A chessboard pixel splitting-based lightweight brain tumor MRI image classification method

Abstract

The application provides a chessboard pixel splitting-based lightweight brain tumor MRI image classification method, and the method is combined with a structure of chessboard pixel splitting shallow feature extraction, progressive refinement coding, accumulation, and gated skip connection and gated multi-scale classification, aiming to improve technical defects of a traditional classification model, such as a large number of parameters, loss of tumor edge details caused by conventional down-sampling, and insufficient recognition accuracy of rare metastatic tumors. In the experimental stage, the optimized network achieves excellent experimental results on the BrainMRI-4Class data set, the classification accuracy reaches 93.66%, the model parameter quantity is only 8.85M, and the calculation quantity is 1.09GFLOPs, compared with the traditional deep network, the recognition accuracy and lightweight performance are improved. This innovative work provides a more accurate and low-computing-power MRI image classification tool for clinical research and diagnosis of brain tumors, and helps to meet the needs of primary medical institutions for lightweight intelligent image analysis technology.

Description

Technical Field

[0001] This invention belongs to the field of medical imaging computer vision and deep learning technology, specifically involving intelligent classification technology for brain tumors in multi-sequence brain MRI images. It is applicable to the early screening of brain tumors, gliomas, meningiomas, metastatic tumors and normal brain tissue in primary medical institutions. Background Technology

[0002] Precise classification of brain tumors is crucial for clinical treatment planning and prognostic assessment. Multi-sequence MRI of the brain, due to its high soft tissue resolution and rich information dimensions, is the primary imaging basis for the clinical diagnosis of brain tumors. Traditional diagnostic methods rely on manual interpretation of images by radiologists, making diagnostic results susceptible to subjective experience. Furthermore, limited imaging resources in primary healthcare institutions make it difficult to meet the clinical needs of large-scale, rapid screening. Therefore, intelligent classification-assisted technologies are urgently needed.

[0003] Existing deep learning methods for brain tumor classification are mainly divided into two categories: heavy-duty models and lightweight models, both of which have significant technical shortcomings. Heavy-duty hybrid networks using 3D convolutions and Transformers achieve high classification accuracy, but have an extremely large number of model parameters and computational overhead, resulting in high hardware dependence and difficulty in achieving lightweight deployment at edge devices. Conventional lightweight networks often use max pooling and large stride convolutions for downsampling, which easily loses high-frequency pathological details such as tumor edges and small metastases, and perform poorly in identifying metastatic tumors with scarce samples.

[0004] Meanwhile, clinical brain tumor datasets commonly suffer from class imbalance. Existing techniques often rely on additional strategies such as weighted loss, label smoothing, and complex data augmentation to optimize models, increasing training difficulty and deployment costs. Currently, there is a lack of lightweight classification schemes that balance low parameter count, low computational cost, native adaptability to imbalanced clinical data, and effective preservation of lesion details. To address these technical shortcomings, this invention proposes a lightweight brain tumor MRI image classification method based on checkerboard pixel decomposition. Summary of the Invention

[0005] To address the shortcomings of existing intelligent classification technologies for brain tumor MRI, such as the difficulty in balancing lightweight models with accuracy, the tendency of traditional downsampling methods to lose lesion details, weak recognition ability of rare tumors, and reliance on additional regularization strategies to alleviate data imbalance, this invention aims to provide a lightweight brain tumor MRI image classification method based on checkerboard pixel decomposition. This invention utilizes a newly designed lossless lightweight feature extraction architecture and adaptive feature fusion, along with a multi-scale balanced classification mechanism. This significantly reduces the number of model parameters and computational overhead while preserving the pathological features of minute lesions, natively adapting to the imbalanced clinical sample distribution, and effectively improving the classification accuracy and robustness of various types of brain tumors.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] This invention achieves accurate four-class classification of normal brain tissue, glioma, meningioma, and metastatic tumors through a complete closed-loop process, including standardized preprocessing of multi-sequence brain MRI images, shallow feature extraction by chessboard pixel splitting, deep feature extraction by three-level progressive refinement coding, cumulative and gated jump feature fusion, and gated multi-scale adaptive classification.

[0008] This invention first performs standardized preprocessing on the acquired raw multi-sequence brain MRI images, resamples all images to a fixed size, and fills in the missing channels in clinical samples to achieve uniformity of the channel dimension of the dataset. It then standardizes the grayscale distribution through pixel normalization and combines it with multi-angle and multi-morphological data augmentation strategies to enrich the diversity of training samples and improve the generalization ability of the model.

[0009] This invention innovatively adopts a chessboard pixel splitting feature extraction structure, which completes lossless splitting based on image parity index and generates two complementary subgraph structures. It achieves downsampling without information loss without introducing additional learning parameters. It completes feature dimensionality enhancement and shallow detail extraction through dual-branch lightweight convolution, effectively preserving high-frequency pathological information such as tumor edges and micrometastases, and solving the problem of detail loss in traditional pooling downsampling.

[0010] Furthermore, this invention constructs a three-level cascaded progressive refinement encoder, which achieves efficient downsampling through a lightweight compression extension structure. It combines a dual attention mechanism of channels and space to adaptively enhance key lesion features and suppress invalid background interference. With the addition of a multi-layer convolutional nonlinear fusion and residual bypass dimension adaptation structure, the semantic feature expression is deepened step by step, enhancing the model's ability to differentiate between different tumor pathological features.

[0011] Meanwhile, this invention designs a gated skip connection fusion mechanism based on accumulation and optimization. Through scale alignment, channel matching, adaptive gate weight generation and row and column accumulation smoothing optimization, it achieves accurate fusion of shallow detailed features and deep semantic features, further compensating for the detail loss caused by deep encoding and improving the feature representation capability of small lesions.

[0012] Finally, this invention constructs a gated multi-scale classification structure, which captures multi-scale lesion features through multi-dilation rate dilated convolution, and adaptively fuses information of different scales by combining a dynamic gated weighting mechanism. It achieves class balance by relying on the network's endogenous structure, and can adapt to the imbalanced distribution of clinical data without relying on additional loss functions, label smoothing and other optimization methods. Ultimately, it achieves high-precision and highly robust four-class classification output of brain tumors.

[0013] Technical effect

[0014] 1. This invention proposes a lightweight MRI classification architecture for brain tumors that integrates chessboard pixel decomposition, progressive coding enhancement and accumulation, and gated multi-scale classification. Relying on a novel non-destructive feature extraction and adaptive balanced classification system, it breaks through the technical bottlenecks of traditional lightweight networks that are prone to losing lesion details during downsampling and heavy models that have high computational overhead. This invention achieves parameter-free and lossless downsampling through a chessboard-like odd-even pixel decomposition mechanism, fully preserving high-frequency pathological details of tumor margins and micrometastases. It further deepens the semantic features of lesions using a three-level channel-space dual-attention coding structure and employs a row-column accumulation and optimized gated skip connection mechanism to achieve high-precision fusion of deep and shallow features, significantly improving the model's ability to identify micro and rare tumor lesions. Simultaneously, it achieves endogenous class balance through a built-in gated multi-scale dynamic weighted classification structure, adapting to imbalanced clinical datasets without relying on weighted loss, label smoothing, or additional data augmentation. While significantly reducing model parameters and computational overhead and simplifying training and deployment processes, it effectively balances lightweight performance with multi-class brain tumor classification accuracy. The model boasts high prediction confidence and precise lesion focusing, aligning with clinical image interpretation and diagnostic logic. It effectively addresses the industry pain points of existing technologies, such as the inability to balance lightweight design with diagnostic accuracy, weak identification of rare tumors, and high implementation costs, demonstrating excellent generalization ability and high clinical applicability and translational value. Attached Figure Description

[0015] Figure 1 This is a flowchart of the overall network architecture of the present invention;

[0016] Figure 2 This is a schematic diagram of the chessboard pixel splitting feature extraction structure of the present invention;

[0017] Figure 3 This is a structural diagram of the single-stage progressive refinement encoder module of the present invention;

[0018] Figure 4 This is a schematic diagram of the accumulation and gated jump connection structure of the present invention;

[0019] Figure 5 This is a block diagram of the internal structure of the gated multi-scale classifier of the present invention;

[0020] Figure 6 This is an example image of four classes of samples from the BrainMRI-4Class dataset of this invention;

[0021] Figure 7 This is a graph showing the ROC performance of the present invention and the comparative model;

[0022] Figure 8 This is a comparison chart of the thermal effects of the present invention and various comparative models, specifically Grad-CAM. Detailed Implementation

[0023] The technical solution of the present invention will be clearly and completely described below with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are only some, not all, embodiments of the present invention. The present invention is not limited to the following embodiments. All embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0024] This embodiment constructs a lightweight brain tumor MRI image classification model based on chessboard pixel decomposition. The overall network flow is as follows: Figure 1 As shown, the entire process sequentially completes image preprocessing, shallow feature extraction, deep feature encoding, adaptive feature fusion, and multi-scale classification output, achieving a four-classification task for brain MRI images: normal brain tissue, glioma, meningioma, and metastatic tumor.

[0025] This embodiment integrates multiple publicly available medical imaging datasets to construct the BrainMRI-4Class standardized dataset. Typical images of various types of samples in the dataset are shown below. Figure 6 As shown, this dataset integrates the BraTS2025 sub-task datasets and the IXI public dataset to construct a four-class baseline dataset. Gliomas, meningiomas, and metastatic tumor samples are taken from different sub-tasks of BraTS2025, while normal brain tissue samples are selected from the IXI dataset. After unified coordinate system registration, resampling, brain parenchyma dissection, and slice quality control, a total of 1174 cases and 45970 standardized multi-sequence MRI slices were finally compiled, fully covering the four categories of samples: normal brain tissue, gliomas, meningiomas, and metastatic tumors. To prevent data leakage, this invention uses a patient-level five-fold cross-validation method to divide the dataset, uniformly classifying all MRI slices of the same patient into the training or test set, ensuring the rigor of data division and the authenticity of experimental results.

[0026] Before formal model training, standardized preprocessing operations were performed on the original MRI images. For the normal samples with dual-sequence IXI in the dataset, supplementary channels were generated through Sigmoid contrast enhancement and edge filtering, uniformly completing the input to a four-channel format. All images were uniformly resampled to a fixed resolution of 240×240, and pixel values ​​were constrained to the [-1,1] interval through min-max normalization. Simultaneously, data augmentation strategies were only enabled on the training set, randomly employing four enhancement methods: ±10° rotation, horizontal flip, σ=0.7 Gaussian blur, and 1.08x contrast adjustment. The probability of enabling any single-class enhancement was set to 50%. No enhancement processing was performed on the validation and test sets to ensure the objectivity and validity of the test results.

[0027] The preprocessed standardized image is input into the shallow feature extraction module. This invention employs a checkerboard pixel splitting structure to achieve lossless shallow feature extraction. The specific structure is as follows: Figure 2As shown. This structure relies on the parity index of the image's rows and columns to perform lossless splitting of the original image, generating two images with the same resolution as the original. Figure Two The complementary sub-images are split into two parts, with no learnable parameters during the splitting process, avoiding the loss of details in traditional downsampling. The two complementary sub-images are extracted using independent lightweight 3×3 convolutions. After being spliced ​​and upgraded in the channel dimension, spatial downsampling is completed by convolution with a stride of 2. Combined with conventional convolutions to expand the feature channels, the final output is a shallow initial feature map that accurately preserves the edges of lesions and the details of small metastases.

[0028] Subsequently, the shallow initial features are input into a three-level cascaded progressive refinement coding structure for deep semantic feature extraction. The single-level encoder module structure of this invention is as follows: Figure 3 As shown, the single-level module internally uses a lightweight downsampling unit composed of 1×1 compressed convolution, max pooling, and 1×1 extended convolution. After each layer operation, a BN layer and ReLU activation function are configured to complete feature normalization and non-linear activation. At the same time, a channel-space dual attention mechanism is integrated. The channel branch relies on global average pooling to build channel dependencies, while the spatial branch fuses pixel mean and extreme values ​​and combines them with 5×5 convolution to extract local lesion spatial features. Non-linear feature fusion is achieved through two stacked 3×3 convolutions. A residual bypass structure is also added. After the input features are pooled and 1×1 convolution to complete dimensional alignment, they are added to the residual of the backbone features. The three-level module compresses the spatial size and expands the feature channels step by step, gradually strengthening the deep semantic features of tumor lesions.

[0029] After completing the deep encoding, adaptive fusion of deep and shallow features is achieved through an accumulation and gated skip connection structure, as shown in the figure below. Figure 4 As shown, firstly, the shallow feature size is scaled using bilinear interpolation, and then matched with 1×1 convolutions to achieve dimensionality matching between deep and shallow feature channels. After concatenating the two types of features, two layers of 1×1 convolutions are used to compress the channels to one-eighth of the original channels, generating an initial spatial gating weight map. Row and column direction cumulative sum operations are then performed sequentially on the weight map to smooth weight fluctuations, and then normalized using the Sigmoid function to obtain accurate gating coefficients. Based on the gating coefficients, deep semantic features and shallow detail features are weighted and fused to compensate for the high-frequency pathological information lost in the deep encoding process.

[0030] The fused high-precision feature input gated multi-scale classifier completes the final classification. The internal structure of the classifier is as follows: Figure 5As shown, the classifier has four dilated convolutional branches with dilation coefficients of 1, 2, 4, and 8. Each branch is pre-compressed with a 1×1 convolution to compress one-quarter of the channels, reducing computational redundancy. The model adaptively generates the weights of each branch based on the input features, and after Softmax normalization, it weights and fuses multi-scale lesion features to adapt to the recognition needs of tumor lesions of different sizes. The feature channels are restored by 1×1 convolution and residual connections are constructed. 3×3 convolutions are used to simplify features and compensate for detail loss. Finally, global average pooling and fully connected layers are used to output the four-class classification prediction results.

[0031] This embodiment uses an RTX5090 32G hardware device and the PyTorch deep learning framework to train the model. The Adam optimizer is used for training, with an initial learning rate of 3e-3, a weight decay coefficient of 1e-4, a training batch size of 32, and 50 rounds of iterative training. The training loss adopts the standard multi-class cross-entropy loss, with the weights of the four classes uniformly configured to 1. Additional optimization strategies such as label smoothing are turned off, and the model relies entirely on its own structure to adapt to the imbalanced dataset.

[0032] After the experiment, the overall classification performance of the model was quantitatively evaluated using ROC curves. The ROC performance curves of this invention and various comparative models are shown below. Figure 7 As shown; simultaneously, the Grad-CAM thermal mapping visualization model's lesion focusing capability is employed, and the thermal effects of different comparative models are compared, for example... Figure 8 As shown, it can be intuitively verified that the model of the present invention can accurately focus on the actual lesion area of ​​the tumor, which is highly consistent with the logic of clinical image reading.

[0033] Experimental results show that, with extremely low parameter and computational requirements, this invention effectively solves the shortcomings of traditional lightweight networks, such as loss of detail, low accuracy in rare tumor identification, and reliance on additional training strategies. The model's classification accuracy, generalization ability, and robustness are significantly better than existing lightweight models of the same type, balancing lightweight deployment performance with clinical diagnostic accuracy.

[0034] The above embodiments only illustrate preferred embodiments of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of protection of the present invention. It should be noted that for those skilled in the art, any modifications, improvements, and substitutions made without departing from the concept of the present invention are all within the scope of protection of the present invention.

Claims

1. A lightweight brain tumor MRI image classification method based on checkerboard pixel decomposition, characterized in that, The specific steps are as follows: S1: Image preprocessing standardization pipeline, which performs standardized processing on the acquired multi-sequence brain MRI original images and outputs a uniform four-channel standardized MRI image; S2: Chessboard splitting shallow feature extraction subsystem, which relies on the chessboard pixel splitting mechanism to complete the shallow feature extraction of the image and generate a shallow initial feature map that preserves the details of the lesions; S3: A progressive deep coding enhancement framework that extracts feature information layer by layer through a three-level cascaded coding module to generate a feature map with deep semantic representation; S4: Accumulation and gated feature fusion unit, which uses a gated jump connection structure to fuse shallow detail features and deep semantic features to generate fused feature data; S5: Gated multi-scale classification prediction module, which uses multi-scale classification structure analysis and fusion features to output four classification results: normal brain tissue, glioma, meningioma, and metastatic tumor.

2. The lightweight brain tumor MRI image classification method based on checkerboard pixel segmentation according to claim 1, characterized in that, The image preprocessing standardization pipeline in S1 specifically includes the following: S1-1: Size uniform resampling unit, which scales all input MRI images to a fixed resolution of 240×240; S1-2: Channel completion generation unit, which uses an image transformation algorithm to generate supplementary channels for normal dual-sequence samples, and uniformly completes them into a four-channel input format; S1-3: Pixel normalization processing unit, which uses the min-max normalization method to constrain all pixels in the image to the value range of [-1, 1]; S1-4: Random Data Augmentation Unit, which enables four random image augmentation strategies only during the model training phase: ±10° rotation, horizontal flip, σ=0.7 Gaussian blur, and 1.08x contrast adjustment.

3. The lightweight brain tumor MRI image classification method based on checkerboard pixel segmentation according to claim 1, characterized in that, The chessboard splitting shallow feature extraction subsystem in S2 specifically includes the following: S2-1: Chessboard parity splitting unit, which cuts the original image according to the parity index of the pixel row and column to obtain two complementary sub-images with half the size. The splitting process does not introduce learnable parameters. S2-2: Dual-branch convolutional upscaling unit, where two subgraphs are spliced ​​together in the channel dimension after undergoing lightweight convolution operations to expand the feature channels; S2-3: Convolutional downsampling unit, which uses convolution with a stride of 2 to compress spatial dimensions, and outputs shallow initial features after expanding the channel dimensions by combining with ordinary convolution.

4. The lightweight brain tumor MRI image classification method based on checkerboard pixel segmentation according to claim 1, characterized in that, The single-level coding module of the progressive deep coding enhancement framework in S3 specifically includes the following: S3-1: Lightweight channel compression downsampling unit, which sequentially connects 1×1 compressed convolution, max pooling, and 1×1 expanded convolution to complete feature size and channel control; S3-2: Dual-channel spatial attention unit. The channel branch relies on global pooling to capture channel correlation, while the spatial branch fuses pixel mean and extreme values ​​and combines 5×5 convolution to extract local spatial features. S3-3: Stacked convolutional feature fusion unit, which uses two consecutive 3×3 convolutions to complete non-linear feature fusion; S3-4: Residual dimension-adapting bypass unit. The original input features are pooled and 1×1 convolution to complete dimension matching, and then superimposed on the backbone features.

5. A lightweight brain tumor MRI image classification method based on checkerboard pixel segmentation according to claim 1, characterized in that, The cumulative and gated feature fusion unit in S4 specifically includes the following: S4-1: Feature scale alignment unit, which uses bilinear interpolation to scale the size of shallow features and combines it with 1×1 convolution to achieve uniformity of deep and shallow feature channels; S4-2: Spatial gating graph generation unit, which concatenates two types of features and generates a weight map through two layers of 1×1 convolution. During the intermediate operation stage, the number of channels is compressed to 1 / 8 of the original number of channels. S4-3: Row and column cumulative smoothing unit, which first accumulates the original gated weight map and then accumulates the columns to smooth spatial weight fluctuations; S4-4: Adaptive weighted fusion unit, which normalizes the gating coefficients using the Sigmoid function and completes feature fusion according to the calculation formula: deep features + gating coefficient × shallow features.

6. A lightweight brain tumor MRI image classification method based on checkerboard pixel segmentation according to claim 1, characterized in that, The gated multi-scale classification prediction module in S5 specifically includes the following components: S5-1: Multi-diffuse convolution branch unit, with four 3×3 dilated convolutions set with dilation coefficients of 1, 2, 4, and 8. Each branch is preceded by a 1×1 convolution to compress the channel to 1 / 4 of its original size. S5-2: Dynamic weight adaptive unit, which generates branch weights based on input features, and then weights and fuses multi-path, multi-scale features after Softmax normalization; S5-3: Residual Channel Restoration Unit, which uses 1×1 convolution to restore feature channels and build residual connections between input and output; S5-4: Feature Refinement Classification Output Unit. It simplifies channels through 3×3 convolution and introduces skip connections to compensate for detail loss. It outputs four-class prediction results through global average pooling and fully connected layers.

Images