A dual-branch lung nodule detection system and method combining domain incremental learning and low-dose CT

Abstract

This invention discloses a dual-branch pulmonary nodule detection system and method combining domain incremental learning and low-dose CT. The continuous domain adaptive detection module of the system includes: a historical domain prior detection model, a current domain adaptive detection model, and a multi-scale inter-domain knowledge transfer framework. Both the historical domain prior detection model and the current domain adaptive detection model embed a channel space collaborative cross-domain optimization module. This invention effectively mines the features of small lesions and cross-domain general anatomical knowledge in the high-noise environment of low-dose CT, solving the catastrophic forgetting and domain drift problems faced by traditional deep learning models in multi-center applications. Compared with existing technologies, this invention can construct a pulmonary nodule detection system with continuous evolution capabilities and strong generalization based on incremental low-dose CT data streams without storing historical private data, thereby achieving efficient, robust, and accurate automated screening of data from different medical institutions and different imaging equipment.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing and computer-aided diagnostic technology, and in particular to a dual-branch pulmonary nodule detection system and method that combines domain incremental learning and low-dose CT. Background Technology

[0002] Lung cancer is one of the leading causes of cancer-related deaths worldwide, posing a significant threat to human health. Early detection, diagnosis, and treatment are crucial for improving the survival rate of lung cancer patients. Accurate detection of pulmonary nodules, as an early imaging manifestation of lung cancer, is vital for clinical treatment planning and improving patient prognosis. With increased public health awareness and widespread screening, low-dose computed tomography (LDCT) has become the international standard for early lung cancer screening due to its significantly lower radiation dose compared to conventional CT.

[0003] However, while low-dose CT reduces radiation risks, it also leads to a decrease in image quality. Due to the reduced number of X-ray photons, LDCT images are often accompanied by higher quantum noise and streak artifacts, which severely affects the clarity of displaying small lung nodules. Distinguishing small nodules from similar structures such as vascular sections and bone artifacts in complex anatomical contexts presents a significant challenge. Currently, clinicians not only face a massive workload of dealing with a vast number of computed tomographic images, but also find manual interpretation highly dependent on their experience, making it susceptible to subjective factors and fatigue, leading to missed or misdiagnosed small nodules and consequently affecting the optimal timing of patient treatment.

[0004] In recent years, artificial intelligence technology, especially computer-aided diagnostic (CAD) systems based on deep learning, has demonstrated outstanding performance in lung nodule detection. Deep learning models can automatically extract image features to assist doctors in rapid screening. However, existing mainstream detection models face severe challenges in actual clinical deployment: First, there is the domain drift problem: different hospitals use different brands of CT equipment, scanning parameters, and reconstruction algorithms, resulting in inconsistent data distribution. A model trained on one dataset often experiences a significant performance drop when directly applied to a new clinical scenario. Second, there is the catastrophic forgetting phenomenon: when models are fine-tuned or updated using newly collected data, they often forget the knowledge learned from the old data, leading to a degradation in the ability to detect the original data distribution. If a method of retraining with a mixture of new and old data is used, it is not only computationally expensive but also difficult to centrally store multi-center historical datasets due to medical data privacy protection policies. In addition, given the high noise characteristics of LDCT, existing convolutional neural networks often struggle to balance local details and global contextual information when extracting features, resulting in insufficient accuracy in locating small targets.

[0005] Therefore, how to build a system that can adapt to new data domains, overcome catastrophic forgetting, and maintain high-precision detection in low-dose noise environments is a key technical challenge that urgently needs to be solved in the field of medical image analysis. Summary of the Invention

[0006] To address the aforementioned issues, this invention provides a dual-branch pulmonary nodule detection system and method that combines domain incremental learning with low-dose CT, thereby improving the accuracy and reliability of detection by fully utilizing imaging features.

[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides a dual-branch pulmonary nodule detection system combining domain incremental learning and low-dose CT, the system comprising: The data acquisition module is configured to acquire low-dose lung CT image data from various sources and corresponding lung nodule annotation information, and construct them into a domain incremental data stream according to the chronological order or data source. The continuous domain adaptive detection module is configured to receive domain incremental data streams from the data acquisition module, perform deep feature extraction, cross-domain knowledge transfer, and lung nodule detection on the input data, and output detection prediction results, including the location coordinates and confidence scores of the nodules, for incremental training and real-time prediction of new domain data; the continuous domain adaptive detection module includes: a historical domain prior detection model, a current domain adaptive detection model, and a multi-scale inter-domain knowledge transfer framework; The historical domain prior detection model includes: an encoder with an embedded channel spatial collaborative cross-domain optimization module, a multi-scale feature construction unit, and a domain-specific detection head; the encoder is configured to perform bi-branch feature decoupling on the input low-dose CT 3D features, extracting cross-domain shared anatomical general features and domain-specific statistical features; the multi-scale feature construction unit is configured to extract multi-scale features from different levels of the encoder, and perform spatial alignment and channel fusion on the multi-scale features to construct a unified multi-scale semantic representation; the domain-specific detection head is configured to output detection prediction results based on the multi-scale feature pyramid, generate auxiliary predictions during the training phase, and participate in pseudo-label screening; The current domain adaptive detection model includes a general feature extraction branch and a domain-specific adaptation branch. The general feature extraction branch is configured to extract cross-domain invariant anatomical semantic features and continuously update them during the continuous incremental learning process to enhance the model's generalization ability to lung nodules under different imaging conditions. The domain-specific adaptation branch is configured to independently model the unique imaging distribution features of the current incremental dataset. The domain-specific adaptation branch is extended using a parameter isolation strategy at different incremental stages to capture domain shifts introduced by different scanning devices and imaging protocols. The multi-scale inter-domain knowledge transfer framework is used to coordinate the transfer of historical knowledge and the adaptation of new domain features during incremental learning. First, the framework extracts hierarchical features from the frozen historical domain prior detection model using multi-scale feature construction units, including shallow texture, mid-level shape, and deep semantic information. It then performs scale alignment and depth fusion on features at different levels to generate a multi-scale joint feature representation for knowledge transfer. Subsequently, it uses double confidence constraints to filter the prediction results of the historical model, ensuring that the retained pseudo-labels simultaneously meet the requirements of high classification confidence and high intersection-union ratio of true labels, effectively suppressing cross-domain noise interference. Finally, it optimizes the distillation loss function based on high-confidence pseudo-labels through knowledge distillation and collaborates with the supervision loss of the current incremental task to optimize the current domain adaptive detection model.

[0008] In one embodiment of the present invention, the workflow of the historical domain prior detection model includes: receiving a lung CT image as input data, and inputting the lung CT image to a historical domain prior encoder and a historical domain prior decoder, wherein the historical domain prior encoder has an embedded channel space collaborative cross-domain optimization module for performing feature extraction and cross-domain feature enhancement processing on the input lung CT image; In the historical domain prior encoder, multi-scale features are extracted from the low-level, mid-level, and high-level layers, respectively. The low-level, mid-level, and high-level features are scale-aligned and fused by the multi-scale feature construction unit to obtain a fused feature representation. The fused features are input into a domain-specific detection head to generate a prediction result corresponding to the historical domain prior detection model. Based on the detection prediction result, sample screening is performed to obtain a high-confidence auxiliary prediction set.

[0009] In one embodiment of the present invention, the channel spatial collaborative cross-domain optimization module includes: a channel interaction modeling unit, a spatial adaptive focusing unit, and a 3D convolutional layer; The channel interaction modeling unit comprises, in sequence: a global average pooling layer, a global max pooling layer, and a shared multilayer perceptron; wherein, the multilayer perceptron comprises, in sequence: a first fully connected layer, a ReLU activation function, a second fully connected layer, and a Sigmoid activation function. For a given input feature tensor F, the formula for calculating the channel attention weights is as follows:

[0010] in, It is the Sigmoid activation function. It is the ReLU activation function. and The weight parameters are learnable; the channel enhancement features are represented as:

[0011] in, This represents element-wise multiplication. The spatial adaptive focusing unit includes: a query mapping convolutional layer, a key mapping convolutional layer, a value mapping convolutional layer, an attention weight calculation unit, and a residual connection unit; the query mapping convolutional layer, key mapping convolutional layer, and value mapping convolutional layer are used to process the channel-enhanced features respectively. Perform convolution mapping to obtain the query mapping. Key mapping Sum value mapping The calculation formula is as follows:

[0012] in, , and This represents the convolutional mapping operation; using the attention weight calculation unit, based on the query mapping... and key mapping Calculate spatial attention weights The calculation formula is as follows:

[0013] Using the residual connection unit, the spatial attention weights are... AND-value mapping Perform aggregation and introduce a learnable scaling factor. Input features Perform residual connections to obtain spatial focusing features. The calculation formula is as follows:

[0014] in, These are learnable parameters used to adjust the weights of global information.

[0015] In one embodiment of the present invention, the domain-specific detection head includes a convolutional layer, a Dropout random deactivation layer, and a convolutional layer connected in sequence, and its expression is:

[0016] in, This represents the joint feature representation of the input. Indicates the kernel size as The three-dimensional feature extraction convolution operation is represented by Dropout, which is a regularization operation with a random deactivation rate of p. Indicates the kernel size as Dimensional projection convolution operation, This is the generated auxiliary prediction result.

[0017] In one embodiment of the present invention, the workflow of the current domain adaptive detection model includes: when the incremental time step t is reached, inputting the lung CT image data corresponding to the current time step into the encoder backbone network, wherein the encoder backbone network includes a general shared feature branch F and multiple domain-specific adaptation branches. Freeze the domain-specific adaptation branches corresponding to historical time steps 1 to t-1. The network parameters only retain the domain adaptation branch corresponding to the current time step. It is in a trainable state; the general shared feature branch F is continuously updated to extract stable anatomical semantic features with cross-domain generalization ability from lung CT images at different time steps and in different domains.

[0018] In one embodiment of the present invention, the workflow of the multi-scale inter-domain knowledge transfer framework is as follows: First, at the current time step t, the previous time step t-1 is trained; the encoder of the historical domain prior detection model extracts multi-scale features at different levels, and the multi-scale feature construction unit performs unified scale alignment and fusion on the multi-scale features to form a joint feature representation for knowledge transfer. Then, the joint feature representation is input into a domain-specific detection head to generate a set of auxiliary prediction results; based on the auxiliary prediction results, a dual-reset confidence constraint strategy is used to filter the predictions of the historical domain prior detection model to obtain a high-quality pseudo-label set, specifically including: For each predicted bounding box, both classification confidence and ground truth consistency constraints are applied. A predicted bounding box is considered a valid pseudo-label when its classification confidence exceeds a dynamic confidence threshold and the intersection-union (IU) ratio between the predicted bounding box and the ground truth label in the current domain exceeds a preset IU threshold. The resulting set of high-quality pseudo-labels is represented as follows:

[0019] in, This represents the i-th prediction result generated by the teacher model. This represents the corresponding classification confidence level. For dynamic confidence thresholds, For the intersection-union ratio threshold, Indicates the actual annotation of the current domain; Finally, based on a high-quality pseudo-label set, knowledge distillation optimization is performed on the current domain adaptive detection model. The current domain real label supervision loss and incremental distillation loss are jointly constrained, learning the current domain feature distribution while maintaining the ability to discriminate key anatomical structures in the historical domain. The overall optimization objective function is defined as:

[0020] Among them, The detection loss for the current domain's true label includes Focal Loss for classification prediction and GIoU Loss for bounding box regression; To address the knowledge distillation loss for high-quality pseudo-labels, this is used to constrain the consistency between the student model and the teacher model in terms of classification probability distribution and regression results; This is an importance coefficient used to balance the learning of new knowledge with the retention of old knowledge.

[0021] In one embodiment of the present invention, the multi-scale feature construction unit includes: a 1×1×1 convolutional layer, a 3×3×3 convolutional layer, an upsampling layer, a downsampling layer, a channel splicing layer, and a feature aggregation convolutional layer; The multi-scale feature construction unit extracts shallow high-resolution features from the frozen historical prior detection model encoder. Mid-layer characteristics and deep low-resolution features ; for shallow high-resolution features Perform downsampling operation on deep, low-resolution features After performing upsampling to unify the spatial resolution of features across all levels, use... Convolutional dimensionality reduction and concatenation along the channel dimension, finally passed through Convolution completes feature aggregation; the expression for the multi-scale feature building unit is:

[0022] in, This indicates a downsampling operation. Indicates an upsampling operation; This represents a 1×1×1 convolution operation used for channel dimensionality reduction. This represents a 3×3×3 convolution operation used for feature aggregation; Indicates a channel splicing operation; This represents the joint feature representation after fusion.

[0023] Secondly, this invention provides a method for detecting bibranched pulmonary nodules using a combination of domain incremental learning and low-dose CT, employing the aforementioned system. The method includes: Step 1: Acquire low-dose lung CT image data and corresponding lung nodule annotation information, and construct a domain incremental data stream in chronological order; Step 2: Use the dual-branch lung nodule detection system to detect lung nodules in the data stream and output the corresponding lung nodule detection results.

[0024] Thirdly, the present invention provides a dual-branch pulmonary nodule detection device combining domain incremental learning and low-dose CT, including a memory and a processor; The memory is used to store computer programs; The processor is configured to implement the method when the computer program is executed.

[0025] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described thereon.

[0026] The beneficial effects of this invention are: The dual-branch lung nodule detection system based on combined domain incremental learning and low-dose CT of this invention can efficiently adapt to the differences in data distribution across different medical institutions and time periods, achieving accurate lung nodule detection even in low-dose, high-noise environments. By employing a continuous domain adaptive detection architecture, a channel-space collaborative cross-domain optimization module, and a domain-specific detection head based on dual-reset confidence constraints, this invention can mine the features of minute lesions and cross-domain general knowledge in low-dose CT images, solving the catastrophic forgetting and domain drift problems faced by traditional deep learning models in multi-center applications. Specifically, this invention designs a channel-space collaborative cross-domain optimization module, adopting a cascaded architecture of channel interaction modeling and spatial adaptive focusing. This module not only optimizes feature information in the channel dimension through adaptive weight allocation, effectively filtering noise and redundant channel information in low-dose CT, but also introduces a self-attention mechanism in the spatial dimension to establish long-distance global dependencies within and between slices. Meanwhile, the multi-scale inter-domain knowledge transfer framework, combined with a pseudo-label screening strategy based on double-reset confidence constraints, effectively integrates the prior knowledge of historical models with the new features of the current domain. Without storing old domain data to protect patient privacy, it improves the model's adaptability to new domain data and the reliability of prediction results.

[0027] Compared to existing technologies, this invention enables efficient and continuous model iteration and optimization without accessing historical private data, providing clinicians with an auxiliary diagnostic tool capable of continuous evolution. The system possesses strong generalization ability and robustness, making it particularly suitable for large-scale early lung cancer screening in multi-center, multi-device environments. It significantly reduces the rates of missed diagnoses and misdiagnoses, driving the intelligent transformation of medical image AI analysis from static, closed systems to dynamic, open systems. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0029] Figure 1 This is a schematic diagram of the operation process of the dual-branch pulmonary nodule detection system based on combined domain incremental learning and low-dose CT provided by the present invention.

[0030] Figure 2 This is a schematic diagram of the structure of the dual-branch pulmonary nodule detection system based on combined domain incremental learning and low-dose CT provided by the present invention.

[0031] Figure 3 This is a schematic diagram of the historical domain prior detection model structure provided by the present invention.

[0032] Figure 4 This is a schematic diagram of the current domain adaptive detection model structure provided by the present invention.

[0033] Figure 5 This is a schematic diagram of the channel space collaborative cross-domain optimization module structure provided by the present invention.

[0034] Figure 6 This is the original image of a low-dose CT lung nodule in Embodiment 3 of the present invention.

[0035] Figure 7 This is a graph showing the prediction results of incremental lung nodules in the dual-branch domain in Embodiment 3 of the present invention. Detailed Implementation

[0036] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are merely illustrative of the technical solution of the present invention and are therefore intended to limit the scope of protection of the present invention.

[0037] Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention; the terms “comprising” and “having”, and any variations thereof, in the specification, claims and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0038] In the description of the embodiments of this invention, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this invention, "multiple" means two or more, unless otherwise explicitly defined.

[0039] In this invention, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least some embodiments of the invention. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this invention can be combined with other embodiments.

[0040] In the description of the embodiments of this invention, the term "and / or" is merely a description of the relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, in this invention, the character " / " generally indicates that the preceding and following associated objects have an "or" relationship.

[0041] Example 1 like Figure 1 As shown, this embodiment provides a dual-branch pulmonary nodule detection system combining domain incremental learning and low-dose CT, including: The data acquisition module is configured to acquire low-dose lung CT image data from diverse sources and corresponding lung nodule annotation information, and construct them into a domain incremental data stream according to chronological order or data source; low-dose lung CT image data can reflect the texture of lung parenchyma, the direction of bronchovascular bundles and the morphological characteristics of potential nodules; lung nodule annotation information includes the spatial coordinates and category labels of nodules, wherein each incremental dataset contains three-dimensional CT image data and lung nodule annotation information; The continuous domain adaptive detection module is configured to receive domain incremental data streams from the data acquisition module, perform deep feature extraction, cross-domain knowledge transfer, and lung nodule detection on the input data, and output detection prediction results, including the location coordinates and confidence scores of the nodules, for incremental training and real-time prediction of new domain data; the continuous domain adaptive detection module includes: a historical domain prior detection model, a current domain adaptive detection model, and a multi-scale inter-domain knowledge transfer framework; The historical domain prior detection model includes: an encoder with an embedded channel spatial collaborative cross-domain optimization module, a multi-scale feature construction unit, and a domain-specific detection head; the encoder is configured to perform bi-branch feature decoupling on the input low-dose CT 3D features, extracting cross-domain shared anatomical general features and domain-specific statistical features; the multi-scale feature construction unit is configured to extract multi-scale features from different levels of the encoder, and perform spatial alignment and channel fusion on the multi-scale features to construct a unified multi-scale semantic representation; the domain-specific detection head is configured to output detection prediction results based on the multi-scale feature pyramid, generate auxiliary predictions during the training phase, and participate in pseudo-label screening; The current domain adaptive detection model includes a general feature extraction branch and a domain-specific adaptation branch. The general feature extraction branch is configured to extract cross-domain invariant anatomical semantic features and continuously update them during the continuous incremental learning process to enhance the model's generalization ability to lung nodules under different imaging conditions. The domain-specific adaptation branch is configured to independently model the unique imaging distribution features of the current incremental dataset. The domain-specific adaptation branch is extended using a parameter isolation strategy at different incremental stages to capture domain shifts introduced by different scanning devices and imaging protocols. The multi-scale inter-domain knowledge transfer framework is used to coordinate the transfer of historical knowledge and the adaptation of new domain features during incremental learning. First, the framework extracts hierarchical features from the frozen historical domain prior detection model through multi-scale feature construction units, including shallow texture, mid-level shape, and deep semantic information. It then performs scale alignment and depth fusion on features at different levels to generate a multi-scale joint feature representation for knowledge transfer. Subsequently, it uses double confidence constraints to filter the prediction results of the historical model, ensuring that the retained pseudo-labels simultaneously meet the requirements of high classification confidence and high intersection-over-union (IoU) of the true labels, effectively suppressing cross-domain noise interference. Finally, it optimizes the distillation loss function based on high-confidence pseudo-labels through knowledge distillation and co-optimizes the current domain adaptive detection model with the supervision loss of the current incremental task, thereby achieving the ability to detect key anatomical structures of the old domain while learning new domain features.

[0042] Example 2 This embodiment provides a dual-branch pulmonary nodule detection system combining domain incremental learning and low-dose CT, including: a data acquisition module and a continuous domain adaptive detection module.

[0043] The data acquisition module acquires low-dose lung CT image data and corresponding lung nodule annotation information, constructing a domain incremental data stream in chronological order. The continuous domain adaptive detection module includes: a historical domain prior detection model, a current domain adaptive detection model, and a multi-scale inter-domain knowledge transfer framework; for details, refer to... Figure 2The system takes low-dose CT image data and lung nodule annotation information as input. The data first passes through a historical domain prior detection model, where multi-scale feature construction units extract and fuse features to obtain a multi-scale joint feature representation. Then, an auxiliary prediction set is generated through a domain-specific detection head. After strict screening by a dual-reset confidence constraint strategy, high-quality pseudo-labels are obtained. At the same time, the current domain adaptive detection model processes the input data in parallel and outputs detection results. These results are jointly calculated with the output of the historical domain prior detection model to calculate the distillation loss, and combined with the high-quality pseudo-labels to calculate the task loss. Through joint optimization, accurate lung nodule detection results are finally obtained.

[0044] The history domain prior detection model is used to provide cross-domain shared general anatomical features and historical knowledge. The structure of the history domain prior detection model is as follows: Figure 3 As shown, it includes: an encoder with an embedded channel space collaborative cross-domain optimization module, a multi-scale feature construction unit, and a domain-specific detection head.

[0045] The workflow of the historical domain prior detection model includes: receiving lung CT images as input data and inputting the lung CT images to a historical domain prior encoder and a historical domain prior decoder, wherein the historical domain prior encoder has an embedded channel space collaborative cross-domain optimization module for performing feature extraction and cross-domain feature enhancement processing on the input lung CT images; In the historical domain prior encoder, multi-scale features are extracted from the low-level, mid-level, and high-level layers, respectively. The low-level, mid-level, and high-level features are scale-aligned and fused by the multi-scale feature construction unit to obtain a fused feature representation. The fused features are input into a domain-specific detection head to generate a prediction result corresponding to the historical domain prior detection model. Based on the detection prediction result, sample screening is performed to obtain a high-confidence auxiliary prediction set.

[0046] The multi-scale feature construction unit includes: a 1×1×1 convolutional layer, a 3×3×3 convolutional layer, an upsampling layer, a downsampling layer, a channel concatenation layer, and a feature aggregation convolutional layer. The multi-scale feature construction unit extracts shallow high-resolution features from the frozen historical prior detection model encoder. Mid-layer characteristics and deep low-resolution features ; for shallow high-resolution features Perform downsampling operation on deep, low-resolution features After performing upsampling to unify the spatial resolution of features across all levels, use... Convolutional dimensionality reduction and concatenation along the channel dimension, finally passed through Convolution completes feature aggregation. The expression for the multi-scale feature building unit is:

[0047] in, This indicates a downsampling operation. Indicates an upsampling operation; This represents a 1×1×1 convolution operation used for channel dimensionality reduction. This represents a 3×3×3 convolution operation used for feature aggregation; Indicates a channel splicing operation; This represents the joint feature representation after fusion.

[0048] The domain-specific detection head generates auxiliary predictions based on multi-scale joint features, including sequentially connected convolutional layers, Dropout random deactivation layers, and convolutional layers, expressed as follows:

[0049] in, This represents the joint feature representation of the input. Indicates the kernel size as The three-dimensional feature extraction convolution operation is represented by Dropout, which is a regularization operation with a random deactivation rate of p. Indicates the kernel size as Dimensional projection convolution operation, This is the generated auxiliary prediction result.

[0050] The current-domain adaptive detection model is used for adaptive learning in response to the current incremental task. The structure of the current-domain adaptive detection model is as follows: Figure 4 As shown, the model includes a general feature extraction branch and a domain-specific adaptation branch. The general feature extraction branch extracts cross-domain invariant anatomical semantic features and continuously updates them during incremental learning. The domain-specific adaptation branch independently models the unique imaging distribution features of the current incremental dataset, and this branch is expanded using a parameter isolation strategy at different incremental stages. The workflow of the current domain adaptive detection model includes: when the incremental time step t is reached, the lung CT image data corresponding to the current time step is input into the encoder backbone network, wherein the encoder backbone network includes a general shared feature branch F and multiple domain-specific adaptation branches. Freeze the domain-specific adaptation branches corresponding to historical time steps 1 to t-1. The network parameters only retain the domain adaptation branch corresponding to the current time step. It is in a trainable state; the general shared feature branch F is continuously updated to extract stable anatomical semantic features with cross-domain generalization ability from lung CT images at different time steps and in different domains.

[0051] To address low-dose CT noise interference and enhance feature representation, a channel-space collaborative cross-domain optimization module is embedded in the encoder. The structure of the channel-space collaborative cross-domain optimization module is as follows: Figure 5 As shown, it includes: a channel interaction modeling unit, a spatial adaptive focusing unit, and a 3D convolutional layer. The 3D convolutional layer includes: a 3×1×1 convolution, a 1×3×1 convolution, and a 1×1×3 convolution connected in sequence.

[0052] The channel interaction modeling unit captures the dependencies between channels and includes a globally average pooling layer, a globally max pooling layer, and a shared multilayer perceptron connected in sequence. The multilayer perceptron comprises a first fully connected layer, a ReLU activation function, a second fully connected layer, and a Sigmoid activation function connected in sequence. For a given input feature tensor F, the channel attention weights are calculated using the following formula:

[0053] in, It is the Sigmoid activation function. It is the ReLU activation function. and These are learnable weight parameters. Channel enhancement features are represented as:

[0054] in, This indicates element-wise multiplication.

[0055] The spatial adaptive focusing unit is used to establish long-range spatial dependencies. It includes a query mapping convolutional layer, a key mapping convolutional layer, a value mapping convolutional layer, an attention weight calculation unit, and a residual connection unit. The query mapping convolutional layer, key mapping convolutional layer, and value mapping convolutional layer are used to enhance the channel-enhanced features, respectively. Perform convolution mapping to obtain the query mapping. Key mapping Sum value mapping The calculation formula is as follows:

[0056] in, , and This represents the convolutional mapping operation. Using the attention weight calculation unit, based on the query mapping... and key mapping Calculate spatial attention weights The calculation formula is as follows:

[0057] Using the residual connection unit, the spatial attention weights are... AND-value mapping Perform aggregation and introduce a learnable scaling factor. Input features Perform residual connections to obtain spatial focusing features. The calculation formula is as follows:

[0058] in, These are learnable parameters used to adjust the weights of global information.

[0059] Furthermore, a multi-scale inter-domain knowledge transfer framework (integrated in...) Figure 2 In the process shown, a dual-reset confidence constraint strategy is used to filter high-quality pseudo-labels to guide the training of the current domain model.

[0060] Workflow of the multi-scale inter-domain knowledge transfer framework: First, at the current time step t, the previous time step t-1 is trained; the encoder of the historical domain prior detection model extracts multi-scale features at different levels, and the multi-scale feature construction unit performs unified scale alignment and fusion on the multi-scale features to form a joint feature representation for knowledge transfer. Then, the joint feature representation is input into a domain-specific detection head to generate a set of auxiliary prediction results; based on the auxiliary prediction results, a dual-reset confidence constraint strategy is used to filter the predictions of the historical domain prior detection model to obtain a high-quality pseudo-label set, specifically including: For each predicted bounding box, both classification confidence and ground truth consistency constraints are applied. A predicted bounding box is considered a valid pseudo-label when its classification confidence exceeds a dynamic confidence threshold and the intersection-union (IU) ratio between the predicted bounding box and the ground truth label in the current domain exceeds a preset IU threshold. The resulting set of high-quality pseudo-labels is represented as follows:

[0061] in, This represents the i-th prediction result generated by the teacher model. This represents the corresponding classification confidence level. For dynamic confidence thresholds, For the intersection-union ratio threshold, Indicates the actual annotation of the current domain; Finally, based on a high-quality pseudo-label set, knowledge distillation optimization is performed on the current domain adaptive detection model. The current domain real label supervision loss and incremental distillation loss are jointly constrained, learning the current domain feature distribution while maintaining the ability to discriminate key anatomical structures in the historical domain. The overall optimization objective function is defined as:

[0062] Among them, The detection loss for the current domain's true label includes Focal Loss for classification prediction and GIoU Loss for bounding box regression; To address the knowledge distillation loss for high-quality pseudo-labels, this is used to constrain the consistency between the student model and the teacher model in terms of classification probability distribution and regression results; This is an importance coefficient used to balance the learning of new knowledge with the retention of old knowledge.

[0063] Once the model is trained, it can be used for domain increment detection of lung nodules. The specific operation process of this system is as follows: First, in the clinical auxiliary diagnosis stage, doctors acquire low-dose lung CT image data and clinical information of patients to be screened, and input them into a trained dual-branch lung nodule detection system combining joint domain incremental learning and low-dose CT. The system retrieves the current continuous domain adaptive detection model, uses the channel space collaborative cross-domain optimization module to enhance and decouple image features, automatically locates and identifies small lesions in the images, and outputs detection results including the three-dimensional spatial coordinates and classification confidence of the nodules. Doctors then formulate subsequent puncture biopsy or follow-up treatment plans based on these prediction results. Subsequently, when the system detects new domain low-dose CT image data from different sources or with updated imaging protocols, it automatically initiates a domain incremental continuous evolution process in the background. The system first inputs the new domain data into the parameter-frozen historical domain prior detection model, using its domain-specific detection head to generate a set of auxiliary prediction results. Then, the system runs a multi-scale inter-domain knowledge transfer framework, applies a dual-confidence constraint strategy to verify each of these auxiliary predictions, and selects prediction boxes that simultaneously meet the requirements of high classification confidence and high realism as high-quality pseudo-labels. Finally, the system uses these high-quality pseudo-labels as supervision signals, calculates the distillation loss between them and the output of the current domain adaptive detection model, and iteratively updates the weight parameters of the current domain adaptive detection model only through backpropagation algorithm, thereby completing the adaptive upgrade of the model to the new imaging style without interrupting clinical services.

[0064] Example 3 like Figure 6 and Figure 7 As shown, in order to verify and illustrate the technical effects of the present invention, this embodiment selects the traditional mainstream deep learning lung nodule detection scheme and the dual-branch lung nodule detection system of the present invention for comparison, so as to verify the real effect of the present invention in a low-dose, multi-center environment.

[0065] This embodiment will perform detection experiments on low-dose CT (LDCT) images from multiple centers. Considering that the core advantage of this invention lies in cross-domain incremental learning, this embodiment establishes a hybrid database containing three different medical institutions, collecting low-dose lung CT images from 1250 patients. The experiment is set up in a time-sequential incremental learning mode, where 875 cases (approximately 70%) are used as incremental training samples in batches according to data source, and the remaining 375 cases (approximately 30%) are used as test samples to evaluate the model's overall performance after encountering the new data domain.

[0066] The dual-branch pulmonary nodule detection system of the present invention, which combines domain incremental learning with low-dose CT, is compared with traditional detection methods and manual diagnosis results by senior radiologists. The comparison results are shown in Table 1 below: Table 1: Comparison Results of the Invention and Traditional Solutions

[0067] The comparative data above demonstrates that the lung nodule detection system of this invention outperforms traditional technologies and expert diagnoses in terms of detection sensitivity and false positive control, and also boasts faster inference time. Crucially, traditional deep learning methods are difficult to directly implement across devices and centers on a large scale; however, the assisted diagnostic system of this invention, with its joint domain incremental learning architecture, can be extended to large-scale, multi-center lung cancer screening scenarios, ultimately yielding superior assisted diagnostic results.

[0068] In summary, this invention effectively mines the characteristics of minute lesions in the high-noise environment of low-dose CT and cross-domain general anatomical knowledge, solving the catastrophic forgetting and domain drift problems faced by traditional deep learning models in multi-center applications. Compared with existing technologies, this invention can build a lung nodule detection system with continuous evolution capability and strong generalization based on incremental low-dose CT data streams without storing historical private data, thereby achieving efficient, robust and accurate automated screening of data from different medical institutions and different imaging equipment.

[0069] Some steps in the embodiments of the present invention can be implemented using software, and the corresponding software program can be stored in a readable storage medium, such as an optical disc or a hard disk.

[0070] Although the invention has been described with reference to preferred embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, the technical features mentioned in the various embodiments can be combined in any manner as long as there is no structural conflict. The invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A dual-branch pulmonary nodule detection system combining domain incremental learning and low-dose CT, characterized in that, The system includes: The data acquisition module is configured to acquire low-dose lung CT image data from various sources and corresponding lung nodule annotation information, and construct them into a domain incremental data stream according to the chronological order or data source. The continuous domain adaptive detection module is configured to receive domain incremental data streams from the data acquisition module, perform deep feature extraction, cross-domain knowledge transfer, and lung nodule detection on the input data, and output detection prediction results, including the location coordinates and confidence scores of the nodules, for incremental training and real-time prediction of new domain data; the continuous domain adaptive detection module includes: a historical domain prior detection model, a current domain adaptive detection model, and a multi-scale inter-domain knowledge transfer framework; The historical domain prior detection model includes: an encoder with an embedded channel spatial collaborative cross-domain optimization module, a multi-scale feature construction unit, and a domain-specific detection head; the encoder is configured to perform bi-branch feature decoupling on the input low-dose CT 3D features, extracting cross-domain shared anatomical general features and domain-specific statistical features; the multi-scale feature construction unit is configured to extract multi-scale features from different levels of the encoder, and perform spatial alignment and channel fusion on the multi-scale features to construct a unified multi-scale semantic representation; the domain-specific detection head is configured to output detection prediction results based on the multi-scale feature pyramid, generate auxiliary predictions during the training phase, and participate in pseudo-label screening; The current domain adaptive detection model includes a general feature extraction branch and a domain-specific adaptation branch. The general feature extraction branch is configured to extract cross-domain invariant anatomical semantic features and continuously update them during the continuous incremental learning process to enhance the model's generalization ability to lung nodules under different imaging conditions. The domain-specific adaptation branch is configured to independently model the unique imaging distribution features of the current incremental dataset. The domain-specific adaptation branch is extended using a parameter isolation strategy at different incremental stages to capture domain shifts introduced by different scanning devices and imaging protocols. The multi-scale inter-domain knowledge transfer framework is used to coordinate the transfer of historical knowledge and the adaptation of new domain features during incremental learning. First, the framework extracts hierarchical features from the frozen historical domain prior detection model using multi-scale feature construction units, including shallow texture, mid-level shape, and deep semantic information. It then performs scale alignment and depth fusion on features at different levels to generate a multi-scale joint feature representation for knowledge transfer. Subsequently, it uses double confidence constraints to filter the prediction results of the historical model, ensuring that the retained pseudo-labels simultaneously meet the requirements of high classification confidence and high intersection-union ratio of true labels, effectively suppressing cross-domain noise interference. Finally, it optimizes the distillation loss function based on high-confidence pseudo-labels through knowledge distillation and collaborates with the supervision loss of the current incremental task to optimize the current domain adaptive detection model.

2. The system according to claim 1, characterized in that, The workflow of the historical domain prior detection model includes: receiving lung CT images as input data, and inputting the lung CT images to the historical domain prior encoder and the historical domain prior decoder, wherein the historical domain prior encoder has an embedded channel space collaborative cross-domain optimization module for performing feature extraction and cross-domain feature enhancement processing on the input lung CT images; In the historical domain prior encoder, multi-scale features are extracted from the low-level, mid-level, and high-level layers, respectively. The low-level, mid-level, and high-level features are scale-aligned and fused by the multi-scale feature construction unit to obtain a fused feature representation. The fused features are input into a domain-specific detection head to generate a prediction result corresponding to the historical domain prior detection model. Based on the detection prediction result, sample screening is performed to obtain a high-confidence auxiliary prediction set.

3. The system according to claim 2, characterized in that, The channel spatial collaborative cross-domain optimization module includes: a channel interaction modeling unit, a spatial adaptive focusing unit, and a 3D convolutional layer; The channel interaction modeling unit comprises, in sequence: a global average pooling layer, a global max pooling layer, and a shared multilayer perceptron; wherein, the multilayer perceptron comprises, in sequence: a first fully connected layer, a ReLU activation function, a second fully connected layer, and a Sigmoid activation function. For a given input feature tensor F, the formula for calculating the channel attention weights is as follows: in, It is the Sigmoid activation function. It is the ReLU activation function. and The weight parameters are learnable; the channel enhancement features are represented as: in, This indicates element-wise multiplication. The spatial adaptive focusing unit includes: a query mapping convolutional layer, a key mapping convolutional layer, a value mapping convolutional layer, an attention weight calculation unit, and a residual connection unit; the query mapping convolutional layer, key mapping convolutional layer, and value mapping convolutional layer are used to process the channel-enhanced features respectively. Perform convolution mapping to obtain the query mapping. Key mapping Sum value mapping The calculation formula is as follows: in, , and This represents the convolutional mapping operation; using the attention weight calculation unit, based on the query mapping... and key mapping Calculate spatial attention weights The calculation formula is as follows: Using the residual connection unit, the spatial attention weights are... AND-value mapping Perform aggregation and introduce a learnable scaling factor. Input features Perform residual connections to obtain spatial focusing features. The calculation formula is as follows: in, These are learnable parameters used to adjust the weights of global information.

4. The system according to claim 3, characterized in that, The domain-specific detection head comprises a convolutional layer, a Dropout random deactivation layer, and a convolutional layer connected in sequence, and its expression is: in, This represents the joint feature representation of the input. Indicates the kernel size as The three-dimensional feature extraction convolution operation is represented by Dropout, which is a regularization operation with a random deactivation rate of p. Indicates the kernel size as Dimensional projection convolution operation, This is the generated auxiliary prediction result.

5. The system according to claim 4, characterized in that, The workflow of the current domain adaptive detection model includes: when the incremental time step t is reached, the lung CT image data corresponding to the current time step is input into the encoder backbone network, wherein the encoder backbone network includes a general shared feature branch F and multiple domain-specific adaptation branches. Freeze the domain-specific adaptation branches corresponding to historical time steps 1 to t-1. The network parameters only retain the domain adaptation branch corresponding to the current time step. It is in a trainable state; the general shared feature branch F is continuously updated to extract stable anatomical semantic features with cross-domain generalization ability from lung CT images at different time steps and in different domains.

6. The system according to claim 5, characterized in that, The workflow of the multi-scale inter-domain knowledge transfer framework: First, at the current time step t, the previous time step t-1 is trained; the encoder of the historical domain prior detection model extracts multi-scale features at different levels, and the multi-scale feature construction unit performs unified scale alignment and fusion on the multi-scale features to form a joint feature representation for knowledge transfer. Then, the joint feature representation is input into a domain-specific detection head to generate a set of auxiliary prediction results; Based on the aforementioned auxiliary prediction results, a dual-reset confidence constraint strategy is used to filter the predictions of the prior detection model in the historical domain, resulting in a high-quality pseudo-label set, specifically including: For each predicted bounding box, both classification confidence and ground truth consistency constraints are applied. A predicted bounding box is considered a valid pseudo-label when its classification confidence exceeds a dynamic confidence threshold and the intersection-union (IU) ratio between the predicted bounding box and the ground truth label in the current domain exceeds a preset IU threshold. The resulting set of high-quality pseudo-labels is represented as follows: in, This represents the i-th prediction result generated by the teacher model. This represents the corresponding classification confidence level. For dynamic confidence thresholds, For the intersection-union ratio threshold, This represents the actual annotation of the current domain; Finally, based on a high-quality pseudo-label set, knowledge distillation optimization is performed on the current domain adaptive detection model. The current domain real label supervision loss and incremental distillation loss are jointly constrained, learning the current domain feature distribution while maintaining the ability to discriminate key anatomical structures in the historical domain. The overall optimization objective function is defined as: Among them, The detection loss for the current domain's true label includes Focal Loss for classification prediction and GIoU Loss for bounding box regression; To address the knowledge distillation loss for high-quality pseudo-labels, this is used to constrain the consistency between the student model and the teacher model in terms of classification probability distribution and regression results; This is an importance coefficient used to balance the learning of new knowledge with the retention of old knowledge.

7. The system according to claim 6, characterized in that, The multi-scale feature construction unit includes: a 1×1×1 convolutional layer, a 3×3×3 convolutional layer, an upsampling layer, a downsampling layer, a channel splicing layer, and a feature aggregation convolutional layer; The multi-scale feature construction unit extracts shallow high-resolution features from the frozen historical prior detection model encoder. Mid-layer characteristics and deep low-resolution features ; for shallow high-resolution features Perform downsampling operation on deep, low-resolution features After performing upsampling to unify the spatial resolution of features across all levels, use... Convolutional dimensionality reduction and concatenation along the channel dimension, finally passed through Convolution completes feature aggregation; the expression for the multi-scale feature building unit is: in, This indicates a downsampling operation. Indicates an upsampling operation; This represents a 1×1×1 convolution operation used for channel dimensionality reduction. This represents a 3×3×3 convolution operation used for feature aggregation; This indicates a channel splicing operation; This represents the joint feature representation after fusion.

8. A method for detecting bibranch pulmonary nodules using combined domain incremental learning and low-dose CT, characterized in that, The method, using the system according to any one of claims 1-7, comprises: Step 1: Acquire low-dose lung CT image data and corresponding lung nodule annotation information, and construct a domain incremental data stream in chronological order; Step 2: Use the dual-branch lung nodule detection system to detect lung nodules in the data stream and output the corresponding lung nodule detection results.

9. A dual-branch pulmonary nodule detection device combining domain incremental learning and low-dose CT, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to implement the method as described in claim 8 when executing the computer program.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program that, when executed by a processor, implements the method as described in claim 8.

Images