A medical image segmentation method and system based on receptive field decoupling and channel adaptive aggregation

Abstract

The present application relates to the technical field of medical image segmentation, in particular to a medical image segmentation method and system based on receptive field decoupling and channel adaptive aggregation, comprising the following steps: constructing a U-shaped encoder-decoder three-dimensional convolutional neural network as a basic segmentation framework, performing multi-scale down-sampling feature extraction on an input 3D CT voxel image through the encoder, and restoring the feature map to the original spatial resolution through the decoder; based on the constructed basic segmentation framework, setting an enabling judgment condition of a conditional decoupling parameterized aggregation block (CDRPA block), and enabling the complete feature refining pipeline of the CDRPA block only when the number of feature channels of the current level of the network reaches a preset channel threshold and the three-dimensional convolution full-dimension step of the level is all 1. The present application discards the traditional optimization idea of external attention plug-in, completes reconstruction from the bottom feature extraction logic, and overcomes three inherent defects existing in three-dimensional organ segmentation from the root.

Description

Technical Field

[0001] This invention relates to the field of medical image segmentation technology, specifically to a medical image segmentation method and system based on receptive field decoupling and channel adaptive aggregation. Background Technology

[0002] In the clinical application of 3D medical imaging, accurate organ and lesion segmentation is the core foundation for quantitative disease analysis, surgical planning, and radiotherapy treatment. Currently, symmetric encoder-decoder 3D convolutional networks, represented by 3DU-Net and nnU-Net, have become the mainstream benchmark framework for medical image segmentation. This type of architecture extracts multi-scale semantic features layer by layer through multi-level downsampling by the encoder, and the decoder gradually restores spatial resolution by combining skip connections. It has shown good generalization ability and robustness in image segmentation tasks of multiple parts such as the abdomen, brain, and lungs. At the same time, the industry is also gradually introducing concepts such as attention mechanisms and grouped convolution into segmentation networks to try to enhance the network's feature selection and multi-scale modeling capabilities, thereby further improving the segmentation effect.

[0003] However, existing 3D medical image segmentation schemes still have several technical shortcomings in practical clinical applications:

[0004] First, the matching degree of receptive fields for organs at multiple scales is insufficient. In abdominal CT images, there is a significant difference in spatial scale between large organs such as the liver and spleen and delicate anatomical structures such as the pancreatic duct and small blood vessels. Traditional networks generally use a single-size convolutional kernel for feature extraction. The fixed receptive field cannot simultaneously take into account the global semantics of large targets and the local texture of small targets, which can easily lead to problems such as missing segmentation of small structures and blurred fitting of organ boundaries, making it difficult to meet the needs of fine segmentation in clinical practice.

[0005] Secondly, the adaptive capability of feature fusion is weak. Existing multi-scale feature aggregation schemes mostly adopt fixed-weight channel splicing or element-wise addition strategies, which cannot dynamically adjust the fusion ratio of features at each scale according to the feature responses of different organs in the input image. This easily leads to the phenomenon of excessively strong feature signals of large targets suppressing the weight polarization of features of small targets, resulting in insufficient adaptability of the model to complex anatomical structures and limited improvement in segmentation accuracy.

[0006] Third, it is difficult to balance computational overhead and segmentation accuracy. Existing accuracy optimization schemes typically embed additional feature processing modules at all network layers, and perform complete fine-grained calculations during the downsampling and dimensionality reduction stage. This results in a significant increase in the number of model parameters and floating-point operations, and a significant decrease in inference speed, making it unsuitable for the efficiency requirements of rapid processing and real-time delineation of large batches of clinical images.

[0007] Fourth, it exhibits poor robustness against cross-sample image heterogeneity. CT images acquired by different patients and different scanning devices show significant differences in grayscale distribution and contrast. Existing network normalization strategies mostly use batch dimension as a benchmark for statistical calibration, which is difficult to effectively smooth out the differences in grayscale distribution of individual images. This can easily lead to a degradation in segmentation accuracy of the model on heterogeneous data, and the generalization ability cannot meet the application requirements of complex clinical scenarios.

[0008] Based on this, the present invention provides a medical image segmentation method and system based on receptive field decoupling and channel adaptive aggregation to solve the above-mentioned technical problems. Summary of the Invention

[0009] The purpose of this invention is to provide a medical image segmentation method and system based on receptive field decoupling and channel adaptive aggregation, thereby solving the problems mentioned in the background art.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] This invention proposes a medical image segmentation method based on receptive field decoupling and channel adaptive aggregation, comprising the following steps:

[0012] S1: Construct a U-shaped encoder-decoder 3D convolutional neural network as the basic segmentation framework. The encoder performs multi-scale downsampling feature extraction on the input 3DCT voxel image, and the decoder upsamples to restore the feature map to the original spatial resolution.

[0013] S2: Based on the basic segmentation framework built by S1, the conditional decoupling parameterized aggregation block, namely the CDRPA block, is set to enable the CDRPA block only when the number of feature channels in the current layer of the network reaches the preset channel threshold and the stride of all dimensions of the three-dimensional convolution of that layer is 1; when the stride of any dimension of the downsampling layer is greater than 1, all internal processing of the CDRPA block is bypassed.

[0014] S3: Based on the activation condition of S2, CDRPA blocks are cascaded in the deep feature processing paths of the encoder and decoder that meet the activation condition. The CDRPA blocks adopt a three-segment serial processing architecture of front-end decoupling extraction segment, middle-segment aggregation segment, and back-end correction segment.

[0015] S4: Based on the CDRPA block of S3, the input feature map is split into two branches according to the preset channel ratio through the front-end decoupling extraction segment. The spatial detail features and channel semantic features are extracted in parallel by 3D group convolution and pointwise convolution, respectively, and the two-way raw features without batch normalization are output.

[0016] S5: Based on the two original features output by S4, batch normalization with learnable parameters is performed through the middle aggregation segment to align the feature distribution. Global feature descriptors are generated based on global average pooling and the fusion weights are adaptively calculated through the normalization exponential function. The dual-branch features are weighted and aggregated into a single multi-scale feature stream.

[0017] S6: Based on the single multi-scale feature stream obtained from S5, random deactivation regularization, instance-normalized grayscale correction and ReLU nonlinear activation are sequentially performed through the back-end correction section to output the corrected refined features;

[0018] S7: Based on the corrected and refined features obtained from S6, the features are passed to the corresponding decoder level through skip connections and fused with the decoder upsampled features. The network parameters are iteratively optimized until the model converges, and the segmentation mask of the 3D medical image is output.

[0019] Preferably, the implementation process of step S1 is as follows:

[0020] S1.1: Construct a symmetrical U-shaped encoder-decoder convolutional neural network framework adapted to 3D voxel images;

[0021] S1.2: In the encoder, multiple cascaded downsampling feature processing layers are set up. Each layer consists of a three-dimensional convolutional layer, a batch normalization layer and a ReLU activation layer in sequence. The downsampling layer adopts a three-dimensional convolution with a stride of 2, and the feature map spatial resolution is compressed to half of the previous layer layer by layer.

[0022] S1.3: In the decoder, multiple upsampling levels symmetrical to the encoder are set up. Three-dimensional transposed convolution is used to restore the spatial resolution of the feature map layer by layer. Before each upsampling level, the output features of the corresponding encoder level are concatenated by channel dimension through skip connections.

[0023] S1.4: Set a 1×1×1 three-dimensional convolutional mapping layer at the end of the decoder to map the multi-channel fused features into a segmentation probability map with the same size as the input voxel space. After normalization by the Softmax activation function, output the organ category prediction result of each pixel.

[0024] Preferably, the implementation process of step S4 is as follows:

[0025] S4.1: Divide the input feature map into a first subset X1 and a second subset X2 according to a preset channel ratio;

[0026] S4.2: Divide the first subset X1 into two subgroups along the channel dimension, and extract local features by using three-dimensional grouped convolution with a kernel size of 3×3×3. The outputs of the two convolutions are concatenated along the channel dimension to obtain the local aggregated feature Z1.

[0027] S4.3: Input the first subset X1 into a 3D convolutional layer with a kernel size of 1×1×1 and perform channel linear mapping to extract cross-channel interaction features Z2;

[0028] S4.4: Add the local aggregation feature Z1 and the cross-channel interaction feature Z2 element by element to obtain the first original feature U1;

[0029] S4.5: Input the second subset X2 into a 3D convolutional layer with a kernel size of 1×1×1 and perform channel linear mapping to obtain the second original feature U2;

[0030] S4.6: The two original features U1 and U2 are not processed by batch normalization and are directly output to the middle aggregation section.

[0031] Preferably, the implementation process of step S5 is as follows:

[0032] S5.1: Input the first original feature U1 and the second original feature U2 into their respective learnable parameter batch normalization layers, perform distribution calibration and alignment, and generate the first calibration feature Y1 and the second calibration feature Y2.

[0033] S5.2: Perform global average pooling on the first calibration feature Y1 and the second calibration feature Y2 respectively to generate the first global descriptor μ1 and the second global descriptor μ2;

[0034] S5.3: Based on the first global descriptor μ1 and the second global descriptor μ2, the adaptive fusion weights α and β are calculated through the normalized exponential function, satisfying α+β=1;

[0035] S5.4: The first calibration feature Y1 and the second calibration feature Y2 are weighted and summed according to the fusion weights α and β to generate a single balanced fusion feature stream Zout.

[0036] Preferably, the implementation process of step S6 is as follows:

[0037] S6.1: Input the balanced fused feature stream Zout into the random deactivation layer, and randomly set some neuron outputs to zero with a preset probability to generate the deactivated feature Dout;

[0038] S6.2: Input the deactivated feature Dout into the instance normalization layer, and perform mean and variance normalization independently on each channel of each sample to generate the grayscale corrected feature Iout;

[0039] S6.3: Input the grayscale-corrected feature Iout into the ReLU activation function for non-linear mapping to generate the corrected and refined feature Rout;

[0040] S6.4: Connect the corrected and refined feature Rout to the skip connection link and transmit it to the corresponding decoder layer.

[0041] Preferably, in step S2, the preset channel threshold is an even number not less than 64; when the stride of any spatial dimension of the three-dimensional convolution of the current sampling level is greater than 1, the current level only performs the standard three-dimensional convolution downsampling operation and does not start the three-stage processing pipeline inside the CDRPA block.

[0042] Preferably, the implementation process of step S7 is as follows:

[0043] S7.1: The corrected and refined features output from each level of CDRPA block are sent to the corresponding stage layer of the decoder through skip connections, and fused with the features upsampled by the decoder at the channel dimension.

[0044] S7.2: The decoder outputs the final segmentation prediction result, and calculates the total loss value by combining it with the segmentation loss function;

[0045] S7.3: Perform backpropagation with the total loss value as the optimization objective, iteratively update all network parameters until the model converges, and output a three-dimensional segmentation mask of multiple organs and small anatomical structures in the abdomen.

[0046] This invention also proposes a medical image segmentation system based on receptive field decoupling and channel adaptive aggregation, which is applied to the aforementioned segmentation method. The system is embedded in the basic segmentation framework of a U-shaped encoder-decoder three-dimensional convolutional neural network, including a conditionally decoupled parameterized aggregation block, namely a CDRPA block.

[0047] The CDRPA block integrates a three-stage serial processing pipeline, which consists of:

[0048] The front-end decoupled extraction unit is used to split the input feature map into two branches according to a preset channel ratio. The spatial detail features and channel semantic features are extracted in parallel by three-dimensional grouped convolution and pointwise convolution, respectively, and the two unbatched normalized original features are output.

[0049] The middle aggregation unit is used to receive the two original features, perform distribution calibration through a batch normalization layer with learnable parameters, generate dual-channel descriptors through global average pooling, and adaptively calculate the fusion weights through a normalized exponential function to converge the weighted sum of the two-branch features into a single balanced feature stream.

[0050] The back-end correction unit is used to sequentially perform random deactivation regularization, instance-normalized grayscale correction and ReLU nonlinear activation on the single balanced feature stream, and output the corrected refined features to the decoder skip connection link.

[0051] The CDRPA block selectively activates the internal complete pipeline based on dual judgment conditions: the three-stage serial processing pipeline is activated only when the number of feature channels in the current layer of the network reaches a preset channel threshold and the stride of all three-dimensional convolutions in that layer is 1; when the stride of any dimension of the current sampling layer is greater than 1, the current layer bypasses all internal processing of the CDRPA block.

[0052] Compared with the prior art, the beneficial effects of the present invention are:

[0053] This invention addresses the challenge of capturing the significant scale differences between large organs and small blood vessels in 3D CT images, which cannot be adequately addressed by a single convolutional receptive field. It employs a decoupled, parallel extraction mechanism using dual channels at the front end of a CDRPA block, simultaneously capturing local texture and global semantics to achieve feature separation and extraction for targets of varying sizes. This effectively improves the segmentation of small pancreatic ducts and blood vessels, preventing missed or incorrect segmentation. Leveraging a learnable batch normalization distribution alignment mechanism built into the middle section of the CDRPA block, it generates fusion weights in real-time based on input features, abandoning the traditional fixed-weight attention structure. This automatically adjusts the proportion of multi-scale features for different organ images, significantly improving channel feature discrimination. Furthermore, this invention features a dual-enablement logic for the CDRPA block, automatically skipping complete streams within the block at the downsampling level. The waterline enables refined feature processing only in the feature refinement layer, separating downsampling and feature refinement operations, significantly reducing the overall computational cost of the network and adapting to real-time clinical delineation scenarios. The CDRPA block post-correction sequence integrates regularization, equalization, and activation operations, which both suppress overfitting in 3D voxel training through Dropout and smooth out differences in CT brightness and contrast among different patients through instance standardization, reducing segmentation bias caused by individual images. The CDRPA integrated block is not limited by the encoder level and can be embedded in all non-downsampling feature layers of the U-shaped network. Compared with optimization schemes that only act on the encoding branches, it fully preserves the fine-grained features in the multi-scale recovery process of the decoder, resulting in better overall segmentation contour and edge restoration effects. Attached Figure Description

[0054] Figure 1 This is a flowchart of the medical image segmentation method based on receptive field decoupling and channel adaptive aggregation according to the present invention. Detailed Implementation

[0055] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0056] Example 1, please refer to Figure 1In practical applications, this embodiment provides a medical image segmentation method based on receptive field decoupling and channel adaptive aggregation. The specific implementation steps are as follows:

[0057] Step 1: Based on the U-shaped encoder-decoder 3D neural network as the basic segmentation framework, a symmetrical 3D encoding and decoding structure is built. Relying on multi-level feature stacking, downsampling compression and upsampling restoration mechanism, multi-scale depth feature extraction and accurate spatial resolution restoration of 3DCT voxel images are completed, providing a highly robust basic feature expression for subsequent refined feature optimization.

[0058] S1.1 Select a symmetrical U-shaped encoder-decoder network adapted to the features of three-dimensional voxel images as the basic segmentation framework to meet the stereo feature representation requirements of 3DCT volume data.

[0059] S1.2 Build a multi-level progressive feature processing hierarchy, with each level configured with a feature aggregation layer, a batch normalization layer and a ReLU activation layer to ensure feature stability and non-linear expressive ability;

[0060] The encoder downsampling layers uniformly adopt three-dimensional convolution with a stride of 2, compressing the feature space resolution layer by layer to achieve multi-scale semantic extraction, and compressing the feature map space resolution layer by layer to half of the previous layer.

[0061] S1.3 Construct a decoder upsampling layer that is completely symmetrical to the encoder. The feature space size is restored layer by layer through three-dimensional transposed convolution. A skip channel splicing structure at the same level is configured to compensate for the loss of fine-grained information during the downsampling process.

[0062] S1.4. Configure a 1×1×1 three-dimensional convolutional mapping layer at the end of the network to perform dimensional mapping on the multi-channel fused features, and map the high-dimensional features into a segmentation probability map with the same size as the input voxel space.

[0063] The feature map output by the three-dimensional convolutional mapping layer is normalized by the Softmax activation function to generate the probability distribution of each pixel in different organ categories, and outputs the prediction result of the organ category to which each pixel belongs.

[0064] Step 2: Set the CDRPA block, i.e., the conditional decoupling parameterized aggregation block, with dual activation conditions. This limits the total number of input channels in the current layer of the network to an even number of not less than 64, and combines the dual constraints of the stride of the three-dimensional convolution to achieve conditional activation.

[0065] The intra-block full feature refinement pipeline is enabled only when the current feature level of the network meets the aforementioned channel threshold and the stride of all three-dimensional convolutions is 1. For downsampling levels with a stride greater than 1 in any dimension, the intra-block parallel processing branch is skipped directly, thereby achieving hierarchical decoupling between the spatial downsampling dimensionality reduction task and the deep feature refinement task, and reducing overall computational redundancy without losing segmentation accuracy.

[0066] The specific conditions for dual activation in this step are: the total number of input channels of the current network feature level is greater than or equal to an even threshold m of 64, and the stride of the length, width and depth dimensions of the three-dimensional convolution of this level is equal to 1, and the sum of the three-dimensional stride parameters is equal to 3.

[0067] If any judgment condition is not met, the current level is determined to be a downsampling compression level. The network directly skips all pipeline operations inside the CDRPA block and only performs standard 3D convolution to complete feature dimensionality reduction, thereby achieving intelligent allocation of computing power.

[0068] Step 3: In the deep feature processing paths of the encoder and decoder that meet the triggering conditions, cascade the CDRPA integrated encapsulation architecture layer by layer.

[0069] The CDRPA block internally constructs a three-stage integrated serial processing pipeline consisting of a front-end decoupling extraction unit, a middle-stage aggregation unit, and a back-end correction unit, forming a complete feature optimization closed loop.

[0070] The pipeline front end uses a front-end decoupling extraction unit to achieve multi-scale receptive field differential extraction through a dual-branch parallel encoding strategy based on channel decoupling. It outputs two unprocessed original features in parallel, fully preserving the original feature distribution information and avoiding the loss of low-level feature details caused by normalization operations.

[0071] In this step of the CDRPA block front-end decoupling extraction process, the input feature map dimension is B×C×H×W×D, the preset channel segmentation ratio parameter r, and the value range of r is 0.2≤r≤0.8;

[0072] The overall input features are split along the channel dimension to obtain a first feature subset X1 and a second feature subset X2, where the number of channels in X1 is ⌊C・r⌋, and X2 retains all remaining channel features, thus achieving a structural decoupling and allocation between local detail channels and global semantic channels.

[0073] The front-end decoupled extraction process performs only pure convolutional mapping operations without introducing any batch normalization, thus preserving the original feature distribution characteristics to the greatest extent. Specific implementation methods include:

[0074] S4.1. The first subset X1 is fed into the local detail extraction branch. X1 is divided into g=2 subgroups, and 3D grouped convolution with a kernel size of 3×3×3 is used to complete the fine extraction of local texture. The output features of the two groups of convolutions are concatenated and fused along the channel dimension to obtain the local aggregated feature Z1=Concat(φ(X 1a ;W1),φ(X1ᵦ;W2)), where φ is the three-dimensional local feature aggregation operator, and W1 and W2 are the independent learnable weight parameters of the two groups of convolutions.

[0075] S4.2. Simultaneously perform 1×1×1 three-dimensional channel linear mapping on the first subset X1 in parallel to realize cross-channel semantic interaction modeling and obtain cross-channel fusion feature Z2=ψ(X1;Wp1), where ψ represents the 1×1×1 three-dimensional convolutional channel linear mapping operator and Wp1 is the independent learnable weight parameter of this branch.

[0076] S4.3. The local aggregation feature Z1 of the same dimension and the cross-channel interaction feature Z2 are added and fused element by element to output the unnormalized first branch original feature U1=Z1+Z2, which is used as the first branch output of the front-end decoupling extraction unit.

[0077] S4.4. Perform a 1×1×1 three-dimensional channel linear mapping directly on the second subset X2 to complete the lightweight encoding of global semantic features and output the unnormalized original features of the second branch U2=ψ(X2;Wp2), where Wp2 is a learnable weight parameter independent of Wp1.

[0078] The two original features U1 and U2 are not processed by batch normalization, and their original feature distribution is fully preserved. They are then input into the aggregation unit in the middle section of the CDRPA block.

[0079] Step 4: The middle aggregation unit inside the CDRPA block serves as the core module for feature fusion. It receives the two original features output from the front-end decoupling extraction unit and sequentially performs dual-channel feature distribution calibration through batch normalization with learnable parameters to eliminate the feature distribution offset problem of different branches.

[0080] Furthermore, by combining three-dimensional global mean statistics and normalized exponential function to achieve adaptive weighted fusion, the multi-scale divergent bi-branch features are dynamically converged and aggregated into a single-path balanced multi-scale feature flow, ensuring feature scale balance and semantic consistency, and the aggregated high-quality features are sent to the back-end correction unit.

[0081] The parameterized distribution alignment adaptive aggregation process in this step specifically includes:

[0082] S5.1 Receive the two original features U1 and U2 output from the front-end decoupling extraction unit, respectively connect them to the independent learnable parameter batch normalization layer, complete the adaptive calibration and alignment of the dual-channel feature distribution, eliminate the difference in the mean and variance distribution of the dual-branch features, and generate the balanced calibration features Y1=BN(U1) and Y2=BN(U2).

[0083] Wherein, BN(・) represents the batch normalization operation, and each batch normalization layer is configured with an independent learnable scaling factor and offset factor.

[0084] S5.2 Perform three-dimensional global mean pooling (GAP) on calibration features Y1 and Y2 respectively, extract global statistical information, and generate dual-channel global feature descriptors μ1=GAP(Y1) and μ2=GAP(Y2);

[0085] The dynamic adaptive fusion weights of the two features are calculated based on two sets of global descriptors using a normalized exponential function. The weight calculation formula is as follows:

[0086] ;

[0087] ;

[0088] The weight coefficients satisfy α+β=1, and the values ​​of α and β are both in the range of [0,1], which ensures the stability and balance of the dual-branch feature fusion.

[0089] S5.3. Based on the dynamic weights, the dual-path calibration features are weighted and fused to obtain a scale-balanced and semantically unified single-path fused feature Zout=α・Y1+β・Y2, thus completing the convergence and aggregation of multi-scale divergent features into a single high-quality feature stream.

[0090] Step 5: Construct a hybrid complementary enhancement strategy of "internal distributed constraints + shell post-processing", and complete feature robustness optimization by relying on the back-end correction unit inside the CDRPA block;

[0091] The aggregated balanced features are sequentially fed into cascaded random deactivation (Dropout), instance normalization (IN), and ReLU nonlinear activation units to form a composite robust correction sequence of regularization, gray-level equalization, and nonlinear enhancement, thereby achieving model overfitting prevention, smoothing of cross-sample gray-level differences, and enhancement of deep feature expression.

[0092] Finally, by combining multi-layer skip connections to complete multi-scale feature iterative fusion, continuously updating the parameters of all network modules, and completing iterative training convergence through a 3D medical image dataset, a high-precision segmentation mask for multiple organs and fine anatomical structures in the abdomen is accurately output.

[0093] The hybrid complementary enhancement strategy of "internal distributed constraints + shell post-processing" in this step specifically includes:

[0094] S7.1 Internal Distribution Constraint: Relying on the learnable batch normalization operator BN in the middle of the CDRPA block, the original features of the two branches are precisely calibrated to constrain the distribution range of channel features, stabilize the output range of adaptive fusion weights, and avoid the problems of polarization of multi-channel weights and imbalance of feature response.

[0095] S7.2 Built-in robust correction sequence cascade: The single equalization feature output by the mid-section convergence aggregation is sent to the post-serial pipeline constructed by the CDRPA block, and random deactivation, instance normalization and nonlinear activation operations are executed in sequence to form a three-in-one equalization regularization activation robust correction sequence of "regularization-grayscale equalization-nonlinear activation".

[0096] S7.3, Functional Division of Labor and Fusion Logic: Among them, the regularized random deactivation operation is used to alleviate the risk of sample overfitting during the training process of 3D medical images and improve the generalization ability of the model.

[0097] The standardized operation of equalization instance performs global grayscale equalization to smooth out individual differences in CT image brightness and contrast caused by different patients and different scanning equipment;

[0098] Activation operations restore the nonlinear expressive power of deep features and enrich the semantic information of features;

[0099] After robust correction, the refined features are connected to the network via a skip connection fusion link and transmitted to the corresponding decoder level. They are then merged with the current level features of the decoder to complete channel splicing and fusion. The spatial attention focus area is preserved throughout the process, reducing the segmentation deviation of abdominal organ edges and small anatomical structures, and participating in the network's iterative optimization throughout the process.

[0100] Example 2: This example provides a medical image segmentation system based on receptive field decoupling and channel adaptive aggregation, applied to the segmentation method described in Example 1. The system is embedded within the U-shaped encoder-decoder three-dimensional convolutional neural network basic segmentation framework, and the core innovative unit of the system is the CDRPA integrated conditional decoupling parameterized aggregation block.

[0101] The CDRPA integrated block, as an overall packaging architecture, selectively enables the internal front-end decoupling extraction unit, mid-section aggregation unit, and back-end correction unit three-stage serial pipeline based on the dual judgment conditions of channel and step size, and dynamically switches between the refinement mode and the lightweight downsampling mode according to the network layer function.

[0102] The CDRPA block's internal front end is a front-end decoupling extraction unit, used to split channel features according to a preset ratio, and output two unnormalized native divergent features through parallel encoding of local group convolution and global 1×1 linear mapping.

[0103] The middle section of the CDRPA block is a middle aggregation unit, which is used to complete dual-channel feature distribution calibration, global statistical modeling and adaptive weighted fusion, and converge the divergent multi-scale features into a balanced single-path feature flow.

[0104] The CDRPA block's internal backend is a backend correction unit that integrates random deactivation, instance normalization, and nonlinear activation operators to complete model regularization optimization and image grayscale adaptive correction, ultimately outputting high-quality refined features to be sent into the network skip fusion link.

[0105] The method of this invention, through a five-fold cross-validation experiment based on the AbdomenCT-1K dataset, comprises the following steps:

[0106] Using abdominal CT scans from 1000 cases, along with labeled real segmentation masks for multiple abdominal organs including the liver, kidneys, spleen, and pancreas, the data covers abdominal images from different patients, using different imaging devices, and with varying degrees of pathological disease. Based on this, five-fold cross-validation was performed.

[0107] The complete dataset is randomly and equally divided into 5 non-overlapping subsets with consistent data distribution.

[0108] Four of the samples were selected as the training set and the remaining one as the validation set. The training and validation were repeated five times, with a different validation subset each time.

[0109] Finally, the average of the various metrics from the five validations is taken as the final performance result of the model.

[0110] On the AbdomenCT-1K abdominal CT dataset, the standard evaluation method of five-fold cross-validation was used to compare the performance indicators of the original baseline network and the network with the CDRPA module of this invention, thereby verifying the improvement effect of this invention in terms of segmentation accuracy, boundary fitting, and model lightweighting.

[0111] The statistical results of the five-fold cross-validation experiment are shown in Tables 1 and 2, as follows:

[0112] Table 1a: Explanation of Segmentation Performance Indicators

[0113] Model / Method The two models compared are: baseline (the original baseline network) and CDRPA (an optimized model incorporating the modules of this invention). Grouping Dimensions Fold The results of single-fold cross-validation (Fold0~Fold4) and the five-fold mean (Mean) Experimental dimensions MeanValidationDice The average Dice coefficient on the validation set measures the spatial overlap between the predicted segmentation region and the ground truth annotation; the closer the value is to 1, the more accurate the segmentation. Core positive metrics, the higher the better. HD95↓ The 95% Hausdorff distance measures the maximum spatial error (usually in mm) between the segmentation boundary and the true boundary, reflecting the accuracy of organ contour fitting. Negative indicators, ↓ means the smaller the better. IOU↑ Intersection over Union (IoU) is the ratio of the intersection to the union of the predicted and actual regions, measuring the overall overlap of the segmentation. Positive indicators, the higher the better. Recall↑ Recall rate, the proportion of the true target region that is correctly segmented, reflects the degree of missed segmentation. Positive indicators, the higher the better. Precision↑ Accuracy, the proportion of regions predicted as targets that are actually targets, reflects the degree of missegmentation. Positive indicators, the higher the better.

[0114] Table 2a: Comparison and Explanation of Computational and Parameter Quantities

[0115] Model / Method Two types of models in comparison Grouping Dimensions FLOPs(G)↓ Floating-point operations, measured in billions of operations per second (G), measure the computational complexity of model inference. The smaller the value, the faster the inference speed and the lower the computing power consumption. Negative indicators, ↓ means the smaller the better. Params(M)↓ The number of learnable parameters in a model, measured in millions (M), measures the size of the model. The smaller the number, the lighter the model and the less GPU memory it uses. Negative indicators, ↓ means the smaller the better.

[0116] Table 1: Comparison of Segmentation Performance Indicators

[0117] Model / Method Fold MeanValidationDice HD95↓ IOU↑ Recall↑ Precision↑ baseline Fold0 0.915166478 32.7528689 0.859502625 0.959705535 0.939764523 CDRPA Fold0 0.920432736 26.0433217 0.867112841 0.955902698 0.955177128 baseline Fold1 0.911656864 37.44815564 0.853288062 0.960070785 0.930252228 CDRPA Fold1 0.922759524 20.47984672 0.869212296 0.954016311 0.954774212 baseline Fold2 0.92279284 27.46239523 0.868385371 0.957983855 0.957027983 CDRPA Fold2 0.926243738 24.84934888 0.873434032 0.967896054 0.949823856 baseline Fold3 0.915566428 48.05424818 0.858307929 0.95708958 0.945169171 CDRPA Fold3 0.922578135 30.02425031 0.868903327 0.963312548 0.948565327 baseline Fold4 0.918883582 34.07713578 0.864439986 0.947227957 0.957181233 CDRPA Fold4 0.921635911 28.94567013 0.868302027 0.95346751 0.956395027 baseline Mean 0.916813238 35.95896075 0.860784794 0.956415542 0.945879028 CDRPA Mean 0.922730009 26.06848755 0.869392905 0.958919024 0.95294711

[0118] Table 2: Comparison of Computational and Parameter Quantities

[0119] Model / Method FLOPs(G)↓ Params(M)↓ baseline 562.1297 30.7884 CDRPA 196.9112 14.2412

[0120] As shown in Tables 1 and 2, compared with existing technologies, the medical image segmentation method and system based on receptive field decoupling and channel adaptive aggregation provided by this invention effectively overcomes the inherent defects of traditional three-dimensional U-shaped segmentation networks through conditional triggering decoupling of an integrated CDRPA block, dual-branch receptive field decoupling extraction, parameterized adaptive aggregation, and robust correction closed-loop design. Experimental results based on five-fold cross-validation on the AbdomenCT-1K dataset demonstrate that this invention significantly improves overall performance compared to the baseline model, specifically possessing the following beneficial effects:

[0121] This invention utilizes a CDRPA block front-end dual-channel receptive field decoupling mechanism to decouple the extraction of local fine textures from global organ semantics, solving the technical challenge that a single convolutional receptive field cannot adapt to the size scale differences of abdominal organs. Experimental data show that the average Dice coefficient of the model improved from 0.9168 to 0.9227, and the Intersection over Union (IOU) improved from 0.8608 to 0.8694; recall and precision were simultaneously improved, effectively reducing the probability of missed or incorrect segmentation of small tissues such as microvessels and pancreatic ducts, and significantly improving the segmentation integrity and accuracy of multiple abdominal organs and fine anatomical structures.

[0122] This invention relies on the CDRPA block mid-segment parametric distribution alignment and aggregation mechanism to eliminate the difference in feature distribution between two branches. Combined with a post-robust correction sequence to optimize feature expression, it effectively corrects the problems of blurred organ edges and distorted contours. Compared with the baseline model, the HD95 Hausdorff distance of this invention is reduced from 35.96 to 26.07, which significantly reduces the spatial offset error of organ edges and significantly optimizes the fitting accuracy of complex adhesion edges and fine structural boundaries of abdominal organs, making it suitable for the needs of fine clinical segmentation.

[0123] This invention decouples the downsampling dimensionality reduction task from the feature refinement task through a dual-channel, step-size conditional triggering mechanism of CDRPA blocks. The downsampling level skips redundant feature processing pipelines, greatly reducing unnecessary computational power consumption. Experimental results show that the floating-point operation volume (FLOPs) of this invention is reduced from a baseline of 562.13G to 196.91G, a computational power reduction of 64.9%; the number of model parameters is compressed from 30.79M to 14.24M, a parameter reduction of over 50%. This invention achieves lightweight model and high-speed inference without reducing segmentation accuracy, perfectly adapting to scenarios requiring rapid processing of large batches of clinical CT images and real-time radiotherapy delineation.

[0124] This invention's CDRPA block incorporates a hybrid enhancement strategy of "internal distribution constraint + outer shell grayscale equalization correction." It uses learnable batch normalization to constrain feature distribution and instance standardization to smooth out individual differences in brightness and contrast in CT images from different patients. Combined with a random inactivation operator, it suppresses overfitting during 3D voxel training. Five-fold cross-validation results in less fluctuation across various metrics, with accuracy and specificity superior to the baseline model. It effectively adapts to clinical CT images acquired from different devices and exhibiting individual patient differences, significantly improving the model's versatility and anti-interference capabilities.

[0125] This invention abandons the traditional fixed attention weight mechanism and relies on the CDRPA block dynamic adaptive fusion strategy to generate fusion weights in real time according to the input image feature response. It can automatically adapt the weight ratio of global features of large organs and local textures of small structures, solve the problem of over-response of large organs and suppression of small target features in traditional models, achieve balanced fusion of multi-scale features, and further ensure the segmentation stability of complex abdominal CT images.

[0126] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0127] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A medical image segmentation method based on receptive field decoupling and channel adaptive aggregation, characterized in that, Includes the following steps: S1: Construct a U-shaped encoder-decoder 3D convolutional neural network as the basic segmentation framework. The encoder performs multi-scale downsampling feature extraction on the input 3DCT voxel image, and the decoder upsamples to restore the feature map to the original spatial resolution. S2: Based on the basic segmentation framework built by S1, the conditional decoupling parameterized aggregation block, i.e. CDRPA block, is set to enable the complete feature refining pipeline of CDRPA block only when the number of feature channels of the current layer of the network reaches the preset channel threshold and the stride of all dimensions of the three-dimensional convolution of that layer is 1. When the step size in any dimension of the current sampling level is greater than 1, bypass all internal processing of the CDRPA block; S3: Based on the activation condition of S2, CDRPA blocks are cascaded in the deep feature processing paths of the encoder and decoder that meet the activation condition. The CDRPA blocks adopt a three-segment serial processing architecture of front-end decoupling extraction segment, middle-segment aggregation segment, and back-end correction segment. S4: Based on the CDRPA block of S3, the input feature map is split into two branches according to the preset channel ratio through the front-end decoupling extraction segment. The spatial detail features and channel semantic features are extracted in parallel by 3D group convolution and pointwise convolution, respectively, and the two-way raw features without batch normalization are output. S5: Based on the two original features output by S4, batch normalization with learnable parameters is performed through the middle aggregation segment to align the feature distribution. Global feature descriptors are generated based on global average pooling and the fusion weights are adaptively calculated through the normalization exponential function. The dual-branch features are weighted and aggregated into a single multi-scale feature stream. S6: Based on the single multi-scale feature stream obtained from S5, random deactivation regularization, instance-normalized grayscale correction and ReLU nonlinear activation are sequentially performed through the back-end correction section to output the corrected refined features; S7: Based on the corrected and refined features obtained from S6, the features are passed to the corresponding decoder level through skip connections and fused with the decoder upsampled features. The network parameters are iteratively optimized until the model converges, and the segmentation mask of the 3D medical image is output.

2. The medical image segmentation method based on receptive field decoupling and channel adaptive aggregation according to claim 1, characterized in that, The implementation process of step S1 is as follows: S1.1: Construct a symmetrical U-shaped encoder-decoder convolutional neural network framework adapted to 3D voxel images; S1.2: In the encoder, multiple cascaded downsampling feature processing layers are set up. Each layer consists of a three-dimensional convolutional layer, a batch normalization layer and a ReLU activation layer in sequence. The downsampling layer adopts a three-dimensional convolution with a stride of 2, and the feature map spatial resolution is compressed to half of the previous layer layer by layer. S1.3: In the decoder, multiple upsampling levels symmetrical to the encoder are set up. Three-dimensional transposed convolution is used to restore the spatial resolution of the feature map layer by layer. Before each upsampling level, the output features of the corresponding encoder level are concatenated by channel dimension through skip connections. S1.4: Set a 1×1×1 three-dimensional convolutional mapping layer at the end of the decoder to map the multi-channel fused features into a segmentation probability map with the same size as the input voxel space. After normalization by the Softmax activation function, output the organ category prediction result of each pixel.

3. The medical image segmentation method based on receptive field decoupling and channel adaptive aggregation according to claim 2, characterized in that, The implementation process of step S4 is as follows: S4.1: Divide the input feature map into a first subset X1 and a second subset X2 according to a preset channel ratio; S4.2: Divide the first subset X1 into two subgroups along the channel dimension, and extract local features by using three-dimensional grouped convolution with a kernel size of 3×3×3. The outputs of the two convolutions are concatenated along the channel dimension to obtain the local aggregated feature Z1. S4.3: Input the first subset X1 into a 3D convolutional layer with a kernel size of 1×1×1 and perform channel linear mapping to extract cross-channel interaction features Z2; S4.4: Add the local aggregation feature Z1 and the cross-channel interaction feature Z2 element by element to obtain the first original feature U1; S4.5: Input the second subset X2 into a 3D convolutional layer with a kernel size of 1×1×1 and perform channel linear mapping to obtain the second original feature U2; S4.6: The two original features U1 and U2 are not processed by batch normalization and are directly output to the middle aggregation segment.

4. The medical image segmentation method based on receptive field decoupling and channel adaptive aggregation according to claim 3, characterized in that, The implementation process of step S5 is as follows: S5.1: Input the first original feature U1 and the second original feature U2 into their respective learnable parameter batch normalization layers, perform distribution calibration and alignment, and generate the first calibration feature Y1 and the second calibration feature Y2. S5.2: Perform global average pooling on the first calibration feature Y1 and the second calibration feature Y2 respectively to generate the first global descriptor μ1 and the second global descriptor μ2; S5.3: Based on the first global descriptor μ1 and the second global descriptor μ2, the adaptive fusion weights α and β are calculated through the normalized exponential function, satisfying α+β=1; S5.4: The first calibration feature Y1 and the second calibration feature Y2 are weighted and summed according to the fusion weights α and β to generate a single balanced fusion feature stream Zout.

5. The medical image segmentation method based on receptive field decoupling and channel adaptive aggregation according to claim 4, characterized in that, The implementation process of step S6 is as follows: S6.1: Input the balanced fused feature stream Zout into the random deactivation layer, and randomly set some neuron outputs to zero with a preset probability to generate the deactivated feature Dout; S6.2: Input the deactivated feature Dout into the instance normalization layer, and perform mean and variance normalization independently on each channel of each sample to generate the grayscale corrected feature Iout; S6.3: Input the grayscale-corrected feature Iout into the ReLU activation function for non-linear mapping to generate the corrected and refined feature Rout; S6.4: Connect the corrected and refined feature Rout to the skip connection link and transmit it to the corresponding decoder layer.

6. The medical image segmentation method based on receptive field decoupling and channel adaptive aggregation according to claim 5, characterized in that, In step S2, the preset channel threshold is an even number not less than 64; when the stride of any spatial dimension of the three-dimensional convolution of the current sampling level is greater than 1, the current level only performs the standard three-dimensional convolution downsampling operation and does not start the three-stage processing pipeline inside the CDRPA block.

7. A medical image segmentation method based on receptive field decoupling and channel adaptive aggregation according to claim 6, characterized in that, The implementation process of step S7 is as follows: S7.1: The corrected and refined features output from each level of CDRPA block are sent to the corresponding stage layer of the decoder through skip connections, and fused with the features upsampled by the decoder at the channel dimension. S7.2: The decoder outputs the final segmentation prediction result and calculates the total loss value by combining it with the segmentation loss function; S7.3: Perform backpropagation with the total loss value as the optimization objective, iteratively update all network parameters until the model converges, and output a 3D segmentation mask of multiple organs and small anatomical structures in the abdomen.

8. A medical image segmentation system based on receptive field decoupling and channel adaptive aggregation, characterized in that, The system, which is applied to the segmentation method of claim 1, is embedded in the U-shaped encoder-decoder three-dimensional convolutional neural network basic segmentation framework and includes a conditionally decoupled parameterized aggregation block, namely a CDRPA block. The CDRPA block integrates a three-stage serial processing pipeline, which consists of: The front-end decoupled extraction unit is used to split the input feature map into two branches according to a preset channel ratio. The spatial detail features and channel semantic features are extracted in parallel by three-dimensional grouped convolution and pointwise convolution, respectively, and the two unbatched normalized original features are output. The middle aggregation unit is used to receive the two original features, perform distribution calibration through a batch normalization layer with learnable parameters, generate dual-channel descriptors through global average pooling, and adaptively calculate the fusion weights through a normalized exponential function to converge the weighted sum of the two-branch features into a single balanced feature stream. The back-end correction unit is used to sequentially perform random deactivation regularization, instance-normalized grayscale correction and ReLU nonlinear activation on the single balanced feature stream, and output the corrected refined features to the decoder skip connection link. The CDRPA block selectively activates the internal complete pipeline based on dual judgment conditions: the three-stage serial processing pipeline is activated only when the number of feature channels in the current layer of the network reaches a preset channel threshold and the stride of all three-dimensional convolutions in that layer is 1; when the stride of any dimension of the current sampling layer is greater than 1, the current layer bypasses all internal processing of the CDRPA block.

Images