CN-122023810-A - Brain tumor MRI image segmentation method based on soft clustering KAN network

CN122023810ACN 122023810 ACN122023810 ACN 122023810ACN-122023810-A

Abstract

The invention discloses a brain tumor MRI image segmentation method based on a soft clustering KAN network, which belongs to the technical field of medical information processing, wherein voxel alignment and intensity normalization are carried out on a multi-mode three-dimensional brain tumor MRI image to construct a unified input sample, hierarchical downsampling and multi-scale feature extraction are realized through a KAN residual error encoder which is formed by alternately cascading a double-layer KAN attention sub-network and a three-dimensional maximum pooling layer, three-dimensional position codes are added to the highest layer features, bottleneck features containing global structure priori are obtained through modeling of a soft K mean value clustering bottleneck layer, then decoding is completed by fusing jump connection features and up sampling features through an attention gating mechanism, and finally, a tumor whole, a tumor core and a multi-subregion three-dimensional segmentation result of an enhanced tumor are output through fusion of a main segmentation branch and a refined branch through a soft clustering regularization and multi-scale depth supervision constraint optimization model.

Inventors

LIU JIAN
MA XINGYU
ZHAO KAI
WANG XUESONG
CHENG YUHU

Assignees

中国矿业大学

Dates

Publication Date: 20260512
Application Date: 20260410

Claims (10)

1. The brain tumor MRI image segmentation method based on the soft clustering KAN network is characterized in that the soft clustering KAN network comprises a KAN residual error encoder, a soft K mean value clustering bottleneck layer and a KAN decoder, and comprises the following steps: Step 1, acquiring a multi-mode three-dimensional brain tumor MRI image, and constructing a multi-mode three-dimensional input sample with aligned voxel levels and uniform intensity distribution based on the multi-mode three-dimensional brain tumor MRI image; Step 2, inputting the multi-mode three-dimensional input sample into a KAN residual error encoder which is constructed based on KAN residual error blocks in advance, wherein the KAN residual error encoder is formed by alternately cascading a double-layer KAN attention sub-network and a three-dimensional maximum pooling layer and is used for executing layer-by-layer downsampling and feature extraction to obtain multi-scale coding features; Step 3, introducing three-dimensional position codes into a final-stage feature map in the multi-scale coding features to supplement spatial position information, and performing soft clustering modeling on the multi-scale coding features by utilizing the soft K mean clustering bottleneck layer to generate a bottleneck feature map containing global structure priori; Inputting the bottleneck characteristic diagram subjected to soft cluster modeling into the KAN decoder, guiding each layer of jump connection characteristics generated by the KAN residual error encoder by using an attention gating mechanism, and performing cascade reconstruction with the upsampling characteristics of the KAN decoder to obtain multi-scale decoding characteristics for recovering resolution step by step; And 5, inputting the first-stage final decoding feature of the multi-scale decoding feature into a multi-category main segmentation branch and an enhanced tumor refinement branch, respectively outputting a main segmentation result and a fine-granularity prediction result aiming at an enhanced tumor region, and carrying out feature fusion on the fine-granularity prediction result and the main segmentation result to obtain a multi-subregion three-dimensional segmentation result covering the whole tumor, the tumor core and the enhanced tumor.
2. The brain tumor MRI image segmentation method based on soft clustering KAN network according to claim 1, further comprising a multi-scale supervision and clustering regular constraint step, wherein in the training process of the soft clustering KAN network, a soft clustering regularization head and a multi-scale depth supervision head are arranged on the final-stage feature map and the middle-stage feature map of the multi-scale decoding feature so as to constrain the decoding feature distribution and the multi-scale segmentation result.
3. The brain tumor MRI image segmentation method based on the soft clustering KAN network according to claim 1, wherein constructing a multi-modal three-dimensional input sample with aligned voxel level and uniform intensity distribution specifically comprises: Performing artifact removal and spatial resampling on the acquired original three-dimensional brain tumor MRI image of each mode, and unifying voxel spacing to a preset three-dimensional spatial resolution; Carrying out rigid or affine registration on each mode three-dimensional brain tumor MRI image to enable the MRI images to realize voxel level alignment under the same space coordinate system; And performing skull stripping and effective visual field cutting to remove non-brain tissue areas, and cutting and normalizing the gray scales of all modes according to a preset percentile threshold value so as to unify the intensity range to a standardized interval.
4. The brain tumor MRI image segmentation method based on the soft clustering KAN network according to claim 1, wherein the multi-modality three-dimensional brain tumor MRI image comprises four modality MRI images of T1 weighted imaging, contrast enhanced T1 weighted imaging, T2 weighted imaging and fluid attenuation inversion recovery imaging.
5. The brain tumor MRI image segmentation method based on the soft clustering KAN network according to claim 1, wherein a KAN residual encoder constructed based on KAN residual blocks comprises five encoding stages, each of which performs an operator through an efficient three-dimensional KAN residual module Operation, operator The operation specifically comprises the following steps: Step 2.1, determining input characteristics of the efficient three-dimensional KAN residual error module in the current encoding stage, establishing a residual error branch in the efficient three-dimensional KAN residual error module, adjusting dimensionality through convolution mapping if the number of input channels is inconsistent with that of output channels, otherwise, performing identity mapping to obtain residual error characteristics and reserving the residual error characteristics; Step 2.2, on the main branch of the high-efficiency three-dimensional KAN residual error module, mapping the input features to a reduced channel space by utilizing a three-dimensional convolution layer, a batch normalization layer and a ReLU activation function to obtain dimension reduction features; Step 2.3, inputting the global context feature vector into an attention subnet built in the efficient three-dimensional KAN residual error module, performing nonlinear transformation and normalization processing on feature elements based on Gaussian radial basis functions by utilizing two KAN linear layers, and outputting a channel attention weight vector for adjusting channel importance by a Sigmoid activation function; step 2.4, carrying out channel-by-channel dot product recalibration on the dimension-reducing feature by utilizing the channel attention weight vector, and then carrying out processing by a three-dimensional convolution layer and a batch normalization layer to obtain a dimension-increasing enhancement feature; The KAN residual encoder obtains the multi-scale coding feature through step-by-step iteration of five coding stages.
6. The brain tumor MRI image segmentation method based on the soft clustering KAN network according to claim 1, wherein the step 3 specifically comprises: step 3.1, receiving a final-stage feature image in the multi-scale coding features as input, constructing a three-dimensional position code generator, and generating a normalized coordinate grid consistent with the spatial dimension of the final-stage feature image; Physically splicing the final-stage feature map and the normalized coordinate grid in the channel dimension, and carrying out feature fusion by utilizing a three-dimensional position fusion convolution layer to output a positioning feature map with explicit space position information; Step 3.2, carrying out nonlinear mapping on the positioning feature map by utilizing a feature extraction subnet to obtain a potential feature map, and flattening the potential feature map into a voxel feature sequence; Step 3.3, converting the Euclidean distance into soft distribution probability by adopting a Softmax function with temperature parameters so as to generate a soft cluster distribution diagram consistent with the input space size; and 3.4, splicing the soft cluster distribution diagram and the positioning feature diagram on the channel dimension to obtain a cluster enhancement combination feature, and mapping the cluster enhancement combination feature back to the target channel dimension by using a cluster fusion convolution layer to output a final bottleneck feature diagram.
7. The brain tumor MRI image segmentation method based on the soft clustering KAN network according to claim 1, wherein the step 4 specifically comprises: Step 4.1, the KAN decoder is composed of four cascaded decoding stages, receives the bottleneck characteristic map as initial input, introduces coding characteristics generated by corresponding levels as jump connection characteristics for each decoding stage, and performs significance screening on the jump connection characteristics by using an attention-gated subnet; step 4.2, in the attention gating sub-network, taking the upper-level decoding characteristic as a gating query signal, mapping the gating query signal and the jump connection characteristic by utilizing a three-dimensional convolution layer respectively, calculating the correlation of the gating query signal and the jump connection characteristic after element-by-element addition and nonlinear processing, mapping the correlation into an attention coefficient matrix, carrying out element-by-element multiplication operation with the original jump connection characteristic, and outputting the jump connection characteristic subjected to saliency weighting; Step 4.3, performing up-sampling operation on the upper-stage decoding feature by using a tri-linear interpolation algorithm to enlarge the spatial resolution and obtain an up-sampling feature, and splicing the up-sampling feature and the saliency weighted jump connection feature in the channel dimension to obtain a decoding combination feature containing multi-scale information; step 4.4, inputting the decoding combination characteristics to an efficient three-dimensional KAN residual error module, and utilizing operators And performing nonlinear mapping and feature reconstruction, outputting the decoding features of the current level and transmitting to the next decoding stage until the first-level final decoding features restored to the original image resolution are output.
8. The brain tumor MRI image segmentation method based on a soft clustering KAN network according to claim 4, characterized in that said method further comprises a step of performing parameter updating on said soft clustering KAN network by means of a joint loss function, said joint loss function being constructed jointly by a main segmentation result and a fine-grained prediction result, said joint loss function being composed in particular of a Dice loss, a Focal loss, a boundary loss, an enhanced tumor auxiliary loss, and a clustering regularization and depth supervision loss weighting.
9. The brain tumor MRI image segmentation method based on the soft clustering KAN network according to claim 2, wherein the multi-scale supervision and clustering regular constraint step specifically comprises the following steps: Selecting middle-level features in the multi-scale decoding features as auxiliary supervision objects, mapping the channel number of the middle-level features into target class number through a three-dimensional convolution layer to obtain an auxiliary prediction graph correspondingly, and up-sampling the spatial dimension of the auxiliary prediction graph to be consistent with multi-mode three-dimensional input samples by utilizing a tri-linear interpolation algorithm to generate multi-scale auxiliary probability distribution; step 6.2, a decoding cluster regularization head is arranged outside the main segmentation head in parallel, a feature projection sub-network is utilized to carry out nonlinear mapping on the first-stage final decoding feature, and projection features are output; And 6.3, defining joint clustering loss consisting of entropy regular terms, cluster balance regular terms and center separation regular terms based on the soft allocation probability, and optimizing decoding characteristic distribution in an unsupervised mode.
10. The brain tumor MRI image segmentation method based on the soft clustering KAN network according to claim 1, wherein the specific process of the step 5 is as follows: step 5.1, constructing a main segmentation prediction head consisting of a three-dimensional convolution layer, receiving a first-stage final decoding feature in the multi-scale decoding features, and mapping the first-stage final decoding feature to a target semantic class space to obtain a main segmentation logarithmic probability map; Step 5.2, constructing special enhanced tumor refinement branches on the first-stage final decoding features in parallel, extracting feature responses of enhanced tumors by utilizing three-dimensional convolution two-class prediction heads to obtain a single-channel enhanced tumor logarithmic probability map; And 5.3, taking the enhanced tumor refinement probability map as an independent focusing channel, carrying out confidence calibration and boundary correction on the corresponding enhanced tumor category in the main segmentation probability distribution map, and outputting the main segmentation probability distribution map and the enhanced tumor refinement probability map by a model at the same time so as to jointly form a brain tumor three-dimensional segmentation result.

Description

Brain tumor MRI image segmentation method based on soft clustering KAN network Technical Field The invention belongs to the technical field of medical information processing, and particularly relates to a brain tumor MRI image segmentation method based on a soft clustering KAN network. Background Brain tumors are a group of diseases formed by abnormal cell proliferation in the brain or central nervous system, and the global incidence rate of the brain tumors is about 7-11 cases per 10 ten thousand people per year, and the brain tumors have the characteristics of high mortality and disability rate. Brain tumors can be divided into two major categories, namely primary and metastatic, wherein primary malignant brain tumors with more aggressiveness, such as glioblastomas, are not very high in annual incidence, but have extremely poor prognosis, the relative five-year survival rate of patients is less than 10%, the median survival period is only 12-15 months, compared with brain metastases, which are more common in practice, have incidence of 10 times that of primary brain tumors, and about 20% -40% of systemic cancer patients can undergo brain metastases, greatly affecting the survival period and quality of life of patients. Accurate diagnosis and treatment planning of brain tumors is highly dependent on multiparameter magnetic resonance imaging techniques (Magnetic Resonance Imaging, MRI) that reveal the internal heterogeneity of the tumor from multiple physiological levels through different pulse sequences. The development of the method is not limited to the improvement of field intensity, and the more central breakthrough is sequence innovation and information mining, on one hand, the quantitative evaluation of tumor cell density, microcirculation perfusion and micro-bleeding is realized through functional sequences such as diffusion weighted imaging, perfusion weighted imaging and magnetic sensitivity weighted imaging, and on the other hand, the deep features which cannot be recognized by human eyes can be extracted from massive images based on artificial intelligence image histology and are fused with clinical and genetic data to construct a prediction model. At present, a tumor segmentation model based on deep learning mostly uses a multi-layer perceptron as a bottom layer framework, is one of the most basic and most classical neural network architectures in the field of deep learning, and a segmentation method based on a traditional convolutional neural network has been remarkably developed, wherein a U-Net network and various varieties thereof are excellent in the field of tumor segmentation due to extremely strong expandability and high efficiency on small sample data. However, with the research of deep learning, the limitation of the traditional multi-layer perceptron, namely parameter redundancy, poor interpretability is difficult to embed deep exploration of physical priors in specific tasks, the inherent 'black box' characteristic makes it difficult to understand the deep basis of model decisions, and the reliability and the robustness are often questioned especially when facing tumor cases with fuzzy boundaries and complex textures. Disclosure of Invention The application aims to provide a brain tumor MRI image segmentation method based on a soft clustering KAN network to enhance the capturing capability of micro focus aiming at the technical problems of complex tumor boundary blurring and complex texture and difficult accurate positioning at present. In order to achieve the technical purpose, the following technical scheme is adopted in the embodiment of the application. The embodiment provides a brain tumor MRI image segmentation method based on a soft clustering KAN network, wherein the soft clustering KAN network comprises a KAN residual error encoder, a soft K mean value clustering bottleneck layer and a KAN decoder, and the method comprises the following steps: Step 1, acquiring a multi-mode three-dimensional brain tumor MRI image, and constructing a multi-mode three-dimensional input sample with aligned voxel levels and uniform intensity distribution based on the multi-mode three-dimensional brain tumor MRI image; Step 2, inputting the multi-mode three-dimensional input sample into a KAN residual error encoder which is constructed on the basis of KAN residual error blocks in advance, wherein the KAN residual error encoder is formed by alternately cascading a double-layer KAN attention sub-network and a three-dimensional maximum pooling layer and is used for executing layer-by-layer downsampling and feature extraction to obtain multi-scale coding features; Step 3, introducing three-dimensional position codes into a final-stage feature map in the multi-scale coding features to supplement spatial position information, and performing soft clustering modeling on the multi-scale coding features by utilizing the soft K mean clustering bottleneck layer to generate a bottleneck feature map containin