A multi-center lung CT nodule precision detection system based on federated learning
By decomposing CT images into physically invariant semantic features and device-specific physical codes in multicenter lung CT nodule detection, the problem of feature incomparability caused by device differences is solved, and a more efficient model performance improvement is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HUAIAN HOSPITAL (HUAIAN CANCER HOSPITAL)
- Filing Date
- 2026-03-13
- Publication Date
- 2026-06-05
AI Technical Summary
Existing federated learning schemes cannot effectively address the inherent physical differences between different CT devices in multi-center lung CT nodule detection, leading to incomparable features. Furthermore, pathological features and physical features are coupled, affecting model performance.
The physical parameter estimation unit is used to obtain imaging physical parameters. The physical-semantic decoupling unit decomposes the CT image into physically invariant semantic features and device-specific physical codes. A hierarchical aggregation strategy is used for model training. The central server performs weighted aggregation of semantic encoders and retention of physical adapters.
It achieves semantic feature consistency and physical property preservation across different CT devices, improves the model's generalization ability and detection performance, and surpasses the performance of single-center trained models.
Smart Images
Figure CN122156169A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical image processing technology, and more specifically, to a multi-center lung CT nodule precision detection system based on federated learning. Background Technology
[0002] Lung CT nodule detection is a key technology for early lung cancer screening. With the widespread use of medical imaging equipment and the development of artificial intelligence, deep learning-based lung nodule detection systems have achieved good results in single-center scenarios. However, in clinical practice, the amount of data from a single medical center is limited, and there is case selection bias, making it difficult to train a detection model with strong generalization ability. Multi-center data collaboration has become an inevitable choice to improve model performance.
[0003] Federated learning, as a privacy-preserving distributed machine learning paradigm, allows multiple medical institutions to collaboratively train models without sharing raw data, and has gained widespread attention in the field of medical image analysis in recent years. Existing federated learning schemes typically adopt a "data-inactive, model-dynamic" architecture, where each center trains locally and uploads the model parameters, and a central server aggregates them through weighted averaging.
[0004] The existing technology has the following technical defects, specifically:
[0005] First, the inherent physical differences are mistakenly treated as statistical distribution differences. Existing methods treat the differences in multi-center CT images as a non-independent and identically distributed problem of data distribution, and use feature normalization, domain adaptation, or weighted aggregation strategies for compensation. However, differences in physical parameters such as detector response, reconstruction algorithm transfer function, and noise characteristics of different CT devices make the images incomparable at the underlying physical semantic level. Treating these inherent physical differences as statistical noise cannot fundamentally solve the feature alignment problem.
[0006] Second, pathological features and physical features are coupled during feature extraction. When extracting image features, convolutional neural networks simultaneously encode the pathological attributes of lesions and the physical artifacts of the equipment. During federated aggregation, these two types of features are mixed together for weighted averaging, resulting in the aggregated model losing both pathological details and retaining equipment-related physical noise. Consequently, the model's performance may actually be lower than that of a single-center trained model. Summary of the Invention
[0007] The purpose of this invention is to provide a multi-center lung CT nodule precision detection system based on federated learning to solve the problems mentioned in the background art.
[0008] To achieve the above objectives, the present invention aims to provide a multi-center lung CT nodule precision detection system based on federated learning, comprising: a central server and multiple clients deployed in different medical centers.
[0009] The client includes:
[0010] The physical parameter estimation unit is used to acquire local lung CT images and estimate the imaging physical parameters corresponding to the CT images based on a preset imaging physical prior knowledge base.
[0011] A physical-semantic decoupling unit, connected to the physical parameter estimation unit, is used to decompose the CT image into physically invariant semantic features and device-specific physical codes based on a reversible neural network; the physically invariant semantic features are used to characterize the pathological attributes of nodules, and the device-specific physical codes are used to characterize physical artifacts introduced by the imaging device.
[0012] The nodule detection unit, connected to the physical-semantic decoupling unit, is used to detect lung nodules based on the physically invariant semantic features and output the detection results.
[0013] The central server includes:
[0014] A global semantic anchor library is used to store typical nodal semantic feature vectors collected from various clients and processed to remove privacy, serving as a reference benchmark for cross-center semantic alignment.
[0015] A hierarchical aggregation unit is used to aggregate model parameters uploaded by each client; wherein, the hierarchical aggregation unit is configured to: perform federated weighted aggregation on the semantic encoder parameters used to extract physically invariant semantic features in the physical-semantic decoupling unit of each client; and not perform cross-center aggregation on the physical adapter parameters used to generate device-specific physical codes in the physical-semantic decoupling unit of each client.
[0016] As a further improvement to this technical solution, the physical parameter estimation unit is further configured such that: the imaging physical prior knowledge base contains physical parameters of various mainstream CT devices, and the physical parameters include at least: detector response function, reconstruction algorithm transfer function, noise model and modulation transfer function curve.
[0017] The physical parameter estimation unit takes a single CT image as input through a lightweight convolutional neural network and outputs the corresponding device type, reconstruction algorithm type, and probability distribution of effective radiation dose.
[0018] As a further improvement to this technical solution, the reversible neural network in the physical-semantic decoupling unit is a reversible residual network or a reversible scaling network architecture.
[0019] The training objectives of the physical-semantic decoupling unit include:
[0020] First constraint: Image reconstruction loss for fully reconstructing the original CT image based on the physically invariant semantic features and the device-specific physical coding.
[0021] The second constraint term: constrains the nodule detection loss to contain sufficient nodule classification information in the physically invariant semantic features;
[0022] The third constraint: the distance between the physically invariant semantic features of the same nodule under different imaging physical parameters is less than the cross-center contrast loss of the distance between different nodules.
[0023] As a further improvement to this technical solution, the method for calculating the cross-center comparison loss includes: for a local anchor nodule feature on the client, comparing it with positive and negative sample features sampled from the global semantic anchor library, and calculating it using a triplet loss function to bring the feature representations of similar nodules closer together under different physical conditions.
[0024] As a further improvement to this technical solution, the construction and updating of the global semantic anchor library includes:
[0025] Initialization phase: Each client selects typical samples that cover the nodule performance patterns unique to this center based on local data, extracts their physically invariant semantic features as candidate anchor points, and uploads them to the central server.
[0026] Dynamic update phase: The central server adds, deletes, and adjusts the weights of anchors in the global semantic anchor library based on the nodule detection performance metrics fed back by each client during the federated training process, as well as the newly discovered nodule subtype features.
[0027] As a further improvement to this technical solution, when the hierarchical aggregation unit performs federated weighted aggregation of the semantic encoder parameters of each client, it calculates the aggregation weight based on the amount of data of that client and the physical diversity index of the client's data.
[0028] As a further improvement to this technical solution, the physical diversity index is specifically implemented as follows:
[0029] The client's imaging physical parameter space is divided into multiple dimensions, including at least the device type dimension, reconstruction algorithm type dimension, and effective radiation dose dimension. The empirical probability distributions of each dimension in the client are statistically analyzed, and a weighted calculation is performed to obtain the client's physical diversity index. The calculation formula is as follows:
[0030]
[0031] in, Indicates the client The physical diversity index, Indicates the client The information entropy of the device type, Indicates the client Information entropy of reconstruction algorithm type Indicates the client The information entropy of the effective radiation dose, , , The preset weighting coefficients satisfy... , Indicates the device type dimension. This indicates the dimension of the reconstruction algorithm type. This represents the effective radiation dose dimension, where the effective radiation dose value is discretized into several preset dose intervals.
[0032]
[0033]
[0034]
[0035] in, Indicates device type The frequency of occurrence Indicates the reconstruction algorithm type The frequency of occurrence This indicates that the effective radiation dose falls within the dose range. The frequency.
[0036] As a further improvement to this technical solution, the nodule detection unit further includes a physical enhancement detection submodule, which is configured to: adaptively adjust the anchor box scale and receptive field of the candidate region generation network according to the imaging physical parameters output by the physical parameter estimation unit, and extract the physical invariant features of the candidate region as the physical signature of the nodule. The physical signature includes attenuation coefficient distribution, edge gradient statistics and three-dimensional shape moments.
[0037] As a further improvement to this technical solution, the system performs the following adaptive process when deployed to the target medical center:
[0038] The client of the target medical center identifies the imaging physical parameters of the local CT equipment through its physical parameter estimation unit.
[0039] Download the trained global semantic encoder parameters from the central server.
[0040] Load physical adapter parameters that match the local imaging physical parameters, or quickly fine-tune the physical adapter based on a small amount of local unlabeled data.
[0041] The semantic encoder and physical adapter are combined to form a complete physical-semantic decoupling unit for lung nodule detection in local CT images.
[0042] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0043] First, this invention solves the problem of incomparability of multi-center data features at the fundamental level of physical imaging. It uses a physical parameter estimation unit to inversely identify the imaging physical parameters of an image, and then a physical-semantic decoupling unit to decompose the image into physically invariant semantic features and device-specific physical codes. Semantic features remain consistent across different devices, while physical codes only characterize device-specific effects, fundamentally resolving the problem of incomparability of multi-center data features caused by differences in physical parameters such as detector response, reconstruction algorithms, and noise characteristics.
[0044] Second, this invention pioneers a hierarchical aggregation strategy that decouples physical and semantic features, preventing pathological features from being contaminated by physical noise. The semantic encoder undergoes federated weighted aggregation to absorb pathological knowledge from each center, while the physical adapter is not aggregated across centers to preserve the physical characteristics of each center. Through decoupled aggregation, the global model learns diverse pathological features while maintaining the physical adaptability of each center, resulting in significantly superior model performance compared to single-center trained models. Attached Figure Description
[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0046] Figure 1 This is a schematic diagram of the system structure connection of the present invention. Detailed Implementation
[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0048] Example: Please refer to Figure 1 As shown, a multi-center lung CT nodule precision detection system based on federated learning is provided, including: a central server and multiple clients deployed in different medical centers.
[0049] The client includes:
[0050] The physical parameter estimation unit is used to acquire local lung CT images and estimate the imaging physical parameters corresponding to the CT images based on a preset imaging physical prior knowledge base.
[0051] In one specific embodiment, the physical parameter estimation unit is further configured such that the imaging physics prior knowledge base contains physical parameters of various mainstream CT devices, and the physical parameters include at least: detector response function, reconstruction algorithm transfer function, noise model, and modulation transfer function curve.
[0052] The detector response function describes the mapping relationship between the incident X-ray photon energy and the detector output signal energy, expressed as:
[0053]
[0054] in, Indicates the energy of the incident photon. This indicates that the detector outputs measured energy. This indicates the energy deposited inside the detector. Indicates incident energy Under the condition of deposition energy is The probability density function is obtained from Monte Carlo simulation. Gaussian-stretched kernel function to describe electron noise and charge sharing effects. For energy-related broadening parameters, Reference energy The expansion of the area This is the energy integral domain; the expression considers that the energy response of the detector to monoenergetic incident photons is not ideal. It is not a function, but a broadened distribution that includes features such as escape peaks and Compton plateaus.
[0055] Reconstruction algorithm transfer function
[0056] The modulation effect of the reconstruction algorithm on the spatial frequency of the image in the frequency domain is expressed as:
[0057]
[0058] in, It represents the frequency amplitude in three-dimensional space;
[0059] This is the product of the ramp filter and the window function;
[0060] In order to rebuild the nuclear Related modulation functions, common forms include:
[0061] For standard cores: ;
[0062] For sharp cores: (High-frequency enhancement);
[0063] For smooth kernels: (High-frequency suppression);
[0064] Represents a three-dimensional Fourier transform;
[0065] In practical applications, the transfer function of the reconstruction algorithm can also be obtained through the Fourier transform of the layer sensitivity curve (SSP): ,in The layer sensitivity distribution is shown along the Z-axis.
[0066] noise model
[0067] The noise variance of CT images is expressed as a function of system parameters and the reconstruction kernel:
[0068]
[0069] The expressions for each item are as follows:
[0070] Quantum noise term (photon statistical noise):
[0071]
[0072] Electronic noise (system thermal noise):
[0073]
[0074] Relevant noise introduced by the reconstruction algorithm:
[0075]
[0076] in, The incident photon flux The linear attenuation coefficient along the path, This refers to the signal-to-noise ratio characteristics of the detector. For scan time, For electronic gain factor, For electronic noise power spectral density, To reconstruct the noise power spectrum of the image, The noise power spectrum of the projected data. The pass function is used for the reconstruction algorithm.
[0077] The relationship between the noise power spectrum and the reconstruction kernel can be further expressed as: That is, the frequency domain distribution of noise is directly modulated by the reconstruction kernel.
[0078] Modulation transfer function
[0079] The modulation transfer function describes the system's ability to transfer contrast across different spatial frequencies, and is expressed as the Fourier transform normalized amplitude of the point spread function (PSF):
[0080]
[0081] in It is the line spread function, which is obtained by integrating the point spread function along a certain direction.
[0082] According to the Gaussian approximation model, MTF can be simplified to a closed-form expression:
[0083]
[0084] in: The Gaussian width parameter of the point spread function is related to the focal spot size and detector aperture. For sampling interval, This is due to the cutoff effect caused by sampling. To rebuild the nuclear Contribution to MTF, Nyquist frequency This is the highest spatial frequency that the system can recover.
[0085] The total MTF of the system is the product of the MTFs of each subsystem:
[0086] .
[0087] Spatial frequency corresponding to 50% modulation transfer function Often used as a key indicator of system resolution, meeting .
[0088] In the above expressions, the symbols for key parameters are standardized as follows: Spatial frequency (unit: lp / cm or cycles / mm) : Reconstruct the kernel type identifier, The standard deviation parameter of a Gaussian distribution. : Transfer function (frequency domain) Point spread function (spatial domain). : Modulation function related to the reconstruction kernel, Sampling interval / pixel size Energy (unit: keV).
[0089] The physical parameter estimation unit takes a single CT image as input through a lightweight convolutional neural network and outputs the corresponding device type, reconstruction algorithm type, and probability distribution of effective radiation dose.
[0090] The preferred architecture is EfficientNet-B0, whose composite scaling formula is: , , ,in The network takes one or more CT images as input and outputs three sets of probability distributions:
[0091] The probability distribution of equipment types covers major CT manufacturers and specific models, including at least GE, Siemens, Philips, Canon, and United Imaging.
[0092] The probability distribution of reconstruction algorithm types includes at least: Filtered Back Projection (FBP), Hybrid Iterative Reconstruction (Hybrid-IR), Model-Based Iterative Reconstruction (MBIR), and Deep Learning Reconstruction (DLR).
[0093] The effective radiation dose probability distribution is calculated through the following steps:
[0094] First, estimate the scanning parameters, including tube voltage kVp, tube current mAs, scanning length L, and pitch p;
[0095] Secondly, calculate the volumetric CT dose index. ,in , , These are empirical constants related to the equipment;
[0096] Then, calculate the dose-length product. ;
[0097] Finally, select the corresponding conversion factor based on the device type and reconstruction algorithm type. Statistical distribution of radiation dose, and calculation of effective radiation dose. It outputs the probability density function.
[0098] A physical-semantic decoupling unit, connected to the physical parameter estimation unit, is used to decompose the CT image into physically invariant semantic features and device-specific physical codes based on a reversible neural network; the physically invariant semantic features are used to characterize the pathological attributes of nodules, and the device-specific physical codes are used to characterize physical artifacts introduced by the imaging device.
[0099] In one specific embodiment, the reversible neural network in the physical-semantic decoupling unit is a reversible residual network (i-ResNet) or a reversible scaling network (NICE) architecture.
[0100] The training objectives of the physical-semantic decoupling unit include:
[0101] First constraint: Image reconstruction loss for fully reconstructing the original CT image based on the physically invariant semantic features and the device-specific physical coding.
[0102] Based on the physical invariant semantic features and device-specific physical codes The original CT image was completely reconstructed through the inverse mapping of a reversible neural network. The reconstruction loss function is defined as:
[0103]
[0104] in: To input the original CT image, To reconstruct the image, For the inverse mapping function of an invertible neural network, L1 norm loss is used to maintain the absolute value accuracy of image pixel values. The structural similarity index is defined as:
[0105]
[0106] in , The mean, , Standard deviation, For covariance, , It is the stability constant. To balance the weights, it is usually taken that... ;
[0107] Based on the principle of invertible residual networks (i-ResNet), the forward mapping is as follows: The reverse mapping is achieved through fixed-point iteration.
[0108]
[0109] in For a residual block, the Lipschitz constant must be satisfied. This is to ensure that the iteration converges.
[0110] The second constraint term: constrains the nodule detection loss to contain sufficient nodule classification information in the physically invariant semantic features;
[0111] Constraining the physically invariant semantic features With sufficient nodule classification information, the detection loss is defined as:
[0112]
[0113] Among them: classification loss (Focal Loss is used to address the nodule sample imbalance problem):
[0114]
[0115] For the first The predicted probability of nodules in each candidate region The labels are true (1 represents a nodule, 0 represents a non-nodule). 4 represents the category balancing weight, which is usually taken as...
[0116] 3 is the focusing parameter, which is usually taken as... This is used to reduce the loss contribution of easily classified samples;
[0117] Regression loss (Used for nodule bounding box regression):
[0118]
[0119] in, For the predicted bounding box parameters, For the actual bounding box parameters, The function is defined as:
[0120]
[0121] This detection loss ensures It encodes sufficient nodule semantic information.
[0122] The third constraint: the distance between the physically invariant semantic features of the same nodule under different imaging physical parameters is less than the cross-center contrast loss of the distance between different nodules.
[0123] Implemented using triplet loss:
[0124]
[0125] in: Anchor samples are specific nodules originating from a particular center. Images that are positive samples, originating from the same case / nodule as the anchor point nodule, but from images with different imaging physical parameters (e.g., different devices, different reconstruction algorithms). This is a negative sample, and it is a different nodule from the anchor nodule. The feature vector extracted from physically invariant semantic features, i.e. , The square of the Euclidean distance. This is the margin parameter, which is usually set to... Control the threshold of the distance difference between positive and negative samples. This refers to the set of triples constructed in the training batch.
[0126] To improve training efficiency, a batch hard sample mining strategy can be adopted: in each training batch, for each anchor sample... Select the farthest positive sample (the hardest positive sample) and the closest negative sample (the hardest negative sample) to construct triples.
[0127] All parameter symbols in the above expression are consistent with those in the other claims of this patent:
[0128] Physically invariant semantic features (corresponding to "physically invariant semantic features" in claim 1);
[0129] Device-specific physical code (corresponding to "device-specific physical code" in claim 1);
[0130] Semantic feature extraction function;
[0131] Lipschitz constant of the residual block (corresponding to the invertibility constraint of i-ResNet);
[0132] : Compare the loss margin.
[0133] In one specific embodiment, the method for calculating the cross-center contrast loss includes: for a local anchor nodule feature on the client, comparing it with positive sample features (different physical manifestations of the same nodule) and negative sample features (different nodules) sampled from the global semantic anchor library, and calculating it through a triplet loss function to bring the feature representations of the same type of nodules closer together under different physical conditions.
[0134] The nodule detection unit, connected to the physical-semantic decoupling unit, is used to detect lung nodules based on the physically invariant semantic features and output the detection results.
[0135] The central server includes:
[0136] A global semantic anchor library is used to store typical nodal semantic feature vectors collected from various clients and processed to remove privacy, serving as a reference benchmark for cross-center semantic alignment.
[0137] In one specific embodiment, the construction and updating of the global semantic anchor library includes: initialization phase: each client selects typical samples covering the nodule performance patterns unique to this center based on local data, extracts their physically invariant semantic features as candidate anchors, and uploads them to the central server.
[0138] Dynamic update phase: The central server adds, deletes, and adjusts the weights of anchors in the global semantic anchor library based on the nodule detection performance metrics fed back by each client during the federated training process, as well as the newly discovered nodule subtype features.
[0139] A hierarchical aggregation unit is used to aggregate model parameters uploaded by each client; wherein, the hierarchical aggregation unit is configured to: perform federated weighted aggregation on the semantic encoder parameters used to extract physically invariant semantic features in the physical-semantic decoupling unit of each client; and not perform cross-center aggregation on the physical adapter parameters used to generate device-specific physical codes in the physical-semantic decoupling unit of each client.
[0140] In one specific embodiment, when the hierarchical aggregation unit performs federated weighted aggregation of the semantic encoder parameters of each client, it calculates the aggregation weight based on the amount of data of that client and the physical diversity index of the client's data.
[0141] The physical diversity index is calculated based on the distribution of imaging physical parameters estimated by the client. The higher the physical diversity of the client, the greater the weight of its semantic encoder during aggregation.
[0142] In one specific embodiment, the physical diversity index is implemented by dividing the client's imaging physical parameter space into multiple dimensions, including at least the device type dimension, reconstruction algorithm type dimension, and effective radiation dose dimension; statistically analyzing the empirical probability distribution of each dimension in the client; and weighting the distributions to obtain the client's physical diversity index. The calculation formula is as follows:
[0143]
[0144] in, Indicates the client The physical diversity index, Indicates the client The information entropy of the device type, Indicates the client Information entropy of reconstruction algorithm type Indicates the client The information entropy of the effective radiation dose, , , The preset weighting coefficients satisfy... , Indicates the device type dimension. This indicates the dimension of the reconstruction algorithm type. This represents the effective radiation dose dimension, where the effective radiation dose value is discretized into several preset dose intervals.
[0145]
[0146]
[0147]
[0148] in, Indicates device type The frequency of occurrence Indicates the reconstruction algorithm type The frequency of occurrence This indicates that the effective radiation dose falls within the dose range. The frequency.
[0149] In one specific embodiment, the nodule detection unit further includes a physical enhancement detection submodule, which is configured to: adaptively adjust the anchor box scale and receptive field of the candidate region generation network according to the imaging physical parameters output by the physical parameter estimation unit, and extract the physical invariant features of the candidate region as the physical signature of the nodule, wherein the physical signature includes attenuation coefficient distribution, edge gradient statistics and three-dimensional shape moments.
[0150] In one specific embodiment, when the system is deployed to a target medical center, it performs the following adaptive process:
[0151] The client of the target medical center identifies the imaging physical parameters of the local CT equipment through its physical parameter estimation unit.
[0152] Download the trained global semantic encoder parameters from the central server.
[0153] Load physical adapter parameters that match the local imaging physical parameters, or quickly fine-tune the physical adapter based on a small amount of local unlabeled data.
[0154] The semantic encoder and physical adapter are combined to form a complete physical-semantic decoupling unit for lung nodule detection in local CT images.
[0155] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.
Claims
1. A multi-center lung CT nodule precision detection system based on federated learning, characterized in that, include: A central server and multiple clients deployed in different medical centers; The client includes: The physical parameter estimation unit is used to acquire local lung CT images and estimate the imaging physical parameters corresponding to the CT images based on a preset imaging physical prior knowledge base. A physical-semantic decoupling unit, connected to the physical parameter estimation unit, is used to decompose the CT image into physically invariant semantic features and device-specific physical codes based on a reversible neural network; the physically invariant semantic features are used to characterize the pathological attributes of nodules, and the device-specific physical codes are used to characterize physical artifacts introduced by the imaging device; The nodule detection unit, connected to the physical-semantic decoupling unit, is used to detect lung nodules based on the physically invariant semantic features and output the detection results. The central server includes: A global semantic anchor library is used to store typical nodal semantic feature vectors collected from various clients and processed to remove privacy, serving as a reference benchmark for cross-center semantic alignment. A hierarchical aggregation unit is used to aggregate model parameters uploaded by each client; wherein, the hierarchical aggregation unit is configured to: perform federated weighted aggregation on the semantic encoder parameters used to extract physically invariant semantic features in the physical-semantic decoupling unit of each client; and not perform cross-center aggregation on the physical adapter parameters used to generate device-specific physical codes in the physical-semantic decoupling unit of each client.
2. The multi-center lung CT nodule precision detection system based on federated learning according to claim 1, characterized in that: The physical parameter estimation unit is further configured such that the imaging physical prior knowledge base contains physical parameters of various mainstream CT devices, and the physical parameters include at least: detector response function, reconstruction algorithm transfer function, noise model and modulation transfer function curve; The physical parameter estimation unit takes a single CT image as input through a lightweight convolutional neural network and outputs the corresponding device type, reconstruction algorithm type, and probability distribution of effective radiation dose.
3. The multi-center lung CT nodule precision detection system based on federated learning according to claim 2, characterized in that: The reversible neural network in the physical-semantic decoupling unit is a reversible residual network or a reversible scalable network architecture; The training objectives of the physical-semantic decoupling unit include: First constraint: Image reconstruction loss for fully reconstructing the original CT image based on the physically invariant semantic features and the device-specific physical coding; The second constraint term: constrains the nodule detection loss to contain sufficient nodule classification information in the physically invariant semantic features; The third constraint: the distance between the physically invariant semantic features of the same nodule under different imaging physical parameters is less than the cross-center contrast loss of the distance between different nodules.
4. The multi-center lung CT nodule precision detection system based on federated learning according to claim 3, characterized in that: The method for calculating the cross-center contrast loss includes: For a local anchor node feature on the client, it is compared with positive and negative sample features sampled from the global semantic anchor library, and the loss function of triplet is used to calculate the feature representation of the same type of nodule under different physical conditions.
5. The multi-center lung CT nodule precision detection system based on federated learning according to claim 4, characterized in that: The construction and updating of the global semantic anchor library includes: Initialization phase: Each client selects typical samples that cover the nodule performance patterns unique to this center based on local data, extracts their physically invariant semantic features as candidate anchor points, and uploads them to the central server. Dynamic update phase: The central server adds, deletes, and adjusts the weights of anchors in the global semantic anchor library based on the nodule detection performance metrics fed back by each client during the federated training process, as well as the newly discovered nodule subtype features.
6. The multi-center lung CT nodule precision detection system based on federated learning according to claim 5, characterized in that: When the hierarchical aggregation unit performs federated weighted aggregation of the semantic encoder parameters of each client, it calculates the aggregation weight based on the amount of data of that client and the physical diversity index of the client's data.
7. The multi-center lung CT nodule precision detection system based on federated learning according to claim 6, characterized in that: The physical diversity index is specifically implemented as follows: The client's imaging physical parameter space is divided into multiple dimensions, including at least the device type dimension, reconstruction algorithm type dimension, and effective radiation dose dimension. The empirical probability distributions of each dimension in the client are statistically analyzed, and a weighted calculation is performed to obtain the client's physical diversity index. The calculation formula is as follows: in, Indicates the client The physical diversity index, Indicates the client The information entropy of the device type, Indicates the client Information entropy of reconstruction algorithm type Indicates the client The information entropy of the effective radiation dose, , , The preset weighting coefficients satisfy... , Indicates the device type dimension. This indicates the dimension of the reconstruction algorithm type. This represents the effective radiation dose dimension, where the effective radiation dose value is discretized into several preset dose intervals. in, Indicates device type The frequency of occurrence Indicates the reconstruction algorithm type The frequency of occurrence This indicates that the effective radiation dose falls within the dose range. The frequency.
8. The multi-center lung CT nodule precision detection system based on federated learning according to claim 7, characterized in that: The nodule detection unit further includes a physical enhancement detection submodule, which is configured as follows: Based on the imaging physical parameters output by the physical parameter estimation unit, the anchor box scale and receptive field of the candidate region generation network are adaptively adjusted, and the physical invariant features of the candidate region are extracted as the physical signature of the nodule. The physical signature includes attenuation coefficient distribution, edge gradient statistics, and three-dimensional shape moments.
9. A multi-center lung CT nodule precision detection system based on federated learning according to claim 8, characterized in that: When the system is deployed to the target medical center, it performs the following adaptive process: The client of the target medical center identifies the imaging physical parameters of the local CT equipment through its physical parameter estimation unit; Download the trained global semantic encoder parameters from the central server; Load physical adapter parameters that match the local imaging physical parameters, or quickly fine-tune the physical adapter based on a small amount of local unlabeled data; The semantic encoder and physical adapter are combined to form a complete physical-semantic decoupling unit for lung nodule detection in local CT images.