifically includes:

[0052] Based on the nodule probability map at the first time point and the nodule probability map at the second time point, as well as the spatial variable point diffusion function distribution map at the first time point and the spatial variable point diffusion function distribution map at the second time point, the first segment of the two-segment conditional flow is constructed under the conditions of the DICOM protocol parameters. The nodule probability is converted into the occupancy rate on a voxel-by-voxel basis through the global baseline mapping subject to monotonically bounded constraints. The monotonically bounded constraints are used to ensure that the occupancy rate is between zero and one and does not decrease monotonically with the nodule probability. The initial value map of the global occupancy rate at the first time point and the initial value map of the global occupancy rate at the second time point are generated.

[0053] Output the initial global occupancy value map at the first time point and the initial global occupancy value map at the second time point.

[0054] Terminology definition:

[0055] The two-stage conditional flow is a model structure consisting of two transformations that maps the input to the output voxel by voxel under given conditional variables (including nodule probability map, spatial variable point diffusion function distribution map and DICOM protocol parameters). The first stage realizes global baseline mapping, and the second stage is used for subsequent local residual calibration.

[0056] The first segment of the two-stage conditional flow is a global mapping module that is uniformly applied to all voxels under the above-mentioned conditional variables, used to convert the nodule probability into an initial estimate of the occupancy rate on a voxel-by-voxel basis.

[0057] The global baseline mapping is a family of parameterized mapping functions from probability to occupancy, implemented by the first segment of a two-segment conditional flow. This mapping maintains a consistent global baseline effect under different DICOM parameters and point spread function conditions.

[0058] The monotonically bounded constraint is a set of functional constraints applied to the probability-to-occupancy mapping, used to ensure that the occupancy output value is in the interval [0,1] and does not decrease monotonically with the nodule probability.

[0059] The occupancy rate is an estimated value of the volume fraction occupied by the target lung nodule tissue per voxel, with a value range of [0,1], and is used for subsequent subvoxel integration and interval calculation of volume.

[0060] The initial global occupancy map is a three-dimensional occupancy field generated from the first segment of the two-segment conditional flow under monotonically bounded constraints on chest CT volume data, serving as the input and reference for subsequent local residual calibration.

[0061] Optionally, step S4 specifically includes:

[0062] Based on the initial global occupancy rate map at the first time point and the initial global occupancy rate map at the second time point, the local area of ​​the nodule is located in the chest CT body data and the local intensity block of the nodule is cropped.

[0063] Under the conditions of the spatial variable point diffusion function distribution map at the first time point, the spatial variable point diffusion function distribution map at the second time point, and the DICOM protocol parameters, the second segment of the two-segment conditional flow is constructed and local residual calibration is implemented. Under monotonically bounded constraints, the quantile results of the occupancy rate of each voxel are directly generated, and the lower quantile occupancy rate map, the median occupancy rate map, the upper quantile occupancy rate map at the first time point, the lower quantile occupancy rate map, the median occupancy rate map, and the upper quantile occupancy rate map at the second time point are generated respectively.

[0064] Output the occupancy rates of the lower quantile at the first time point, the median at the first time point, the upper quantile at the first time point, the lower quantile at the second time point, the median at the second time point, and the upper quantile at the second time point.

[0065] Terminology definition:

[0066] The nodule local region is a three-dimensional region of interest in the chest CT volume data, which includes the target lung nodule and its surrounding safety boundary, and is used to carry out subsequent local calibration and quantile estimation based on the global occupancy initial value map.

[0067] The nodule local intensity block is a three-dimensional sub-volume data cropped from chest CT volume data with the nodule local region as the center. It retains the original intensity of the voxels in the region and aligns with the relevant conditional variables, and is used as the second input of the two-stage conditional flow.

[0068] The second stage of the two-stage conditional flow is a local transformation module that performs conditional residual mapping on the initial global occupancy rate within the nodule local intensity block, and outputs the quantiles of the occupancy rate per voxel based on the spatially variant point spread function distribution map and DICOM protocol parameters;

[0069] The local residual calibration is a process in the second stage of the two-stage conditional flow that takes the nodule local intensity block as the input and the initial global occupancy rate as the reference, learns and corrects the local deviation of the occupancy rate, and generates a quantile output by fitting the residual distribution between the initial value and the true occupancy rate and being subject to a monotonic bounded constraint;

[0070] The quantile result is a set of statistical values of the conditional distribution of the occupancy rate per voxel at a preset quantile level, including at least the median, the lower quantile, and the upper quantile, and is used to form various occupancy rate quantile maps;

[0071] The lower quantile occupancy rate map is a three-dimensional scalar field aligned with the chest CT volume data, and the voxel value thereof is the estimated value of the occupancy rate of the voxel at the preset lower quantile level q_low (0 < q_low < 0.5), and is used to represent the lower limit;

[0072] The median occupancy rate map is a three-dimensional scalar field aligned with the chest CT volume data, and the voxel value thereof is the estimated value of the occupancy rate of the voxel at the 0.5 quantile, and is used to represent the median level;

[0073] The upper quantile occupancy rate map is a three-dimensional scalar field aligned with the chest CT volume data, and the voxel value thereof is the estimated value of the occupancy rate of the voxel at the preset upper quantile level q_high (0.5 < q_high < 1), and is used to represent the upper limit;

[0074] The quantile level is a set of probability quantile parameters for generating the quantile result, including at least q_low, 0.5, and q_high, and the values are preset by the algorithm or determined during the training process and the specific values are not limited.

[0075] Optionally, step S5 is specifically:

[0076] Based on the median occupancy rate map at the first time point and the median occupancy rate map at the second time point, sub-voxel integration is performed on the anisotropic voxels under the voxel size and slice thickness conditions defined by the DICOM protocol parameters to generate the volume point estimate at the first time point and the volume point estimate at the second time point;

[0077] Calculate the volume interval at the first time point based on the lower quantile occupancy rate map at the first time point and the upper quantile occupancy rate map at the first time point under the same integration conditions, and calculate the volume interval at the second time point based on the lower quantile occupancy rate map at the second time point and the upper quantile occupancy rate map at the second time point under the same integration conditions;

[0078] Based on the dispersion between occupancy quantile plots, an occupancy quantile dispersion index is generated, and based on the statistical characteristics of the point spread function confidence index, a point spread function confidence index is generated.

[0079] Output the volume point estimate at the first time point, the volume point estimate at the second time point, the volume interval at the first time point, the volume interval at the second time point, the occupancy rate quantile dispersion index, and the point spread function confidence index.

[0080] Terminology definition:

[0081] The anisotropic voxels are voxel settings in three-dimensional volume data where the voxel spacing in each principal axis direction is not equal. The voxel spacing in the x and y directions comes from DICOM's PixelSpacing, and the voxel spacing in the z direction comes from SliceThickness or SpacingBetweenSlices.

[0082] The voxel size is the set of physical lengths of the voxel in the x, y, and z directions, given by the pixel pitch and layer thickness fields of DICOM, and used to determine the physical scale and integral weight of the voxel.

[0083] The layer thickness is a physical thickness value along the stacking direction, derived from the DICOM field SliceThickness or its equivalent field, and the unit is millimeters;

[0084] The subvoxel integral is a process of subdividing and integrating the occupancy rate within each voxel under anisotropic voxel geometry to estimate the physical volume contribution. Specifically, it can be achieved through regular subdivision within voxels, random sampling, or analytical approximation, but is not limited to these methods.

[0085] The volume point estimate is a single-value estimate obtained by multiplying the occupancy rate of each voxel within the nodule space by the physical volume of the corresponding voxel and summing the results. The unit is cubic millimeters or milliliters.

[0086] The volume range is an ordered pair of the lower and upper limits of the volume obtained by performing subvoxel integration on the lower quantile occupancy map and the upper quantile occupancy map respectively under the same integration conditions, used to characterize the uncertainty range of the volume.

[0087] The same integration condition means that the same voxel geometric parameters, integral grid subdivision scale and integration method are used for different quantile occupancy rate maps at the same time point to ensure the comparability and consistency of the upper and lower limit calculations.

[0088] The quantile dispersion index is a statistical measure of the degree of difference between the lower quantile, median and upper quantile occupancy maps within the nodule region. It is usually obtained by aggregating (such as mean, quantile or weighted) the voxel-level quantile difference (such as the difference between the upper quantile and the lower quantile). The larger the value, the higher the uncertainty.

[0089] The point spread function confidence statistic is a statistical measure obtained by summarizing the point spread function confidence maps within the nodule region. It is used to reflect the overall confidence of the PSF estimate and includes, but is not limited to, the mean, median value, quantiles, and the proportion of low-confidence voxels.

[0090] Optionally, step S6 specifically includes:

[0091] Based on the volume point estimates at the first and second time points and the acquisition time interval, VDT point estimates are generated according to the calculation relationship that VDT point estimates are calculated as the natural logarithm of the acquisition time interval multiplied by a constant 2 divided by the natural logarithm of the ratio of the volume point estimates at the second and first time points.

[0092] Based on the volume intervals at the first and second time points and the data collection time interval, the corresponding VDT lower limit and VDT upper limit are calculated by pairing the lower and upper limits of the volume intervals to generate the VDT intervals.

[0093] Output VDT point estimates and VDT intervals.

[0094] Terminology definition:

[0095] The volume doubling time (VDT) is an indicator of the time required for the nodule volume to double, calculated based on the volume at two time points and the time interval between acquisitions. It uses the natural logarithm to measure the growth ratio, and the unit is time.

[0096] The VDT point estimate is a single-value result calculated based on the volume point estimate at the first time point, the volume point estimate at the second time point, and the acquisition time interval, according to VDT = acquisition time interval × ln(2) / ln(volume point estimate at the second time point / volume point estimate at the first time point);

[0097] The VDT interval is an ordered pair of lower and upper limits calculated by substituting the lower and upper limits of each time point volume interval into the VDT formula, based on the first time point volume interval, the second time point volume interval, and the collection time interval.

[0098] The lower limit of VDT is the lower endpoint of the interval calculated by pairing the lower limit of the volume interval at the first time point with the lower limit of the volume interval at the second time point according to the VDT formula.

[0099] The upper limit of VDT is the upper endpoint of the interval calculated by pairing the upper limit of the volume interval at the first time point with the upper limit of the volume interval at the second time point according to the VDT formula.

[0100] The volume ratio is the ratio of the volume point estimate at the second time point to the volume point estimate at the first time point, and is used in the logarithmic term of the VDT formula;

[0101] The pairing calculation is a rule that substitutes the same type of endpoints (lower limit to lower limit, upper limit to upper limit) of the volume intervals of the first time point and the second time point into the VDT formula for calculation.

[0102] Optionally, step S7 specifically includes:

[0103] The width of the VDT interval is calculated based on the VDT point estimate and the VDT interval.

[0104] Based on the comparison of the quantile dispersion index of the occupancy rate and the confidence statistical index of the point spread function with preset thresholds respectively, when the quantile dispersion index of the occupancy rate is greater than the first threshold, or the confidence statistical index of the point spread function is lower than the second threshold, or the VDT interval width is greater than the third threshold, the quality control state is set to the triggered state; otherwise, it is set to the non-triggered state, and the quality control state is generated.

[0105] When the quality control status is triggered, a conservative volume range is generated based on the volume range at the first time point and the volume range at the second time point according to the conservative strategy of expanding the upper and lower limits of the range. The conservative volume range includes the conservative volume range at the first time point and the conservative volume range at the second time point. The upper and lower limits of VDT are recalculated based on the conservative volume range at the first time point and the conservative volume range at the second time point to generate a conservative VDT range. At the same time, a manual review prompt is generated.

[0106] When the quality control status is non-triggered, conservative volume range, conservative VDT range, and manual review prompts are not generated.

[0107] The output is in quality control mode. When triggered, it simultaneously outputs the conservative volume range, the conservative VDT range, and a manual review prompt.

[0108] Terminology definition:

[0109] The VDT interval width is a non-negative number representing the difference between the upper and lower limits of the VDT interval, used to measure the magnitude of VDT uncertainty;

[0110] The preset threshold is a set of numerical parameters used for quality control judgment, which can be set by training and validation statistics, clinical priors or system configuration.

[0111] The first threshold is a trigger threshold for comparison with the occupancy quantile dispersion index;

[0112] The second threshold is a trigger threshold for comparison with the confidence statistic of the point spread function;

[0113] The third threshold is a trigger threshold that is compared with the width of the VDT interval;

[0114] The preset triggering conditions are a set of logical rules used to determine whether the quality control state is set to the triggering state, including at least one of the following conditions: the occupancy rate quantile dispersion index is greater than the first threshold, the point spread function confidence statistical index is lower than the second threshold, or the VDT interval width is greater than the third threshold.

[0115] The quality control status is the result of the judgment indicating whether to enter the conservative expansion and manual review process, and the value includes triggered status and non-triggered status;

[0116] The trigger state is a value of the quality control state, indicating that the preset trigger conditions are met, conservative expansion should be performed and a manual review prompt should be generated;

[0117] The non-triggering state is a value of the quality control state, indicating that the preset triggering conditions are not met, and neither conservative expansion is executed nor a manual review prompt is generated.

[0118] The conservative strategy expansion is a set of rules that expand the upper and lower limits of the volume range outward in order to reduce the risk of underestimation under the triggered state. These rules include, but are not limited to, increasing the range width by a fixed ratio, applying a non-positive offset to the lower limit and a non-negative offset to the upper limit, or setting an expansion coefficient according to the index layer. The expansion shall satisfy that the lower limit after expansion is not greater than the original lower limit and the upper limit is not less than the original upper limit.

[0119] The upper and lower limits of the expansion range are the new lower and upper limits of the volume range obtained after expansion using a conservative strategy.

[0120] The conservative volume interval is a set of volume intervals obtained by expanding the volume intervals at two time points respectively under the triggered state using a conservative strategy, including the conservative volume interval at the first time point and the conservative volume interval at the second time point.

[0121] The first time point conservative volume interval is the interval obtained by expanding the first time point volume interval outward according to a conservative strategy;

[0122] The second time point conservative volume interval is the interval obtained by expanding the second time point volume interval outward according to a conservative strategy;

[0123] The conservative VDT interval is the upper and lower limits of VDT calculated by pairing the same type of endpoints of the conservative volume interval at the first time point and the conservative volume interval at the second time point and substituting them into the VDT formula.

[0124] The manual review prompt is a reminder message output when the quality control status is in the triggered state, intended for manual review. It is used to indicate that manual inspection, confirmation or correction is required, and can take the form of interface marker, message or task item.

[0125] The beneficial effects of this invention are:

[0126] 1. Calibrable monotonic mapping from probability to occupancy rate: Through global baseline mapping and local residual calibration of two-stage conditional flow, and by applying monotonic bounded constraints, the systematic bias of volume and volume doubling time is significantly reduced under anisotropic voxels and different reconstruction kernel conditions, thereby improving the consistency and accuracy of cross-time point calculations.

[0127] 2. Imaging Physics Adaptation and Robustness Enhancement: Based on the spatially variable point spread function distribution with orientation conditionalization, and introducing consistency constraints between the edge spread function and the modulation transfer function, the uncertainty and point spread function confidence map are output to achieve adaptive modeling of differences between devices and acquisition protocols, thereby enhancing robustness across devices and protocols.

[0128] 3. Uncertainty Quantification and Quality Control Closed Loop: Directly output the occupancy quantile to form volume range and volume doubling time range. Combined with the quality control triggering mechanism of quantile dispersion and point spread function confidence, it automatically performs conservative expansion and prompts manual review, providing interpretable uncertainty management and reducing clinical risk. Attached Figure Description

[0129] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0130] Figure 1 This is a flowchart of a deep learning-based VDT method for lung nodules. Detailed Implementation

[0131] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0132] refer to Figure 1 A deep learning-based VDT method for lung nodule subvoxel volume, characterized by comprising:

[0133] S1. Based on chest CT body data at at least two time points and the corresponding DICOM protocol parameters, generate nodule probability maps at the first and second time points, detect the edges of the blood vessel wall and airway wall and calculate the edge direction, generate the edge direction field at the first and second time points, and calculate the acquisition time interval according to the acquisition date and time in the DICOM protocol parameters.

[0134] S2. Using the edge direction field as a condition, self-supervised learning is implemented under the chest CT body data and DICOM protocol parameters. The consistency constraint between the edge spread function in the spatial domain and the modulation transfer function in the frequency domain is adopted to establish the spatial variable point spread function distribution in a directional conditional manner, and generate the spatial variable point spread function distribution map and point spread function confidence map at the first time point and the second time point.

[0135] S3. Using the nodule probability map and the spatial variable point diffusion function distribution map at the first and second time points as conditions, and under the DICOM protocol parameters, construct the first segment of the two-segment conditional flow. Under monotonically bounded constraints, map the nodule probability to the occupancy rate on a voxel-by-voxel basis to obtain the initial value map of the global occupancy rate at the first and second time points.

[0136] S4. Based on the initial global occupancy rate map, locate the local area of ​​the nodule in the chest CT volume data and crop the local intensity block of the nodule. With the spatial variable point diffusion function distribution map and DICOM protocol parameters as conditions, construct the second segment of the two-segment conditional flow and perform local residual calibration. Under monotonically bounded constraints, directly generate the quantile results of the occupancy rate of each voxel, which are the lower quantile occupancy rate map, median occupancy rate map and upper quantile occupancy rate map at the first time point and the second time point, respectively.

[0137] S5. Under the anisotropic voxel conditions limited by the DICOM protocol parameters, perform sub-voxel integration on the median occupancy map to generate volume point estimates for the first and second time points. Calculate the volume intervals for the first and second time points based on the lower and upper quantile occupancy maps. Simultaneously, based on the statistical characteristics of the dispersion between occupancy quantile maps and the point spread function confidence map, generate occupancy quantile dispersion index and point spread function confidence statistical index.

[0138] S6. Calculate VDT point estimates based on volume point estimates and acquisition time intervals, and calculate VDT intervals based on volume intervals and acquisition time intervals.

[0139] S7. When the quality control status meets the preset triggering conditions, expand the volume interval between two time points according to the conservative strategy and recalculate the conservative VDT interval, while generating a manual review prompt.

[0140] In this specific embodiment, S1 specifically refers to:

[0141] The system first reads chest CT volume data from at least two time points and aligns them in the patient coordinate system. At the same time, it parses the set of DICOM protocol parameters associated with each time point so that subsequent processing can be carried out on a unified voxel grid.

[0142] Time index is Representing chest CT data as a three-dimensional intensity field (Voxel intensity is in HU), the DICOM protocol parameter set is represented as follows: (Includes fields such as pixel pitch, layer thickness / layer spacing, voxel orientation and position, reconstruction / convolution kernel, tube voltage / current, pitch and rotation time, equipment manufacturer / model, and series date / time or acquisition date / time).

[0143] When generating the nodule probability map, a deep learning ensemble segmentation model is used, denoted as . Under parameter conditions Below Perform inference, and output the value that is aligned with the original voxel grid and has a range of values. Nodule probability plot The calculation relationship is as follows:

[0144] ;

[0145] in This represents an ensemble model that combines multiple segmentation sub-models. Indicates the first Three-dimensional CT intensity field at time point Indicates the first DICOM parameter set at a given time point Indicates the first Nodal probability plot at time points;

[0146] Subsequently, chest CT scan data Detecting the edges of blood vessel walls and airway walls and constructing edge voxel masks For each voxel location belonging to the edge (3D voxel coordinates) Calculate the unit normal vector along the direction of maximum grayscale increase to obtain the edge direction field. Its definition is:

[0147] ;

[0148] in Indicates the location Intensity gradient at the location, Represents the Euclidean norm, Indicates the first Edge mask at time point, Indicates the first Edge direction field at a given time point;

[0149] Non-edge voxels are masked to preserve the physical meaning of the orientation field;

[0150] At the same time from Extract valid acquisition time (SeriesDate / SeriesTime is preferred; if missing, AcquisitionDate / AcquisitionTime or similar available fields are used, and all are converted to real time in hours or days.) The data collection time interval is calculated accordingly.

[0151] ;

[0152] in Indicates the effective data collection time at the first time point. Indicates the effective data collection time at the second time point. This indicates the time interval between two data collection points.

[0153] The final output is the nodule probability map at the first time point. Second time point nodule probability plot Field at the edge direction at the first time point Field at the edge of the second time point and the time interval between data collection .

[0154] In this specific embodiment, S2 specifically refers to:

[0155] The system uses the edge direction field at the first and second time points, chest CT body data, and DICOM protocol parameters as conditions. Through self-supervised learning, under the consistency constraints of the edge spread function in the spatial domain and the modulation transfer function in the frequency domain, it establishes the distribution expression of the spatially variable point spread function in a directional conditional manner and outputs the point spread function confidence map.

[0156] Time index is Chest CT data are represented as a three-dimensional intensity field. (Voxel intensity is in HU), the DICOM protocol parameter set is represented as follows: (Including fields such as pixel pitch, layer thickness or interlayer spacing, voxel orientation and position, reconstruction or convolution kernel, tube voltage or current, pitch and rotation time, device manufacturer and model, and series date or time or acquisition date or time), the edge orientation field is represented as (At voxels identified as edges, a unit normal vector along the direction of maximum grayscale increase is given; non-edge voxels are left empty using a mask). The edge voxel mask is represented as follows: (A collection of marginal voxels that identify the walls of blood vessels and airways).

[0157] Under the above conditions, a point spread function distribution estimator is constructed and inference is performed based on voxel positions to obtain the mean and covariance of the parameter distribution. The core relationship is as follows:

[0158] ;

[0159] in Indicates the location The mean vector of the spread function parameters at a point in (3D voxel coordinates), Indicates the location The covariance matrix of the point spread function parameters Indicates and Conditional self-supervised estimation model Indicates the voxel location for estimation;

[0160] To construct a self-supervised signal, the system in Inner edge The direction is correct The edge spread function of the observation is obtained by sampling, and based on the... The predicted edge spread function is derived from the determined local kernel, and the predicted modulation transfer function is obtained by the Fourier transform amplitude of the kernel. The observed modulation transfer function is estimated by the spectrum of the local block or by the numerical derivative of the observed edge spread function and the Fourier relation.

[0161] The joint optimization objective is adopted as a weighted sum of space-frequency consistency terms and space regularization terms:

[0162] ;

[0163] in Indicates total loss under self-monitoring, Indicates the weight of spatial consistency loss. The weighted mean squared error loss between the observed and predicted marginal spread functions. Indicates the frequency domain consistency loss weights, This represents the weighted mean square error loss between the observed and predicted modulation transfer functions. Indicates space regularization loss weights, Indicates to Penalty terms for applying spatial smoothing or prior constraints;

[0164] During training, uncertainty and consistency residuals are simultaneously quantified and estimated, and a point spread function confidence plot is constructed, defined as:

[0165] ;

[0166] in Indicates the location Confidence of point spread function at point Represents exponential functions, Represents the non-negative weighting coefficients for parameter uncertainty. Represents the trace operation of a matrix. Represents the parameter covariance matrix, Represents the non-negative weighting coefficient for consistency residuals. This represents the non-negative residual index obtained by normalizing the space-frequency consistency error;

[0167] After the optimization converges, and The output, in the form of a set of all voxel locations, is a spatially variable point diffusion function distribution map of the first and second time points, and... The output, in the form of a set of all voxel locations, is a confidence plot of the point spread function at the first and second time points.

[0168] In this specific embodiment, S3 specifically refers to:

[0169] Based on the nodule probability maps at the first and second time points, the spatial variable point diffusion function distribution map, and the DICOM protocol parameters, the system constructs the first segment of the two-segment conditional flow as a global baseline mapping module. Under monotonically bounded constraints, the nodule probability is converted into an initial estimate of the occupancy rate on a voxel-by-voxel basis, and the initial value map of the global occupancy rate is output.

[0170] Time index is The voxel position is (3D voxel coordinates), nodule probability map is (No. Time point in location The probability of nodules, taking values ​​in the range of The mean and covariance of the point spread function distribution parameters are respectively and (No. Time point in location The center and discreteness of the uncertainty distribution of the point spread function parameters), the DICOM protocol parameter set is (No. (Set of geometric and imaging physical parameters at time points);

[0171] The global baseline mapping employs a constrained family of monotone bounded functions, with the core relation written as:

[0172] ;

[0173] in Indicates the first Time point in location Initial value of global occupancy This represents the Sigmoid function with a range of 0. Guaranteed bounded Indicates the condition variable The generated non-negative slope ensures monotonicity. This indicates that the calibration offset is generated from the same condition variable to achieve consistency across devices and acquisition protocols;

[0174] To clearly define the constraints, the system explicitly applies:

[0175] and as well as ;

[0176] in This indicates the boundedness of the occupancy output. This indicates that the probability of nodules does not decrease monotonically. This indicates that the slope is non-negative to ensure monotonicity;

[0177] This mapping is uniformly applied to all voxels and... Conditionalize its shape parameters to obtain consistent global baseline calibration under anisotropic voxels and different reconstruction kernel conditions, ultimately... The output, in the form of a set of all voxel positions, is a global occupancy initial value map for the first and second time points.

[0178] In this specific embodiment, S4 specifically refers to:

[0179] Under the DICOM protocol parameters, based on the initial global occupancy map at the first and second time points, chest CT volume data, and spatial variable point diffusion function distribution map, the system first locates the local area of ​​the nodule and cuts the local intensity block of the nodule as the second input of the two-stage conditional flow, thereby completing the local residual calibration under monotonically bounded constraints and directly outputting the occupancy quantile result of each voxel.

[0180] Time index is The voxel position is (3D voxel coordinates), chest CT volume data is (No. (3D intensity field at time point), initial global occupancy rate map is as follows ( (The first one obtained from step S3) Time point in location (Initial value of occupancy rate), the mean and covariance of the point spread function distribution parameter field are respectively and (No. The mean vector field and covariance matrix field of the spatially variable point spread function parameter distribution at time points implicitly reflect the mean vector field and covariance matrix field of the spread function parameter distribution at time points. (Changes), the DICOM protocol parameter set is (No. (Set of geometric and imaging physical parameters at time points);

[0181] The local region of the nodule is obtained by thresholding the initial occupancy rate and applying a morphological safety boundary. The core relationship is:

[0182] ;

[0183] in Indicates the first voxel set of local region of nodule at time point, Indicated by radius Morphological dilation operator, This indicates the occupancy threshold used for initial screening. This represents the non-negative expansion radius used to preserve the safety boundary;

[0184] In the region After clipping and aligning the condition variables, we get:

[0185] and ;

[0186] in Indicates will Limited to Local strength blocks of nodules Indicates will Limited to The initial value block of local occupancy rate;

[0187] Subsequently, using the point spread function distribution parameter field and DICOM protocol parameters as conditions, the second-stage mapping operator of the two-stage conditional flow is constructed and applied to... Perform local residual calibration to directly output the estimate of the occupancy of each voxel at the preset quantile level. The core relation is written as follows:

[0188] ;

[0189] in The second mapping operator represents the two-stage conditional flow. Indicates the first Time point in location The lower quantile occupancy rate (corresponding quantile level) ), Indicates the first Time point in location Median occupancy (corresponding to 0.5 percentile) Indicates the first Time point in location Upper quantile occupancy (corresponding quantile level) );

[0190] To satisfy the monotone bounded constraint and ensure the consistent order of quantiles, the system explicitly imposes the following on the output:

[0191] ;

[0192] in Indicates quantile level index Under the occupancy rate Indicates to The partial derivative is used to characterize the property that it is monotonically non-decreasing with the initial value of the occupancy rate;

[0193] Ultimately and The data is output as a three-dimensional scalar field aligned with the chest CT volume data, and as lower quantile occupancy map, median occupancy map, and upper quantile occupancy map for the first and second time points, respectively.

[0194] In this specific embodiment, S5 specifically includes:

[0195] Based on the median occupancy rate map, lower quantile occupancy rate map, and upper quantile occupancy rate map at the first and second time points, the system performs sub-voxel integration on anisotropic voxels under the voxel size and layer thickness conditions limited by the DICOM protocol parameters to generate volume point estimates and volume intervals, and summarizes the occupancy rate quantile dispersion index and point spread function confidence index in the local region of the nodule.

[0196] Time index is The voxel position is (3D voxel coordinates), the voxel set of the local region of the nodule is (Region of interest determined by step S4), median occupancy rate The lower quantile occupancy rate The occupancy rate of the upper quantile is The DICOM protocol parameter set is ;

[0197] from Extract the physical length of voxels in each principal axis direction and represent it at the voxel level. and Therefore, the physical volume of a voxel is defined as:

[0198] ;

[0199] in Indicates the first Time point in location The physical volume of voxels and These represent the results obtained by mapping PixelSpacing to SliceThickness or SpacingBetweenSlices, respectively. Directional voxel physical length, Indicates the first The set of DICOM parameters at a given time point;

[0200] In the local area of ​​the nodule Within this, subvollar integration is performed on the median occupancy to obtain a volume point estimate. The core relationship is:

[0201] ;

[0202] in Indicates the first Volume point estimation at time points, This represents the summation operator over the positions of voxels within a set. Indicates the location The median occupancy rate Indicates the location The physical volume of voxels This represents the set of local regions of nodules used for integration;

[0203] Under the same integration conditions, the occupancy rates of the lower and upper quantiles are used to calculate the endpoints of the volume interval, written as:

[0204] ;

[0205] in Indicates the first The lower limit of the volume range at a given time point. Indicates the first Upper limit of volume range at a given time point and They represent the positions respectively. The occupancy rates of the lower and upper quantiles;

[0206] To quantify the uncertainty of occupancy quantiles, a volume-weighted occupancy quantile dispersion index is constructed, written as:

[0207] ;

[0208] in Indicates the first The occupancy quantile dispersion index at a given time point has the numerator being the weighted sum of the voxel-level quantile differences and voxel volumes, and the denominator being the weighted sum of voxel volumes.

[0209] Simultaneously, based on the statistical characteristics of the point spread function confidence map within the local region of the nodule, a volume-weighted point spread function confidence statistic is defined, written as:

[0210] ;

[0211] in Indicates the first The confidence statistic of the point spread function at a given point in time. Indicates the location The confidence level of the point spread function. and The meaning is the same as before;

[0212] Ultimately and The output is the volume point estimate for the first time point and the second time point. and The output is the volume range between the first and second time points, and also outputs... and For use in subsequent quality control.

[0213] In this specific embodiment, S6 specifically refers to:

[0214] The system calculates the point estimate and interval of the volume doubling time based on the volume point estimate and volume interval at the first and second time points, combined with the collection time interval.

[0215] Time index is The volume point estimates at the first and second time points are denoted as follows: and (Single-valued result obtained from the subvocetary integral of the median occupancy rate, in units of volume), the corresponding lower and upper limits of the volume interval are denoted as follows: and (The endpoints of the interval are obtained by subvoxel integrals of the occupancy rates of the lower and upper quantiles, in units of volume), and the acquisition time interval is denoted as... (The difference in effective acquisition time calculated from two DICOM timestamps, in hours or days), the natural logarithm operator is denoted as... (by (logarithm to base), the constant 2 represents the proportion by which the volume doubles;

[0216] Under the above notation, the point estimate of the volume doubling time is calculated as follows:

[0217] VDT ;

[0218] Where VDT represents the point estimate of the volume doubling time, Indicates the time interval for data collection. Represents the natural logarithm of the constant 2. The natural logarithm of the ratio of the volume at the second time point to the volume at the first time point;

[0219] To obtain an interval estimate of the volume doubling time, the lower and upper limits are calculated separately according to the rule of pairing endpoints of the same type, written as:

[0220] ;

[0221] in The lower limit of the interval representing the time for volume doubling. The upper limit of the interval representing the time for volume doubling; and Indicates the lower limit of the volume interval between two time points, and This indicates the upper limit of the volume range between two time points;

[0222] The above calculations are performed under the premise that the volume is positive and the logarithmic terms are well defined. The output VDT is a point estimate of the volume doubling time, and is expressed as follows: As an interval estimate of the volume doubling time.

[0223] In this specific embodiment, S7 specifically refers to:

[0224] The system uses the volume doubling time interval and combines the occupancy rate quantile dispersion index and the point spread function confidence statistical index to trigger quality control and conservative expansion;

[0225] Time index is The lower and upper limits of the time interval for volume doubling are respectively denoted as... and The interval width is defined as:

[0226] ;

[0227] in The interval width representing the time of volume doubling. Indicates the upper limit of the interval, Indicates the lower limit of the interval;

[0228] Occupancy rate quantile dispersion index is marked as (No. (Statistical analysis of volume-weighted quantile differences within a local region of the nodule at a given time point), and statistical analysis of point spread function confidence, indicated by... (No. (Volume-weighted confidence statistic of the local region of the nodule at a given time point), quality control status is denoted as Indicates the trigger state, (Indicating non-triggered state), preset trigger thresholds are denoted as follows: and (used with) and Compare);

[0229] Quality control triggering decisions use logical OR rules, with the core relationship being:

[0230] if ,otherwise (in (represents the logical "OR" operation).

[0231] when At that time, the system expands outwards across the volume interval between the two time points using a conservative strategy, assuming the expansion coefficient is... and (corresponding to the non-positive offset ratio of the lower limit of the interval and the non-negative amplification ratio of the upper limit of the interval, respectively), in the... The time point expands the volume range to:

[0232] ;

[0233] in and They represent the first Lower and upper limits of the volume range at a given time point and These represent the lower and upper limits of the volume range after conservative expansion, respectively;

[0234] And recalculate the conservative volume doubling time interval within the expanded volume range, written as:

[0235] ;

[0236] in This indicates the data acquisition time interval calculated in step S1. Represents the natural logarithm operator, Represents the natural logarithm of the constant 2. and (These represent the lower and upper limits of the conservative range, respectively), and a manual review prompt is generated to indicate that manual review and confirmation are required.

[0237] when At that time, the system will not implement interval expansion and will not generate a manual review prompt;

[0238] Final output quality control status In the triggered state, the conservative volume ranges of the first and second time points are output simultaneously. and and the conservative volume doubling time range .

[0239] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0240] This invention addresses the technical problem of calibrable mapping of depth model probability output to physical volume. It estimates the spatially variable point spread function distribution under DICOM parameters and edge orientation field conditions, and ensures imaging physical consistency through consistency constraints between the edge spread function and modulation transfer function. Subsequently, a two-stage conditional flow maps the probability to voxel occupancy under monotonically bounded constraints. Residual calibration and quantile output are then performed in the local region of nodules, combined with the point spread function distribution. Finally, sub-voxel integration and volume doubling time calculations are performed under anisotropic voxel conditions. This combination couples segmentation probability, imaging physics, and voxel geometry, significantly reducing volume deviations caused by differences in partial volume and reconstruction kernel, thus forming interpretable volume and volume doubling time intervals.

[0241] In terms of algorithm structure, this case proposes several improvements to address technical issues: a spatially variable point spread function neural field with directional conditionalization is used instead of a fixed kernel model, with the output mean and covariance characterizing uncertainty; spatial and frequency domain consistency is used as a self-supervised signal to improve robustness across devices and acquisition protocols; the probability-to-occupancy calibration is decomposed into global baseline mapping and local residual calibration, and monotonically bounded constraints and quantile heads are introduced to ensure that the occupancy rate does not decrease with increasing probability and directly provides the uncertainty interval; a quality control closed loop is constructed, using quantile dispersion, point spread function confidence, and interval width to trigger conservative expansion and manual review. Through these structural improvements, calibrable mapping and robust volume estimation are better achieved, improving the reliability and clinical usability of volume doubling time.

Claims

1. A deep learning based lung nodule sub-voxel volume VDT method, characterized in that, The method comprises the following steps: S1, based on the chest CT volume data and the corresponding DICOM protocol parameters of at least two time points, generate the nodule probability maps of the first time point and the second time point, detect the edges of the blood vessel wall and the airway wall and calculate the edge direction, generate the edge direction field of the first time point and the second time point, and calculate the acquisition time interval according to the acquisition date and time in the DICOM protocol parameters; S2, under the condition of the edge direction field, implement self-supervised learning in the chest CT volume data and the DICOM protocol parameters, adopt the consistency constraint of the edge extension function in the spatial domain and the modulation transfer function in the frequency domain, directionally conditionally establish the spatially variable point spread function distribution, and generate the spatially variable point spread function distribution map and the point spread function confidence map of the first time point and the second time point; S3, under the condition of the nodule probability map and the spatially variable point spread function distribution map of the first time point and the second time point and in the DICOM protocol parameters, construct the first section of the two-section conditional flow, and map the nodule probability to the occupancy rate voxel by voxel under the monotonic bounded constraint to obtain the global occupancy rate initial value map of the first time point and the second time point; S4, according to the global occupancy rate initial value map, locate the nodule local area in the chest CT volume data and crop the nodule local intensity block, and under the condition of the spatially variable point spread function distribution map and the DICOM protocol parameters, construct the second section of the two-section conditional flow and implement local residual correction, directly generate the quantile result of each voxel occupancy rate under the monotonic bounded constraint, which are the lower quantile occupancy rate map, the median occupancy rate map and the upper quantile occupancy rate map of the first time point and the second time point, respectively; S5, under the anisotropic voxel condition defined by the DICOM protocol parameters, implement sub-voxel integration on the median occupancy rate map to generate the volume point estimate of the first time point and the second time point, and calculate the volume interval of the first time point and the second time point according to the lower quantile occupancy rate map and the upper quantile occupancy rate map, and generate the occupancy rate quantile dispersion index and the point spread function confidence statistical index according to the dispersion degree between the occupancy rate quantile maps and the statistical characteristics of the point spread function confidence map; S6, calculate the VDT point estimate according to the volume point estimate and the acquisition time interval, and calculate the VDT interval according to the volume interval and the acquisition time interval; S7, when the quality control state meets the preset triggering condition, expand the volume interval of the two time points and recalculate the conservative VDT interval according to the conservative strategy, and generate a manual review prompt.

2. The deep learning-based lung nodule sub-voxel volume VDT method of claim 1, wherein, S1 specifically comprises: According to the chest CT volume data and the corresponding DICOM protocol parameters of at least two time points, the first time point nodule probability map and the second time point nodule probability map are generated by using a deep learning integrated segmentation model; According to the chest CT volume data, the edges of the blood vessel wall and the airway wall are detected and the edge direction is calculated, and the first time point edge direction field and the second time point edge direction field are generated; The acquisition time interval is generated according to the acquisition date and time in the DICOM protocol parameters; The first time point nodule probability map, the second time point nodule probability map, the first time point edge direction field, the second time point edge direction field and the acquisition time interval are outputted.

3. The deep learning-based lung nodule sub-voxel volume VDT method of claim 1, wherein, S2 specifically comprises: According to the first time point edge direction field and the second time point edge direction field, self-supervised learning is implemented under the condition of chest CT volume data and DICOM protocol parameters, and the spatial edge expansion function consistency constraint and the frequency domain modulation transfer function consistency constraint are jointly optimized to directionally conditionally establish the distribution expression of the spatial variable point spread function, generate the first time point spatial variable point spread function distribution map and the second time point spatial variable point spread function distribution map, wherein the point spread function distribution of each voxel is parameterized by mean and covariance; At the same time, according to the uncertainty evaluation in the estimation process, the first time point point spread function confidence map and the second time point point spread function confidence map are generated; Output the first time point spatial variable point spread function distribution map, the second time point spatial variable point spread function distribution map, the first time point point spread function confidence map and the second time point point spread function confidence map.

4. The deep learning-based lung nodule sub-voxel volume VDT method of claim 1, wherein, S3 is specifically: According to the first time point nodule probability map and the second time point nodule probability map, and the first time point spatial variable point spread function distribution map and the second time point spatial variable point spread function distribution map, a first segment of a two-segment conditional flow is constructed under the condition of DICOM protocol parameters, and the nodule probability is converted into the occupancy rate voxel by voxel through a global baseline mapping constrained by a monotone bounded constraint, wherein the monotone bounded constraint is used to ensure that the occupancy rate value is between zero and one and is monotonically non-decreasing with the nodule probability, and the first time point global occupancy rate initial value map and the second time point global occupancy rate initial value map are generated; Output the first time point global occupancy rate initial value map and the second time point global occupancy rate initial value map.

5. The deep learning-based lung nodule sub-voxel volume VDT method of claim 1, wherein, S4 is specifically: According to the first time point global occupancy rate initial value map and the second time point global occupancy rate initial value map, the nodule local area is located in the chest CT volume data and the nodule local intensity block is cut; The second segment of the two-segment conditional flow is constructed under the condition of the first time point spatial variable point spread function distribution map and the second time point spatial variable point spread function distribution map and the DICOM protocol parameters, and the local residual error is calibrated, and the quantile result of each voxel occupancy rate is directly generated under the monotone bounded constraint, and the first time point lower quantile occupancy rate map, the first time point median occupancy rate map, the first time point upper quantile occupancy rate map, the second time point lower quantile occupancy rate map, the second time point median occupancy rate map and the second time point upper quantile occupancy rate map are generated; Output the first time point lower quantile occupancy rate map, the first time point median occupancy rate map, the first time point upper quantile occupancy rate map, the second time point lower quantile occupancy rate map, the second time point median occupancy rate map and the second time point upper quantile occupancy rate map.

6. The deep learning-based lung nodule sub-voxel volume VDT method of claim 1, wherein, S5 is specifically: According to the first time point median occupancy rate map and the second time point median occupancy rate map, sub-voxel integration is performed on the anisotropic voxels under the condition of the voxel size and the layer thickness defined by the DICOM protocol parameters, and the first time point volume point estimate and the second time point volume point estimate are generated; According to the first time point lower quantile occupancy rate map and the first time point upper quantile occupancy rate map, the first time point volume interval is calculated under the same integration condition, and according to the second time point lower quantile occupancy rate map and the second time point upper quantile occupancy rate map, the second time point volume interval is calculated under the same integration condition; The occupancy fraction quantile dispersion index is generated according to the dispersion between the occupancy fraction quantile maps, and the point spread function confidence statistical index is generated according to the statistical characteristics of the point spread function confidence map; The first time point volume point estimate, the second time point volume point estimate, the first time point volume interval, the second time point volume interval, the occupancy fraction quantile dispersion index and the point spread function confidence statistical index are outputted.

7. The deep learning-based lung nodule sub-voxel volume VDT method of claim 1, wherein, S6 specifically is: The VDT point estimate is generated according to the first time point volume point estimate and the second time point volume point estimate and the collection time interval, according to the calculation relationship that the VDT point estimate is the natural logarithm of the collection time interval multiplied by 2 divided by the natural logarithm of the ratio of the second time point volume point estimate to the first time point volume point estimate; The VDT interval is generated according to the first time point volume interval and the second time point volume interval and the collection time interval, and the corresponding VDT lower limit and VDT upper limit are calculated by pairing the lower limit and the upper limit of the volume interval, respectively; The VDT point estimate and the VDT interval are outputted.

8. The deep learning-based lung nodule sub-voxel volume VDT method of claim 1, wherein, S7 specifically is: The VDT interval width is calculated according to the VDT point estimate and the VDT interval; The quality control state is set to the trigger state when the occupancy fraction quantile dispersion index is greater than the first threshold value or the point spread function confidence statistical index is lower than the second threshold value or the VDT interval width is greater than the third threshold value, or is set to the non-trigger state, and the quality control state is generated according to the comparison between the occupancy fraction quantile dispersion index and the point spread function confidence statistical index and the preset threshold value, respectively; When the quality control state is the trigger state, the conservative volume interval is generated according to the first time point volume interval and the second time point volume interval according to the conservative strategy, wherein the conservative volume interval includes the first time point conservative volume interval and the second time point conservative volume interval, and the upper and lower limits of the VDT are recalculated according to the first time point conservative volume interval and the second time point conservative volume interval to generate the conservative VDT interval, and the artificial review prompt is generated at the same time; When the quality control state is the non-trigger state, the conservative volume interval, the conservative VDT interval and the artificial review prompt are not generated; The output is the quality control state, and the conservative volume interval, the conservative VDT interval and the artificial review prompt are outputted at the same time in the trigger state.

Images