Medical image segmentation method based on adaptive regularization network module

Abstract

This invention discloses a medical image segmentation method based on an adaptive regularization network module, relating to the interdisciplinary field of medical image processing and computer vision. The method includes: S1, constructing an adaptive regularization network module MARN, which includes an anatomical perceptual space regularization submodule and a multi-scale channel modulation (MFACM) submodule; S2, inserting MARN into all basic blocks of a deep learning network HRNet, inputting the preprocessed medical image into the ASR submodule for processing, and outputting weighted fused regularized features; S3, fusing the regularized features with the original features to obtain an anatomical region quality score; S4, modulating the input into the MFACM submodule to obtain modulated output features; and S5, inputting the input into the HRNet task head to generate the medical image segmentation result through convolution and activation functions. This invention solves the problems of unstable feature distribution and large statistical bias in traditional normalization mechanisms under conditions of small samples, low contrast, and high noise in medical scenarios.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of medical image processing and computer vision. More specifically, this invention relates to a medical image segmentation method based on an adaptive regularization network module. Background Technology

[0002] Medical image segmentation is a core technology in computer-aided diagnosis, and its accuracy directly determines the reliability of disease diagnosis. HRNet, due to its ability to maintain high-resolution features throughout the process, has become the mainstream model for medical image segmentation. However, the batch normalization (BN) mechanism relied upon by traditional HRNet has two major flaws that severely limit its segmentation performance:

[0003] First, the statistical reliability is insufficient: batch normalization relies on batch statistical information. In small sample medical data, the batch mean and variance cannot reflect the global data distribution, which is prone to statistical bias and leads to unstable feature distribution. Moreover, batch normalization uses a uniform statistic for all spatial locations of the feature map, which cannot adapt to the local differences between lesions and background. When the proportion of lesions is extremely low, background statistical information will dominate the normalization and weaken the lesion features.

[0004] Secondly, it has weak anti-interference ability: In low-contrast, high-noise medical images, noise will interfere with the estimation of BN statistics, further reducing the segmentation accuracy.

[0005] To address the normalization problem, existing techniques such as Instance Normalization (IN), Adaptive Instance Normalization (AdaIN), and CrossNorm (CN) have been improved, but they still have the following shortcomings: Instance Normalization only normalizes a single sample and cannot utilize the structural correlation information of batch data, making it easy to lose local features in small sample scenarios; Adaptive Instance Normalization has large parameter estimation biases and low reliability of statistical information in small sample scenarios; CrossNorm only focuses on the mixture of statistics between samples and does not optimize for the anatomical structure of medical images, making it difficult to highlight the normalization effect of key lesion areas.

[0006] Therefore, existing technologies have not simultaneously solved the two core problems of "feature normalization stability" and "insufficient anatomical structure modeling ability", and lack regularization schemes adapted to the characteristics of medical data. There is an urgent need for a network module that can collaboratively optimize the processing of spatial and channel dimension features. Summary of the Invention

[0007] One object of the present invention is to solve at least the above-mentioned problems and / or defects, and to provide at least the advantages described below.

[0008] To achieve these objectives and other advantages of the present invention, a medical image segmentation method based on an adaptive regularization network module is provided, characterized by comprising:

[0009] S1. Construct an adaptive regularized network module MARN, which includes: an anatomical perception spatial regularization (ASR) submodule focusing on the spatial dimension and a multi-scale channel modulation (MFACM) submodule focusing on the channel dimension.

[0010] S2. After processing the preprocessed medical image in the ASR submodule, the weighted fused regularized features are output. :

[0011]

[0012] In the above formula, λ is the feature fusion weight that controls the regularization strength. For the original features x Features after standardization and normalization, original features x Refers to the initial input characteristics received by the ASR submodule. These are the enhanced features generated after global fusion of the ASR submodules;

[0013] S3. Regularize the features using the following formula. With original features x By fusion, the quality scoring characteristics of the anatomical region are obtained. :

[0014]

[0015] In the above formula, To control the learnable balance factor between global information and local details, It is a range mapping activation function;

[0016] S4, will The input is modulated in the MFACM submodule to obtain the modulation output characteristics using the following formula. :

[0017]

[0018] In the above formula, Scaling factor As the offset factor, To The feature values ​​after channel attention enhancement for The channel statistical mean;

[0019] S5, will The task head is fed into the deep learning network HRNet to generate segmentation results for medical images through convolution and activation functions.

[0020] Preferably, in S2, the enhanced features Obtained through the following formula:

[0021] In the above formula, Blend ( ) represents the feature fusion function. It is a stable benchmark feature generated based on global statistical information from similar medical datasets. For the original features x The transformation features after normalization, and , , For each channel c The mean and standard deviation obtained after performing medical characteristic statistics , These are the reference mean and reference standard deviation, respectively. mask For the mask, This refers to the local features of the key anatomical regions after adaptive cropping.

[0022] Preferably, each channel c mean It is characterized by the following formula:

[0023]

[0024] In the above formula, N is the batch size. c H represents the number of channels, and H and W represent the spatial dimensions. Original feature map x Pixel-level feature values ​​at corresponding samples, corresponding channels, and corresponding spatial locations;

[0025] Each channel c standard deviation It is characterized by the following formula:

[0026]

[0027] In the above formula, To avoid small, unstable adjustment values, and .

[0028] Preferably, in S4, Obtained through the following formula:

[0029]

[0030] In the above formula, For the average attention of the channels, For channel attention, and , Indicates to Perform global average pooling. ReLU represents the convolution operation, and ReLU() represents the activation function. For channel attention, use the sigmoid activation function. To control and Organizational sensitivity weighting for integration.

[0031] Preferably, the deep learning network HRNet includes the following four-stage feature extraction process:

[0032] The first stage consists of four cascaded bottleneck blocks, each of which includes: a 1×1 convolutional dimensionality reduction layer, a 3×3 convolutional layer, a 1×1 convolutional dimensionality increase layer, and a residual connection layer.

[0033] The second phase includes one high-resolution module;

[0034] Phase 3: Contains 4 high-resolution modules;

[0035] Phase 4: Contains 3 high-resolution modules;

[0036] Each high-resolution module branch's base block is embedded with MARN;

[0037] In the deep learning network HRNet, the input feature map has dimensions of N×C×H×W, where N is the batch size, C is the number of channels, and H and W are the spatial dimensions.

[0038] Beta distribution parameters in the ASR submodule β =1.0, cropping region area threshold bbx_thres=0.1, feature fusion weight λ=0.5, learnable balance factor The initial value is 0;

[0039] Tissue sensitivity weights in the MFACM submodule The value is 0.8, and the convolutional layer uses a 3×3 convolutional kernel.

[0040] Preferably, the scaling factor offset factor They are represented by the following formulas respectively:

[0041]

[0042] In the above formula, For 1×1 convolutional layer weights, , All are BatchNorm1d layers with stable statistical distributions. , These are the bias terms, It is the sigmoid activation function. This indicates element-wise multiplication, and s is the statistical vector obtained by concatenating the mean and standard deviation along the channel dimension and integrating the dissected regions.

[0043] This invention offers at least the following beneficial effects: For small-sample segmentation scenarios in medical images, this invention constructs a dual-module collaborative architecture of "Anatomical Perception Spatial Regularization (ASR) + Multi-Scale Feature Channel Modulation (MFACM)," which not only solves the statistical bias of small samples but also achieves a technical breakthrough of "medical anatomy prior guidance + spatial-channel dual-dimensional dynamic optimization." Its effects include:

[0044] Firstly, this invention adopts a "beta distribution anatomical clipping + batch statistical association" mechanism. The clipping ratio generated by the beta distribution is adapted to the common proportion of medical lesions. At the same time, the batch statistics are used to correct the statistical deviation of the anatomical region of the unit sample, thus solving the industry pain point of "insufficient extraction of anatomical features" under small sample conditions.

[0045] Secondly, this invention proposes a modulation factor generation logic of "spatial quality feedback - channel weight adaptation". The MFACM module receives the "anatomical region quality score" output by ASR and dynamically adjusts the gy / fy modulation factor to achieve dual-dimensional optimization of "spatial enhancement of lesions and channel optimization of texture features", which is different from the single-dimensional modulation of existing technologies.

[0046] Third, by embedding MARN into HRNet, multi-resolution feature extraction and medical feature optimization can be carried out simultaneously, forming a three-layer feature representation of "high-resolution features - anatomical features - statistically stable features".

[0047] Fourth, the module of this invention does not require reconstruction of the main model and can be directly embedded into the feature processing flow of HRNet at each stage, resulting in low development cost, easy promotion and application, and greater convenience.

[0048] Other advantages, objectives and features of the present invention will become apparent in part from the following description, and in part from those skilled in the art through study and practice of the invention. Attached Figure Description

[0049] Figure 1 This is a flowchart illustrating the processing of the basic block of the present invention;

[0050] Figure 2 This is a flowchart of the anatomical perception spatial regularization process of the present invention;

[0051] Figure 3This is a flowchart of the medical feature adaptive channel modulation process of the present invention;

[0052] Figure 4 A schematic diagram comparing the segmentation performance of the MARN module of this invention with existing mainstream segmentation models on the Kvasir-SEG dataset of intestinal polyps;

[0053] Figure 5 A schematic diagram comparing the segmentation performance of the MARN module of this invention with existing mainstream segmentation models on the intestinal polyp CVC-ClinicDB dataset;

[0054] Figure 6 A schematic diagram comparing the segmentation performance of the MARN module of this invention with existing mainstream segmentation models on the ISIC-2016 dataset of skin lesions;

[0055] Figure 7 A schematic diagram comparing the segmentation performance of the MARN module of this invention with existing mainstream segmentation models on the BIOHit tongue dataset;

[0056] Figure 8 A schematic diagram showing the comparison of the segmentation performance of the MARN module of this invention with existing mainstream segmentation models on the ISIC-2017 dataset of skin lesions. Detailed Implementation

[0057] The present invention will now be described in further detail with reference to the accompanying drawings, so that those skilled in the art can implement it based on the description.

[0058] This invention proposes a Medical Adaptive Regularization Network (MARN) module, which is integrated into a medical image segmentation model through a "space-channel" dual-dimensional collaborative optimization strategy. The core of this module is a dual-module collaborative mechanism, consisting of two sub-modules: Anatomical Perception Spatial Regularization (ASR) and Medical Feature Adaptive Channel Modulation (MFACM). In application, the ASR sub-module focuses on the spatial dimension, optimizing local feature extraction based on the anatomical structural characteristics of medical images; while the MFACM sub-module focuses on the channel dimension, dynamically adapting the feature distribution. These two modules complement each other, achieving full-dimensional feature regularization. Furthermore, the MARN module can be directly embedded into the feature processing flow of mainstream segmentation models such as HRNet without reconstructing the main model architecture, exhibiting seamless integration and strong adaptability.

[0059] To better explain the MARN module, the workflows of ASR and MFACM are explained below.

[0060] The work of Anatomical Perception Spatial Regularization (ASR) includes:

[0061] 1. Adaptive anatomical region cropping: A cropping ratio is generated using a beta distribution. Local regions in the feature map whose area exceeds 0.1 times the feature map area are randomly cropped, forcing the model to focus on key anatomical regions (lesions and boundaries). Existing technologies using fixed ratios or purely random cropping do not consider the anatomical pattern in medical images where "lesions typically occupy a small proportion of the image." This invention selects a cropping ratio generated by this beta distribution parameter that precisely covers the common proportion range of lesions.

[0062] 2. Channel-level statistical foundation: First, calculate the mean and standard deviation of each channel of the input feature map (dimensions N×C×H×W), and then introduce a small adjustment value ( To avoid numerical instability, this invention differs from batch-based overall statistics and instance-based single-sample channel statistics in batch normalization mechanisms. It focuses on the core characteristic of medical images—"anatomical correlations exist between channels (e.g., lesion areas often correspond to specific channel feature responses)"—and captures commonalities in cross-sample anatomical features through the channel mean / standard deviation of multiple batches of data. It is not a general, conventional setting, but rather the result of experiments verified by 5 different medical datasets (including skin lesions, intestinal polyps, etc.) to achieve the optimal balance between "avoiding numerical overflow" and "preserving channel statistical details", specifically adapted to the channel statistical fluctuation problem caused by low contrast and high noise in medical images;

[0063] 3. Multi-modal feature fusion: Features are fused according to the cropping mode (full image / region), and finally the enhanced features are weighted and fused with the original features (learnable parameters (also known as learnable balancing factors)). (Controlling the balance between global information and local details). Unlike the fixed weights or single fusion modes of existing technologies, the learnable parameters of this invention... This allows the model to autonomously adapt to different medical scenarios, such as in skin lesion datasets (with blurred lesion boundaries). The convergence value is 0.6-0.7 (enhancing local detail weights), and 0.4-0.5 in the intestinal polyp dataset (where the contrast between lesions and background is obvious) (balancing global and local aspects). The two modes provided are full-image and region, which are adapted to the anatomical morphology of "small lesions distributed throughout the image" and "large lesions clustered locally". This design is based on the inductive analysis of anatomical morphology of 5 types of medical images, rather than a simple overlay of modes.

[0064] The work of Medical Feature Adaptive Channel Modulation (MFACM) includes:

[0065] 1. Medical-Oriented Channel Attention Learning: This method learns channel attention weights for medical tissue features by using global average pooling, convolutional layers, and sigmoid activation to learn the output features of the anatomical perception space regularization submodule. Existing channel attention methods often employ "global average pooling + 1×1 convolution," focusing only on channel feature intensity and neglecting the functional correlation between channels in medical tissues. This invention uses 3×3 convolutional layers to effectively capture the correlation features between adjacent channels. Simultaneously, the global average pooling range is limited to the "effective region where lesions may exist" through ASR pruning, avoiding interference from background regions and specifically adapting to the channel co-expression characteristics of medical tissues.

[0066] 2. Organizational Sensitivity Weighted Fusion: The learned channel attention is weighted and fused with the average channel attention (organizational sensitivity). =0.8), enhancing the response to medical tissue features. Existing attention fusion methods mostly use equal weights or random weights, without considering the differences in the impact of different types of medical tissues on segmentation accuracy;

[0067] 3. Dynamic Statistical Modulation: Features are scaled and shifted at the channel level using adaptively generated statistical modulation factors, replacing the fixed statistics of traditional batch normalization mechanisms and dynamically adapting to the feature distribution of medical data. Unlike adaptive instance normalization, which generates modulation factors based solely on single-sample statistics, this invention uses modulation factors (modulation factors are also known as scaling factors). and offset factor The scaling factor integrates two dimensions of information: channel-level statistics from the ASR submodule and global statistics from batch data. Combining the channel mean fluctuation range (reflecting the stability of medical characteristics), the offset factor It combines the anatomical feature offset of batch data (reflecting the commonality of features across samples); and the modulation factor adopts independent channel-level calculation to address the characteristic of "large differences in the magnitude of features in different channels" in medical images, avoiding the suppression of key channel features caused by the traditional batch normalization mechanism of "single batch statistics covering all channels", and specifically solves the statistical bias problem in small sample medical scenarios.

[0068] The MARN module designed in this invention can solve the problems of unstable feature distribution and large statistical bias in traditional normalization mechanisms under medical scenarios with small samples, low contrast, and high noise; it makes up for the shortcomings of existing technologies in modeling the global correlation of anatomical structures in medical images, and enhances the model's ability to capture local details and boundaries of lesions; it enables the module to be universally adaptable to multiple medical image segmentation scenarios, and improves the robustness of the model under different lesion types and data quality.

[0069] Example:

[0070] In this embodiment, a network module is designed to optimize the feature normalization and anatomical structure modeling capabilities of medical image segmentation models. This module can be seamlessly integrated into mainstream medical image segmentation models such as HRNet, and is specifically adapted for accurate segmentation tasks of various types of medical images, such as skin lesions, intestinal polyps, and tongues. It has significant advantages, especially for medical scenarios with small samples, low contrast, and high noise.

[0071] In this embodiment, since the MARN module needs to be integrated into the feature processing flow of HRNet, the specific parameter settings are as follows: Input feature map dimension: N (batch size) = 6, C (number of channels) is 48 / 96 / 192 / 384 depending on the HRNet stage, H / W (spatial size) = 480×480 (after uniform scaling); ASR submodule: Beta distribution parameter β = 1.0, cropping region area threshold (bbx_thres) = 0.1, feature fusion weight λ = 0.5, initial value of learnable balance factor γ = 0.5; MFACM submodule's tissue sensitivity weight α = 0.8, the convolutional layer uses a 3×3 convolutional kernel, and the channel attention activation function is sigmoid.

[0072] The workflow for applying the MARN module to image segmentation is as follows:

[0073] Step 1: Medical Image Preprocessing

[0074] Medical images were uniformly scaled to 480×480 pixels. Bilinear interpolation was used to preserve texture, and nearest neighbor interpolation was used for labels to ensure the integer property of the categories. Mean subtraction standardization was performed using the BGR channel mean of the ImageNet dataset.

[0075] Step 2: HRNet Feature Extraction and MARN Embedding

[0076] The preprocessed image is input into HRNet, and features are extracted in a staged manner. The feature extraction process for each stage is as follows:

[0077] The first stage consists of four cascaded bottleneck blocks, each with a structure of "1×1 convolution dimensionality reduction → 3×3 convolution → 1×1 convolution dimensionality increase + residual connection", and 256 output channels.

[0078] The second stage includes one high-resolution module (HighResolutionModule), and the basic block of each module branch is embedded with MARN, outputting 48 / 96 multi-resolution feature channels; it should be noted that because the basic blocks of the second, third, and fourth stages in HRNet embed MARN (such as... Figure 1 As shown), therefore, taking the second stage as an example, the working method of each sub-module in MARN will be explained below:

[0079] (a) The ASR submodule processing flow is as follows: Figures 2-3 As shown, the main steps include:

[0080] S210, Adaptive cropping region generation, through... Generate cropping ratio Calculate the cutting width and height (cx, cy) are the randomly selected cropping center coordinates, and (bbx1, bby1, bbx2, bby2) represent the top left and bottom right corners of the cropping area, ensuring that the cropping area ratio exceeds the preset threshold bbx_thres (default 0.1).

[0081]

[0082] This step mainly involves adaptive spatial cropping and feature fusion of ASR, forcing the model to focus on local details of the lesions. On the BIOHit tongue image dataset (small sample size, low contrast), the HRNet integrated with MARN achieved an mIoU of 98.20% and an mDice of 99.09%, which are improvements of 0.97% and 0.5% respectively compared to the original HRNet. On the Kvasir-SEG intestinal polyp dataset, the mIoU and mDice were improved by 9.2% and 5.94% respectively.

[0083] The adaptive cropping mode adaptation logic in S210 is based on the anatomical characteristics of medical images, such as "diverse lesion morphologies and large differences in background proportions." After the adaptive cropping region is generated, this invention achieves accurate region selection through the following four types of cropping modes. The correspondence between each mode and the parameter crop, as well as their applicable scenarios, are as follows:

[0084] 1. Reference cropping: using reference features Global statistical distribution Based on the selection criteria, within the cropping range generated by S210, only those items that match... Distribute matching anatomical regions (such as regions with statistical characteristics consistent with similar lesions) and remove background and noise interference regions.

[0085] 2. Content cropping: based on original features x The selection criteria are based on local content features (such as lesion texture and brightness gradient), retaining local response values ​​above a threshold within the cropping range (the default is the channel mean). The area (1.2 times larger than the original sample) is used to focus on the specific lesions in the current sample.

[0086] 3. Dual cropping: Reference cropping and content cropping are performed simultaneously. First, reference cropping is used to select regions that meet the commonalities of global anatomy. Then, content cropping is used to further retain the local high information density regions within the selected regions. Finally, the ROI after dual filtering of "global commonality + local specificity" is output.

[0087] 4. No cropping: Disables the region filtering logic, does not perform the above cropping operation, and directly uses the original features. x The entire image is used as the processing object (crop='neither'); or only features are style-aligned (crop='style', based on...). The style statistics correct the brightness and contrast distribution of the original features, but do not change the spatial range.

[0088] All cropping modes ultimately generate local features that focus on the key anatomical regions. The model includes a corresponding mask (mask=1 represents the ROI region, mask=0 represents the non-ROI region), providing clean, background-free input for subsequent medical feature extraction. The model has fixed parameters for double pruning, which can be switched according to the data type.

[0089] S211. Medical Feature Statistical Extraction: Calculate the input original feature map. x (Dimensions are N×C×H×W, where N is the batch size, C is the number of channels, and H / W is the spatial size) Mean value over each channel c and standard deviation Introducing minute values ​​into the standard deviation. (Default 10) −5 This avoids numerical instability and provides a statistical basis for subsequent adaptive feature fusion.

[0090]

[0091]

[0092] In the above formula, N is the batch size. c H represents the number of channels, and H and W represent the spatial dimensions. Original feature map x Pixel-level feature values ​​for a specific sample, a specific channel, and a specific spatial location; To avoid small, unstable adjustment values, and .

[0093] These channel-level statistics provide a statistical basis for subsequent adaptive feature fusion, enabling the network to dynamically normalize and modulate according to the importance of each channel, thereby enhancing robustness and preserving lesion-related details in medical image segmentation.

[0094] S212, Adaptive Feature Fusion, Original Features x First, by subtracting its mean... Divide by its standard deviation Standardization was performed, followed by the use of reference characteristic statistics ( For reference average, The standard deviation is used as a reference for renormalization, and the formula is:

[0095]

[0096] In the above formula, This represents the normalized transformation feature, which effectively improves the consistency of cross-sample or cross-regional features in medical images by keeping the distribution of content features consistent with that of reference features.

[0097] S213, Medical Adaptive Regularization, mainly generates enhanced features based on pruning patterns. :

[0098]

[0099] In the above formula, Blend ( ) represents the feature fusion function. As a reference feature, it is a stable benchmark feature generated based on global statistical information from similar medical datasets. For the original features x The transformation features after normalization, and , , For each channel c The mean and standard deviation obtained after performing medical characteristic statistics , These are the reference mean and reference standard deviation, respectively. mask For the mask, This refers to the local features of the key anatomical regions after adaptive cropping.

[0100] 1) Global fusion: Directly apply the feature fusion function Blend() to the entire image, where, x Indicates original features, These are reference features used for feature fusion.

[0101] 2) Region fusion: The mask is used to fuse only the cropped Region of Interest (ROI) while preserving the original features of the non-ROI regions.

[0102] Finally, output Original features x With enhanced features The weighted fusion (with weight λ controlling the regularization strength) is shown in the following equation:

[0103]

[0104] In the above formula, λ is the feature fusion weight that controls the regularization strength. For the original features x Features after standardization and normalization, original features x Refers to the initial input characteristics received by the ASR submodule. This refers to the enhanced features generated after global fusion of the ASR submodules.

[0105] (II) The processing flow of the MFACM submodule is as follows: Figure 3 As shown, the main steps include:

[0106] S220. First, fuse the output of the anatomical perception spatial regularization with the original feature x, as shown in the following formula:

[0107]

[0108] In the above formula, To control the learnable parameters that balance global information and local details, To learnable parameters The value of is compressed and mapped to the interval [0,1], which is a commonly used activation function. The original features x Preserve the original high-resolution global information of HRNet;

[0109] S221, Channel Attention Learning, for Global average pooling (AvgPool) is performed, followed by two convolutions (3×3) and the ReLU activation function and sigmoid function. Channel attention is generated by calculating global average pooling features:

[0110]

[0111] S222, Attention Medicine Enhancement, Average attention of the channel Weighted fusion, (Default value 0.8) is the control and The formula for the fusion of organizational sensitivity weights is:

[0112]

[0113] S223, Feature channel modulation, through statistical modulation. and For feature scaling offset, For 1×1 convolutional layer weights, , It is a BatchNorm1d layer (stable statistical distribution). , These are the bias terms, It is the sigmoid activation function. This indicates element-wise multiplication. This represents the enhanced channel attention in S222. s is the statistical vector of the integrated anatomical region, concatenating the mean and standard deviation along the channel dimension.

[0114] scaling factor offset factor They are represented by the following formulas respectively:

[0115]

[0116] Final modulation output:

[0117]

[0118] In the above formula, for The channel statistical mean.

[0119] During each forward propagation, scaling factors and offset weights are calculated online based on the statistics of the current features (mean and standard deviation per sample and per channel), and these weights change with random pruning and instance pairing. MFACM completely solves the statistical bias problem in small sample scenarios by replacing the fixed statistics of the batch normalization mechanism with dynamically generated scaling factors and offset weights. On the CVC-ClinicDB dataset, the validation loss of HRNet integrated with MARN is reduced by about 50% compared with the original HRNet, and the training convergence speed is faster, reaching the loss level of the original HRNet in 60th rounds by the 30th round.

[0120] The third stage includes four high-resolution modules (HighResolutionModule), and the basic block of each module branch is embedded in MARN, with output channels of 48 / 96 / 192.

[0121] The fourth stage includes three high-resolution modules (HighResolutionModule), each with a basic block embedded in MARN, and output channels of 48 / 96 / 192 / 384.

[0122] Step 3: Split the output of MARN. The input HRNet task head is used to generate medical image segmentation results through convolution and activation functions.

[0123] The adaptive spatial cropping and feature fusion of this invention's ASR forces the model to focus on local details of lesions. On the BIOHit tongue image dataset (small sample, low contrast), the HRNet integrated with MARN achieves an mIoU of 98.20% and an mDice of 99.09%, which are improvements of 0.97% and 0.5% respectively compared to the original HRNet. On the Kvasir-SEG intestinal polyp dataset, the mIoU and mDice are improved by 9.2% and 5.94% respectively.

[0124] Furthermore, to illustrate the universality advantage of this invention across multiple scenarios, the method of this invention, along with existing mainstream methods such as U-Net, UNet++, and Deeplabv3+, was applied to five datasets: skin lesions (ISIC-2016, ISIC-2017), intestinal polyps (CVC-ClinicDB, Kvasir-SEG), and tongue (BIOHit). The results are shown in Tables 1 and 2. Figures 4-8 It can be seen that the segmentation performance of the model integrating MARN is better than that of existing mainstream methods such as U-Net, UNet++, and Deeplabv3+, and its adaptability and robustness are at the leading level.

[0125] Table 1: Comparison of segmentation performance of the present invention and existing technologies on the BIOHit (300), CVC-ClinicDB (612), and ISIC-2017 (2150) three-phase datasets.

[0126]

[0127] Table 2: Comparison of segmentation performance of the present invention and existing technologies on the Kvasir-SEG and ISIC-2016 three-phase datasets.

[0128]

[0129] As can be seen from the above description, the present invention is fundamentally different from existing normalization techniques and related patents, as described below:

[0130] I. Compared with Batch Normalization (BN) mechanism: It breaks through the limitations of existing BN-type "static statistics + global processing" and deeply binds dynamic modulation with medical anatomical features; existing dynamic weights only focus on multi-scale weight allocation, while this invention focuses on the root cause of statistical bias in small samples and achieves statistical self-adaptation through modulation factor generation; it is the first to create a collaborative mechanism of spatial anatomical regularization and channel dynamic modulation, filling the gap of BN-type technology in medical spatial feature optimization.

[0131] II. Compared with Instance Normalization (IN) / Adaptive Instance Normalization (AdaIN): Unlike the limitation of IN / AdaIN being "single-sample dominant", this invention solves the problem of small sample parameter estimation through batch statistical correlation; existing medical image segmentation related technologies only focus on morphological boundaries, while this invention delves into the statistical characteristic modeling of anatomical regions; it breaks the industry convention of "independent optimization of space and channel" and achieves two-dimensional collaboration under the guidance of anatomical prior.

[0132] Third, compared with CrossNorm: it breaks through the limitations of CrossNorm-like "general statistical mixing" and constructs a medical-specific statistical optimization logic; existing multimodal fusion only focuses on the fusion of segmentation results, while this invention optimizes statistics from the source of feature extraction through anatomical priors; it achieves "content-aware statistical modulation", which is different from the "indiscriminate statistical mixing" of existing technologies and is suitable for various medical scenarios such as skin lesions and intestinal polyps.

[0133] The above solution is merely an illustration of a preferred example and is not limited thereto. When implementing this invention, appropriate substitutions and / or modifications can be made according to the user's needs.

[0134] Although embodiments of the present invention have been disclosed above, they are not limited to the applications listed in the specification and embodiments. It can be applied to various fields suitable for the present invention. Other modifications can be readily made by those skilled in the art. Therefore, without departing from the general concept defined by the claims and their equivalents, the present invention is not limited to the specific details and examples shown and described herein.

Claims

1. A medical image segmentation method based on an adaptive regularization network module, characterized in that, include: S1. Construct an adaptive regularized network module MARN, which includes: an anatomical perception spatial regularization (ASR) submodule focusing on the spatial dimension and a multi-scale channel modulation (MFACM) submodule focusing on the channel dimension. S2. After processing the preprocessed medical image in the ASR submodule, the weighted fused regularized features are output. : In the above formula, λ is the feature fusion weight that controls the regularization strength. For the original features x Features after standardization and normalization, original features x Refers to the initial input characteristics received by the ASR submodule. These are the enhanced features generated after global fusion of the ASR submodules; S3. Regularize the features using the following formula. With original features x By fusion, the quality scoring characteristics of the anatomical region are obtained. : In the above formula, To control the learnable balance factor between global information and local details, It is a range mapping activation function; S4, will The input is modulated in the MFACM submodule to obtain the modulation output characteristics using the following formula. : In the above formula, Scaling factor As the offset factor, To The feature values ​​after channel attention enhancement for The channel statistical mean; S5, will The task head is fed into the deep learning network HRNet to generate segmentation results of medical images through convolution and activation functions; In S2, enhanced features Obtained through the following formula: In the above formula, Blend ( ) represents the feature fusion function. It is a stable benchmark feature generated based on global statistical information from similar medical datasets. For the original features x The transformation features after normalization, and , , For each channel c The mean and standard deviation obtained after performing medical characteristic statistics , These are the reference mean and reference standard deviation, respectively. mask For the mask, This is to focus on the local features of key anatomical regions after adaptive cropping; scaling factor offset factor They are represented by the following formulas respectively: In the above formula, For 1×1 convolutional layer weights, , All are BatchNorm1d layers with stable statistical distributions. , These are the bias terms, It is the sigmoid activation function. This indicates element-wise multiplication, and s is the statistical vector obtained by concatenating the mean and standard deviation along the channel dimension and integrating the dissected regions.

2. The medical image segmentation method based on an adaptive regularization network module as described in claim 1, characterized in that, Each channel c mean It is characterized by the following formula: In the above formula, N is the batch size. c H represents the number of channels, and H and W represent the spatial dimensions. Original feature map x Pixel-level feature values ​​for a specific sample, a specific channel, and a specific spatial location; Each channel c standard deviation It is characterized by the following formula: In the above formula, To avoid small, unstable adjustment values, and .

3. The medical image segmentation method based on an adaptive regularization network module as described in claim 1, characterized in that, In S4, Obtained through the following formula: In the above formula, For the average attention of the channels, For channel attention, and , Indicates to Perform global average pooling. ReLU represents the convolution operation, and ReLU() represents the activation function. For channel attention, use the sigmoid activation function. To control and Organizational sensitivity weighting for integration.

4. The medical image segmentation method based on an adaptive regularization network module as described in claim 1, characterized in that, The deep learning network HRNet includes the following four-stage feature extraction process: The first stage consists of four cascaded bottleneck blocks, each of which includes: a 1×1 convolutional dimensionality reduction layer, a 3×3 convolutional layer, a 1×1 convolutional dimensionality increase layer, and a residual connection layer. The second phase includes one high-resolution module; Phase 3: Contains 4 high-resolution modules; Phase 4: Contains 3 high-resolution modules; Each high-resolution module branch's base block is embedded with MARN; In the deep learning network HRNet, the input feature map has dimensions of N×C×H×W, where N is the batch size, C is the number of channels, and H and W are the spatial dimensions. Beta distribution parameters in the ASR submodule β =1.0, cropping region area threshold bbx_thres=0.1, feature fusion weight λ=0.5, learnable balance factor The initial value is 0; Tissue sensitivity weights in the MFACM submodule The value is 0.8, and the convolutional layer uses a 3×3 convolutional kernel.

Images