MRI brain tumor image segmentation method based on sparse deconvolution enhanced preprocessing
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIHANG UNIV
- Filing Date
- 2026-03-30
- Publication Date
- 2026-07-21
Smart Images
Figure CN122434829A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary fields of medical image processing, computer vision and artificial intelligence, and more specifically, relates to a method and system for segmenting MRI brain tumor images based on sparse deconvolution enhancement preprocessing. Background Technology
[0002] Accurate segmentation of brain tumors is fundamental for clinical diagnosis, surgical planning, radiotherapy target delineation, and efficacy evaluation. Current automated segmentation methods primarily rely on deep learning, typically employing convolutional neural networks (CNNs) or Transformer structures such as U-Net and its variants. These methods involve supervised training on a large number of labeled MRI images before segmentation prediction on the test data. However, in real-world clinical settings, high-quality segmentation annotation requires experienced physicians and significant time investment, making it difficult to expand the sample size. This results in small training data sets, significant distributional biases, and a tendency for the model to overfit and exhibit insufficient generalization ability.
[0003] Furthermore, MRI images are often limited in resolution, have blurred boundaries, and low contrast due to factors such as k-space sampling bandwidth, magnetic field inhomogeneity, motion artifacts, and noise during acquisition. When tumor boundaries are unclear or the internal texture of the tumor is overwhelmed by noise, the effective feature learning of the neural network is weakened, further amplifying the difficulty of training with small samples.
[0004] Existing medical image preprocessing generally includes normalization, denoising, artifact removal, histogram equalization, etc., but most methods are general filtering or empirical enhancement, which are difficult to maintain lesion boundaries while suppressing noise, and also difficult to significantly improve small sample segmentation performance without relying on additional labeled data. Summary of the Invention
[0005] Based on the above problems, this invention aims to solve the following problems: (1) When the number of labeled samples is small, the existing neural network segmentation model is not trained sufficiently, converges slowly, and has blurred prediction boundaries. The segmentation accuracy is highly sensitive to the size of the training set. (2) MRI images have limited resolution and noise interference, which weakens the edge and internal texture information of lesions at the input end, limiting the learnable features of the subsequent segmentation network. (3) An enhancement preprocessing scheme is needed that does not rely on additional labeled data, is compatible with multiple segmentation networks, and can stably improve image quality within a limited parameter range, so as to improve the segmentation effect and training efficiency under small sample conditions.
[0006] To address the aforementioned deficiencies or improvement needs of existing technologies, as a first aspect of this invention, the present invention provides a method for segmenting MRI brain tumor images based on sparse deconvolution enhancement preprocessing, comprising: S1. Prepare an MRI image dataset containing brain tumor annotation masks, complete the dataset division, and perform standardized preprocessing on the MRI images of each dataset to obtain basic preprocessed images; S2. Based on the k-space band-limited sampling characteristics of MRI imaging, a convolution-noise degradation model is constructed, and the equivalent point spread function corresponding to the MRI image is derived. Taking the basic preprocessed image as input, a sparse deconvolution reconstruction optimization model is constructed that integrates data fidelity term, Hessian continuity prior regularization term and L1 sparse prior regularization term. A parameter strategy adapted to MRI two-dimensional slices is configured and the model is solved iteratively. Sparse deconvolution enhancement processing is uniformly performed on the basic preprocessed images of the training set, validation set and test set to obtain the enhanced MRI image. S3. Select a medical image segmentation neural network as the brain tumor segmentation model, build the training framework of the segmentation model and configure a standardized training optimization strategy; take basic preprocessed images and sparse deconvolution enhanced MRI images as model inputs respectively, and perform model training for training sets with different sample sizes under unified training parameters and data augmentation strategies. Use the core evaluation index of medical image segmentation as the judgment basis, save the optimal weight of the model under each training mode to obtain the optimal segmentation model. S4. Perform standardized basic preprocessing on the test set MRI images in the same way as in the training phase. Use the original input inference process and the sparse deconvolution enhanced inference process respectively. Input the processed images into the corresponding optimal segmentation model to obtain the brain tumor prediction mask. Evaluate the image reconstruction effect through contour consistency quantification index and evaluate the model segmentation effect through pixel-level segmentation performance quantification index. Output the final MRI brain tumor image segmentation result and comprehensive performance evaluation index.
[0007] Furthermore, the preprocessing process in S1 includes, but is not limited to: unifying the size of the input image, cropping it, and correcting its offset; and converting the single-channel image into a 3-channel input according to the network input requirements.
[0008] Furthermore, the equivalent point spread function in S2 is specifically: The MRI spatial domain image is considered as the inverse Fourier transform result after k-space finite bandwidth sampling. Let M(k) be the bandwidth sampling window in k-space. The process of band-limited sampling of the ideal image x and superimposing noise to obtain the observed image is as follows: ; in X(k) is the inverse Fourier transform from k-space to spatial domain, where X(k) is the k-space representation of the ideal image x, and n is the noise term. The process is transformed by an equivalent method, and the equivalent point spread function h is defined as follows: This transforms the bandwidth-limited effect of MRI imaging into a spatial domain convolutional form. , This represents the convolution operation.
[0009] Furthermore, the sparse deconvolution reconstruction optimization model in S2 is specifically as follows: Using the preprocessed image as the image to be reconstructed, a sparse deconvolution reconstruction optimization model is constructed, jointly constrained by a data fidelity term, a Hessian continuity prior regularization term, and an L1 sparse prior regularization term. The expression of this model is: in, For the reconstructed enhanced MRI image, The weighting coefficients for data fidelity items. The pixel value matrix of the preprocessed image as the basis. The convolution operator defined for the equivalent point spread function, To smooth the background item, For the squaring operation of the L2 norm, This is a Hessian-based continuity regularization term used to suppress artifacts caused by noise amplification and improve image structural continuity. The weighting coefficients are for L1 sparse priors. This is an L1 norm operation used to balance the extraction of high-frequency information from images with the preservation of weak structures.
[0010] Furthermore, the iterative solution process in S2 is as follows: First, a parameter strategy adapted to MRI two-dimensional slices is configured for the sparse deconvolution reconstruction optimization model. The parameter strategy includes turning off the background estimation and upsampling steps, setting the axial continuity term to 0, setting the number of sparse reconstruction iterations to no less than 100, and using the RL form for the iterative deconvolution step and configuring a preset number of iterations. Then, based on the parameter strategy, the sparse deconvolution reconstruction optimization model is iteratively solved. During the iteration process, the content-aware parameters of image fidelity, sparsity, and number of deconvolution iterations are determined only on the training set and validation set through grid search and visual inspection. All determined parameters are fixed and reused in the sparse deconvolution enhancement process of the test set. The solution is terminated after completing a preset number of iterations to obtain the enhanced MRI image.
[0011] Furthermore, the original input inference process in S4 uses the basic preprocessed image as input; the sparse deconvolution enhanced inference process uses the enhanced image as input.
[0012] Furthermore, the contour consistency quantization process in S4 is as follows: The pixel set of the predicted contour after MRI image reconstruction and the pixel set of the real contour are extracted. The 95% Hausdorff distance and the mean surface distance between the two point sets are calculated respectively. The contour consistency is quantitatively evaluated by the two indicators. The 95% Hausdorff distance is the 95th percentile of all unidirectional distances between the predicted contour point set and the true contour point set, used to measure the maximum deviation level of the contour boundary; the average surface distance is the average Euclidean distance between the two sets of contour point sets, used to reflect the overall fitting accuracy of the contour.
[0013] Furthermore, the process of evaluating the model's segmentation effect using pixel-level segmentation performance quantification metrics in S4 is as follows: Based on the pixel-level comparison between the brain tumor prediction mask and the real labeled mask, the number of true positive pixels (TP), the number of false negative pixels (FN), the number of false positive pixels (FP), and the number of true negative pixels (TN) were statistically obtained. The performance of the segmentation model was quantified by four core indicators: Intersection over Union (IoU), Dice similarity coefficient, Sensitivity, and Specificity. The comprehensive quantitative evaluation of the model's segmentation effect was completed through the calculation results of the four indicators. The Intersection over Union (IoU) ratio reflects the degree of overlap between the predicted region and the actual region, and its calculation formula is as follows: ; The Dice similarity coefficient is used to characterize the overlap consistency of segmentation results and has good adaptability to class imbalance scenarios. The calculation formula is as follows: ; Sensitivity measures a model's ability to identify real positive samples, and is calculated using the following formula: ; Specificity measures a model's ability to identify true negative samples, and is calculated using the following formula: ; The values of IoU, Dice similarity coefficient, sensitivity, and specificity are all in the range of [0,1]. The higher the value, the better the model segmentation performance.
[0014] As a second aspect of the present invention, an MRI brain tumor image segmentation system based on sparse deconvolution enhancement preprocessing is also provided, comprising: The dataset preprocessing unit is used to prepare MRI image datasets containing brain tumor annotation masks, complete dataset partitioning, and perform standardized preprocessing on the MRI images of each dataset to obtain basic preprocessed images. The sparse deconvolution enhancement unit is used to construct a convolution-noise degradation model based on the k-space band-limited sampling characteristics of MRI imaging, and derive the equivalent point spread function corresponding to the MRI image. Taking the basic preprocessed image as input, a sparse deconvolution reconstruction optimization model is constructed that integrates data fidelity term, Hessian continuity prior regularization term and L1 sparse prior regularization term. The parameter strategy adapted to MRI two-dimensional slices is configured and the model is solved iteratively. Sparse deconvolution enhancement processing is uniformly performed on the basic preprocessed images of the training set, validation set and test set to obtain the enhanced MRI image. The segmentation model training unit is used to select a medical image segmentation neural network as a brain tumor segmentation model, build the training framework of the segmentation model and configure a standardized training optimization strategy; the model is trained on training sets with different sample sizes under the same training parameters and data augmentation strategy, using basic preprocessed images and sparse deconvolution enhanced MRI images as model inputs respectively; the optimal weights of the model under each training mode are saved based on the core evaluation indicators of medical image segmentation to obtain the optimal segmentation model. The inference and performance evaluation unit performs standardized basic preprocessing on the test set MRI images in a manner consistent with the training phase. It employs both the original input inference process and the sparse deconvolution enhanced inference process to input the processed images into the corresponding optimal segmentation model, thereby obtaining a brain tumor prediction mask. The image reconstruction effect is evaluated using a contour consistency quantification index, and the model segmentation effect is evaluated using a pixel-level segmentation performance quantification index. The final MRI brain tumor image segmentation result and comprehensive performance evaluation index are then output.
[0015] As a third aspect of the invention, a computer-readable storage medium is also provided, on which a computer program is stored, which is executed by a processor as described in any one of the claims: an MRI brain tumor image segmentation method based on sparse deconvolution enhancement preprocessing.
[0016] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: 1. The MRI brain tumor image segmentation method based on sparse deconvolution enhancement preprocessing of the present invention provides a rigorous mathematical modeling foundation for sparse deconvolution enhancement by constructing a convolution-noise degradation model of MRI imaging and deriving the equivalent point spread function. Taking the preprocessed images as input, a sparse deconvolution reconstruction optimization model integrating data fidelity terms, Hessian continuity priors, and L1 sparse priors is constructed. A parameter strategy adapted to MRI two-dimensional slices is configured and iteratively solved to achieve unified unsupervised enhancement of training, validation, and test set images. This step effectively improves the spatial resolution and boundary contrast of MRI images without increasing the cost of manual annotation, enhances the texture information and structural continuity of the tumor region, provides a high-quality feature input source for subsequent neural network segmentation, and significantly reduces the segmentation error caused by image quality degradation.
[0017] 2. The MRI brain tumor image segmentation method based on sparse deconvolution enhancement preprocessing of the present invention selects a medical image segmentation neural network as the brain tumor segmentation model, builds a standardized training framework and configures a unified optimization strategy; using basic preprocessed images and sparse deconvolution enhanced MRI images as model inputs respectively, under the same training parameters and data augmentation strategy, model training is performed on training sets with different sample sizes, and the optimal weights of each training mode are saved based on the core evaluation indicators of medical image segmentation; this step verifies and locks the improvement effect of sparse deconvolution enhancement on small sample scenarios through dual-input mode comparison and multi-sample gradient testing, ensuring that the optimal segmentation model obtained by training has both generalization ability and robustness, and provides reliable model support for test set inference.
[0018] 3. The MRI brain tumor image segmentation method based on sparse deconvolution enhancement preprocessing of the present invention performs standardized basic preprocessing on the test set MRI images in the same manner as the training phase. It then employs both the original input inference process and the sparse deconvolution enhancement inference process to input the processed images into the corresponding optimal segmentation model to obtain a brain tumor prediction mask. The image reconstruction effect is evaluated using a contour consistency quantification index, and the model segmentation effect is evaluated using a pixel-level segmentation performance quantification index, resulting in a comprehensive evaluation. This step enables a quantitative comparison of the segmentation effects before and after enhancement, accurately quantifying the actual improvement effect of sparse deconvolution enhancement on boundary recovery, region overlap, and category recognition capabilities, providing a quantitative verification basis for the effectiveness and practicality of the method. Attached Figure Description
[0019] Figure 1 This is a flowchart of an MRI brain tumor image segmentation method based on sparse deconvolution enhancement preprocessing according to an embodiment of the present invention; Figure 2The following are the simulation results of the embodiments of the present invention. a) Comparison of the original image, degraded image, and weakly sparse deconvolution-enhanced super-resolution reconstructed image of the geometric image, along with magnified views of the three regions; b) Canny edge detection results and HD95 and MSD analysis of the degraded geometric image and the weakly sparse deconvolution-enhanced super-resolution reconstructed image; c) Comparison of the original image, degraded image, and weakly sparse deconvolution-enhanced super-resolution reconstructed image of the brain map, along with magnified views of the two regions; d) Multi-threshold contour analysis results and mean HD95 analysis of the degraded brain map image and the weakly sparse deconvolution-enhanced super-resolution reconstructed image.
[0020] Figure 3 The following are brain MRI image reconstruction results according to an embodiment of the present invention: a) is a comparison of the original brain MRI image and its weakly sparse unconvolution-enhanced super-resolution reconstructed image, where the red area is the tumor mask marker; b) is a magnified view of the two regions in a before and after reconstruction; c) is the tumor region predicted by the MKU-net and U-net models trained with the datasets before and after reconstruction, respectively, where the red box is the region marked by the original mask in a.
[0021] Figure 4 This invention presents an analysis of the prediction results of MKU-net and U-net models for brain MRI images. Example a shows the Dice analysis of the prediction results of MKU-net and U-net models trained with pre- and post-reconstruction datasets at different training set sizes. Example b shows the Dice data change curves of the pre- and post-reconstruction datasets during the U-net model training process with a training set of 400 units, where the thick line represents the higher-order fit of the training curve. Example c shows the IoU, Sensitivity, and Specificity analysis of the U-net model's prediction results for the pre- and post-reconstruction datasets with a training set of 400 units. Example d shows the IoU, Sensitivity, and Specificity analysis of the MKU-net model's prediction results for the pre- and post-reconstruction datasets with a training set of 400 units.
[0022] Figure 5 This is a schematic diagram of the system units in an embodiment of the present invention. Detailed Implementation
[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.
[0024] Example 1 Please refer to Figure 1This embodiment 1 provides a method for segmenting MRI brain tumor images based on sparse deconvolution enhancement preprocessing, including: S1. Prepare an MRI image dataset containing brain tumor annotation masks, complete the dataset division, and perform standardized preprocessing on the MRI images of each dataset to obtain basic preprocessed images; S2. Based on the k-space band-limited sampling characteristics of MRI imaging, a convolution-noise degradation model is constructed, and the equivalent point spread function corresponding to the MRI image is derived. Taking the basic preprocessed image as input, a sparse deconvolution reconstruction optimization model is constructed that integrates data fidelity term, Hessian continuity prior regularization term and L1 sparse prior regularization term. A parameter strategy adapted to MRI two-dimensional slices is configured and the model is solved iteratively. Sparse deconvolution enhancement processing is uniformly performed on the basic preprocessed images of the training set, validation set and test set to obtain the enhanced MRI image. S3. Select a medical image segmentation neural network as the brain tumor segmentation model, build the training framework of the segmentation model and configure a standardized training optimization strategy; take basic preprocessed images and sparse deconvolution enhanced MRI images as model inputs respectively, and perform model training for training sets with different sample sizes under unified training parameters and data augmentation strategies. Use the core evaluation index of medical image segmentation as the judgment basis, save the optimal weight of the model under each training mode to obtain the optimal segmentation model. S4. Perform standardized basic preprocessing on the test set MRI images in the same way as in the training phase. Use the original input inference process and the sparse deconvolution enhanced inference process respectively. Input the processed images into the corresponding optimal segmentation model to obtain the brain tumor prediction mask. Evaluate the image reconstruction effect through contour consistency quantification index and evaluate the model segmentation effect through pixel-level segmentation performance quantification index. Output the final MRI brain tumor image segmentation result and comprehensive performance evaluation index.
[0025] This embodiment 1 further elaborates on the above steps.
[0026] (1) Dataset preprocessing Before performing neural network segmentation on MRI brain tumor images, it is necessary to prepare and standardize the image dataset. This provides standardized and high-quality basic image data for subsequent sparse deconvolution enhancement and model training, and solves the problems of inconsistent clinical MRI image acquisition specifications and bias that affect the subsequent processing results.
[0027] First, an MRI image dataset containing brain tumor annotation masks was prepared. In this embodiment, the BRISC dataset was selected, which contains approximately 6,000 contrast-enhanced T1-weighted MRI slices and corresponding tumor annotation masks. When dividing the dataset, 500 images were fixed as the test set, and 500, 1,000, and 1,500 images were extracted from the remaining samples to form annotation training pools. Each annotation training pool was divided into training and validation sets in an 8:2 ratio, resulting in training sample sizes of 400, 800, and 1,200 respectively. Different sample size gradients were set in this way for subsequent model training and validation.
[0028] After the dataset is divided, standardized preprocessing operations are performed on the MRI images of each set. First, the input images are sized uniformly according to the actual processing requirements. At the same time, necessary cropping and offset correction are carried out to eliminate the offset deviation generated during image acquisition and to regularize the effective area of the image. Then, according to the input requirements of the subsequent neural network, the single-channel MRI images are converted into a 3-channel input format. Through standardized preprocessing operations, the MRI images of each set reach a uniform input specification, and finally the basic preprocessed images are obtained to ensure the smooth progress of subsequent weak sparse deconvolution enhancement processing and model training.
[0029] To address the channel dimension conversion of single-channel MRI images, this embodiment, based on the adaptability requirements of the grayscale features of MRI images and neural network channel inputs, adopts a multi-channel mapping reconstruction method for single-channel grayscale information to convert it into a 3-channel input form: First, the pixel grayscale value matrix of the single-channel MRI image is extracted as the core feature basis, and this grayscale feature matrix is synchronously mapped to the pixel space of the three channels, keeping the pixel grayscale information of each channel completely consistent with the original single-channel image. At the same time, each channel is given an independent pixel dimension identifier, so that the three-channel image not only retains all the grayscale texture and lesion feature information of the original MRI image, but also forms a three-dimensional channel structure that conforms to the neural network input specifications.
[0030] This conversion method avoids the image information distortion problem caused by random assignment or feature tampering in multiple channels. While meeting the requirements of the neural network for the 3-channel input format, it maximizes the preservation of the original anatomical structure and tumor region features of the single-channel MRI image, ensuring the integrity of the feature input for subsequent weak sparse deconvolution enhancement processing. At the same time, it matches the feature dimension of the three-channel image with the extraction dimension of the convolution kernel of the neural network, improving the perception and learning efficiency of the convolution layer for key features such as tumor boundaries and textures, and providing format-adapted and information-complete basic image data for the training of subsequent segmentation models.
[0031] (2) Sparse deconvolution enhancement Constructing a convolution-noise degradation model based on the k-space band-limited sampling characteristics of MRI imaging and deriving the equivalent point spread function corresponding to the MRI image is the core mathematical foundation for sparse deconvolution enhancement of MRI images and a key link connecting the physical characteristics of MRI imaging with subsequent image reconstruction. In the MRI imaging process, the image is essentially a spatial domain result obtained by inverse Fourier transform after k-space sampling with limited bandwidth. Limited by the sampling bandwidth of the imaging device, complete k-space information cannot be obtained. This characteristic directly leads to bandwidth-limited degradation in the imaging result. Simultaneously, noise interference is introduced during the imaging process, and these two factors together constitute the core degradation factors of MRI images. Therefore, the MRI spatial domain image is considered as the inverse Fourier transform result after k-space finite bandwidth sampling. Let M(k) be the bandwidth sampling window of k-space. The process of band-limited sampling of the ideal image x and superimposing noise to obtain the observed image is as follows: ; in X(k) is the inverse Fourier transform from k-space to spatial domain, where X(k) is the k-space representation of the ideal image x, and n is the noise term. The process is transformed by an equivalent method, and the equivalent point spread function h is defined as follows: This transforms the bandwidth-limited effect of MRI imaging into a spatial domain convolutional form. , This represents the convolution operation.
[0032] Meanwhile, using standardized preprocessed MRI images as input, a weakly sparse deconvolution reconstruction optimization model is constructed, fusing data fidelity terms, Hessian continuity prior regularization terms, and L1 sparse prior regularization terms. This is the core step in achieving a gentle improvement in MRI image quality. The "weak" aspect of this scheme is reflected in the fact that, while maintaining the core framework of the sparse deconvolution algorithm, only mild edge and detail enhancement is applied to the images by constraining the number of iterations and the enhancement intensity. This avoids artifacts and information distortion caused by over-repair, preserving the anatomical integrity of the original image while appropriately enhancing the texture and boundary features of the tumor region. The data fidelity term ensures the consistency of core information between the reconstructed image and the original preprocessed image, preventing the loss of original features due to over-reconstruction. The Hessian continuity prior regularization term focuses on the structural continuity of the image, effectively suppressing artifacts caused by noise amplification during deconvolution, allowing the tumor region boundaries and texture structure of the reconstructed image to remain natural and coherent. The L1 sparse prior regularization term balances the need for high-frequency information extraction and the preservation of weak structures by constraining the sparseness of image pixels, appropriately enhancing the key features of the tumor region without destroying the overall detail features of the image. The model expression is: in, For the reconstructed enhanced MRI image, The weighting coefficients for data fidelity items. The pixel value matrix of the preprocessed image as the basis. The convolution operator defined for the equivalent point spread function, Smooth background option is optional. For the squaring operation of the L2 norm, This is a Hessian-based continuity regularization term used to suppress artifacts caused by noise amplification and improve image structural continuity. The weighting coefficients are for L1 sparse priors. This is an L1 norm operation used to balance the extraction of high-frequency information from images with the preservation of weak structures.
[0033] Furthermore, configuring parameter strategies adapted to MRI 2D slices and iteratively solving the sparse deconvolution reconstruction optimization model directly determines the enhancement effect and model adaptability. Considering the imaging characteristics and data properties of MRI 2D slices, the parameter strategy configuration involves disabling background estimation and upsampling steps to avoid invalid processing of non-target areas and unnecessary information interpolation distortion; setting the axial continuity term to zero to fit the processing scenario of 2D slices and reduce interference from irrelevant parameters in the reconstruction process; setting no fewer than 100 sparse reconstruction iterations to ensure sufficient convergence of the reconstruction process and achieve the desired reconstruction effect; and using an RL (Regression-Based Algorithm) approach with a preset number of iterations to reduce the probability of noise amplification and edge ringing, ensuring the clarity and coherence of the reconstructed image.
[0034] During the iterative solution process, content-aware parameters such as image fidelity, sparsity, and the number of iterative deconvolutions are determined only on the training and validation sets through a combination of grid search and visual inspection. This avoids over-reliance on and leakage of information from the test set, ensuring the generalizability and rationality of the parameter settings. The fidelity constraint is above 300, the sparsity constraint is below 3, and the number of iterative deconvolutions is constrained to be between 1 and 5.
[0035] After the parameters are determined, they are fixed and reused in the sparse deconvolution enhancement processing of the training set, validation set and test set. The solution process is completed according to the preset number of iterations, and finally high-quality enhanced MRI images are obtained, realizing the effective optimization of the original preprocessed images.
[0036] (3) Training of segmentation model After preprocessing and sparse deconvolution enhancement of MRI images, the training phase of the brain tumor segmentation model is initiated. This phase aims to verify the impact of sparse deconvolution enhancement on segmentation performance with different sample sizes and network structures through a standardized training framework and unified optimization strategy. Firstly, U-Net and MKU-Net, two typical medical image segmentation neural networks, are selected as evaluation networks for brain tumor segmentation. The model is built and implemented using the PyTorch framework. These two types of networks possess encoder-decoder structures adapted for medical image segmentation, effectively extracting tumor features from MRI images and achieving pixel-level segmentation.
[0037] To ensure the standardization of the training process and the comparability of the results, the labeled training pool was first divided into training and validation sets in an 8:2 ratio, constructing a standardized training and validation system. During the training phase, to improve the model's generalization ability, data augmentation operations such as flipping, rotating, and color jittering were performed on the input images to avoid overfitting due to a single sample size. Simultaneously, a unified training optimization strategy was configured, using the Adam optimizer for parameter updates and the BCDiceLoss loss function to measure the difference between the predicted results and the true annotations. This loss function combines the pixel-level classification characteristics of binary cross-entropy with the regional overlap characteristics of the Dice coefficient, better meeting the task requirements of brain tumor segmentation. Regarding training parameter settings, the initial learning rate was set to 0.001, and the learning rate was decayed by a factor of 10 every 10 training epochs, with a total of 100 training epochs. An early stopping mechanism was enabled to prevent overtraining and ensure a balance between training efficiency and model performance.
[0038] Building upon this foundation, models were trained using preprocessed images (Raw input) and sparsely deconvolutionally enhanced MRI images (weakly sparsely deconvolutionally enhanced input) as inputs, under completely uniform training parameters and data augmentation strategies. Training sets with different sample sizes (400, 800, and 1200 samples) were used. For each training sample size and network architecture, both the Raw input model and the weakly sparsely deconvolutionally enhanced input model were trained separately. This ensured that the only difference between the two models was the input image preprocessing method, eliminating interference from other factors such as training parameters and network architecture, and guaranteeing that subsequent performance differences could be accurately attributed to sparsely deconvolutionally enhanced preprocessing. During training, the Dice similarity coefficient of the validation set was used as the core evaluation metric, and model performance changes were monitored in real time. When the validation set Dice coefficient reached its optimum, the corresponding model weights were saved. Ultimately, the optimal segmentation model under different input modes and sample sizes was obtained, providing a unified model foundation for subsequent test set inference and performance evaluation.
[0039] (4) Reasoning and performance evaluation After training the segmentation model and determining the optimal weights, the next stage is test set inference and performance quantification evaluation. This stage is the core step in verifying the effect of sparse deconvolution enhancement on improving the segmentation effect of MRI brain tumors. Through standardized inference procedures and multi-dimensional quantitative indicators, the effectiveness of the method is objectively verified. First, standardized basic preprocessing is performed on the MRI images of the test set, which is completely consistent with that of the training stage. This ensures that the specifications and quality of the test set images are consistent with those of the training set, eliminates variable interference in the preprocessing stage, and provides consistent basic data for the subsequent inference process.
[0040] Based on the preprocessed test set images, two inference workflows were used for segmentation prediction: the original input inference workflow directly inputs the preprocessed images into the corresponding optimal segmentation model to obtain a brain tumor prediction mask; the sparse deconvolution enhancement inference workflow first performs sparse deconvolution enhancement on the preprocessed images, and then inputs the enhanced images into the corresponding optimal segmentation model to generate a prediction mask. The two workflows differ only in the image enhancement stage; all other inference parameters and model weights remain the same, allowing for a precise comparison of the actual impact of sparse deconvolution enhancement on the segmentation results.
[0041] After generating the predictive mask, a quantitative evaluation is conducted from two dimensions: image reconstruction effect and model segmentation effect. For image reconstruction effect evaluation, the image reconstruction / contour consistency evaluation uses the 95th percentile Hausdorff distance (HD95) and mean surface distance (MSD) as quantification metrics. HD95 is defined as the 95th percentile of all unidirectional distances between the predicted contour point set and the true contour point set, used to measure the maximum deviation level of the contour boundary, as described below: The 95% Hausdorff distance is the core quantitative indicator that characterizes the degree of deviation between the predicted contour and the true contour boundary. The set of pixels for predicting the contour consists of the two-dimensional coordinates of all pixels located at the edge of the contour in the brain tumor segmentation mask output by the model, and is the set of boundary features of the model segmentation result. The set of pixels representing the true contour consists of the two-dimensional coordinates of all pixels located at the contour edge in the manually annotated gold standard mask for brain tumors, serving as a benchmark for contour deviation evaluation. This is a Euclidean distance calculation operator used to solve for the straight-line distance between any two points in a pixel set in a two-dimensional image space. This indicates traversing the predicted contour point set. Each pixel in Calculate the distance from this point to the set of true contour points. All pixels Find the Euclidean distance and take the minimum value, then combine all pixels. The corresponding minimum Euclidean distances are compared, and the maximum value is taken. This value represents the maximum one-way deviation of the predicted contour from the true contour. This indicates traversing the set of real contour points. Each pixel in Calculate the distance from the point to the predicted contour point set. All pixels Find the Euclidean distance and take the minimum value, then combine all pixels. The corresponding minimum Euclidean distances are compared, and the maximum value is taken. This value represents the maximum one-way deviation of the true contour from the predicted contour. The operator for the maximum value is selected to extract the larger of the two one-way maximum deviations, thus obtaining the two-way maximum deviation between the predicted contour and the true contour in a single sample. The 95th percentile calculation function is used to statistically sort the maximum bidirectional deviation values of all samples in the test set. The value at the 95th percentile after sorting is selected as the final 95th Hausdorff distance result. This calculation method can effectively eliminate the interference of extreme outliers on the evaluation results and achieve robust quantification of the maximum deviation level of the contour boundary.
[0042] MSD calculates the average Euclidean distance between two sets of contour points, reflecting the overall fitting accuracy of the contour, as shown below: The average surface distance is the core quantitative indicator that characterizes the overall fitting accuracy between the predicted contour and the actual outline. and These represent the number of pixels in the predicted contour point set and the true contour point set, respectively, and are used to normalize the total distance and eliminate the impact of differences in point set size on the evaluation results. This indicates traversing the predicted contour point set. Each pixel in Calculate the distance from this point to the set of true contour points. All pixels Find the Euclidean distance and take the minimum value, then combine all pixels. The corresponding minimum Euclidean distances are summed to obtain the total one-way distance from the predicted outline to the actual contour; This indicates traversing the set of real contour points. Each pixel in Calculate the distance from the point to the predicted contour point set. All pixels Find the Euclidean distance and take the minimum value, then combine all pixels. The corresponding minimum Euclidean distances are summed to obtain the total one-way distance from the true contour to the predicted contour. The normalization coefficient is obtained by dividing the sum of the two unidirectional total distances by the total number of pixels in the two sets of points to get the minimum Euclidean distance from each contour point to the other contour, which reflects the overall fitting accuracy of the contour. The smaller the value, the higher the overall fitting degree between the predicted contour and the real contour. In terms of evaluating the segmentation performance of the model, a pixel-level multi-dimensional index system is used for comprehensive quantification: First, based on the pixel-level comparison between the brain tumor prediction mask and the real labeled mask, the number of true positive pixels (TP), the number of false negative pixels (FN), the number of false positive pixels (FP), and the number of true negative pixels (TN) are statistically obtained. The performance of the segmentation model is quantified by four core indicators: Intersection over Union (IoU), Dice similarity coefficient, Sensitivity, and Specificity. The comprehensive quantitative evaluation of the model segmentation performance is completed through the calculation results of the four indicators. The Intersection over Union (IoU) ratio reflects the degree of overlap between the predicted region and the actual region, and its calculation formula is as follows: ; The Dice similarity coefficient is used to characterize the overlap consistency of segmentation results and has good adaptability to class imbalance scenarios. The calculation formula is as follows: ; Sensitivity measures a model's ability to identify real positive samples, and is calculated using the following formula: ; Specificity measures a model's ability to identify true negative samples, and is calculated using the following formula: ; The values of IoU, Dice similarity coefficient, sensitivity, and specificity are all in the range of [0,1]. The higher the value, the better the model segmentation performance.
[0043] Please refer to Figure 2 This embodiment also provides a specific process for verifying the effectiveness of sparse deconvolution in restoring boundary details under the conditions of "bandwidth-limited + noise interference": the experiment uses the process of ground truth image → simulated degradation → sparse deconvolution reconstruction → geometric / contour error quantization for verification.
[0044] (1) Data and ground truth images: Two types of simulated images are used as ground truth images (GT). The first type is a geometric image generated by a self-written program (containing clear boundaries and regular structures) to evaluate the geometric shift and boundary diffusion caused by degradation; the second type is a grayscale structure image of brain tissue (containing continuous grayscale changes and complex textures), and the MNI152 standard brain template can be further used as an anatomical structure reference to evaluate the morphological fidelity under continuous grayscale structures.
[0045] (2) Simulation of degradation process: The geometric image uses frequency domain bandwidth truncation to simulate MRI imaging degradation, that is, the band-limited function H(u,v) is introduced in k space to suppress high frequency components to obtain the degradation image; the brain tissue image uses Gaussian blur (e.g. r=2) to simulate edge diffusion caused by resolution limitation, and superimposes noise terms consistent with the MRI scene to form degradation input.
[0046] (3) Sparse deconvolution reconstruction: Sparse deconvolution reconstruction is performed on the degraded image. The reconstruction uses an equivalent PSF / OTF (given by a band-limited function or parameterized kernel), and ringing noise is suppressed and structural boundaries are enhanced by continuity prior and L1 sparse prior constraints. The reconstruction hyperparameters are adjusted within a preset range, and a smaller number of iterations is used to avoid noise amplification.
[0047] (4) Evaluation indicators and analysis methods: Canny edge detection was performed on the geometric reconstruction results. After aligning the detected boundaries with the ground truth boundaries, the 95% Hausdorff distance (HD95) and mean surface distance (MSD) were calculated to quantify the maximum boundary deviation and overall offset. Multi-threshold contour analysis was performed on the brain tissue reconstruction results. The contour error under different thresholds was calculated and summarized as meanHD95 to measure the morphological consistency of continuous grayscale structures.
[0048] (5) Experimental phenomena and conclusions: such as Figure 2 As shown, in the geometric graph experiment, sparse deconvolution significantly suppresses ringing and artifacts caused by band-limited degradation and makes edge transitions sharper, with HD95 and MSD decreasing by approximately 21% and 14%, respectively. In the brain tissue graph experiment, sparse deconvolution can stabilize and shrink the edge diffusion caused by degradation, improving the consistency between the contour and the ground truth, with mean HD95 decreasing by approximately 69%. This example illustrates that controlled redistribution of band-limited information under structural prior constraints can improve the segmentability of medical images without introducing additional annotations.
[0049] Please refer to Figure 3 as well as Figure 4 This embodiment also provides a specific process for a brain tumor MRI three-group volume sample segmentation experiment, as follows: (1) Dataset and Grouping: The BRISC brain tumor MRI dataset was used for validation. A fixed set of 500 images was used as the test set; three additional training / validation sets were constructed with 500, 1000, and 1500 images respectively, divided into training and validation sets at an 8:2 ratio, corresponding to training sample sizes of 400, 800, and 1200. To ensure fairness in the comparison, the same data partitioning and annotation mask were used for all different processes.
[0050] (2) Basic preprocessing: Spatial alignment and size unification (resampling), intensity normalization / standardization, necessary bias field correction and cropping are performed on the MRI images to obtain the input images for segmentation training. According to the network input requirements, the single-channel MRI is converted into a three-channel input (e.g., copied to three channels or generated as a three-channel stack according to the preprocessing strategy).
[0051] (3) Sparse deconvolution enhancement: After basic preprocessing, sparse deconvolution enhancement is performed on the training set, validation set and test set respectively to obtain enhanced images. The enhancement process uses continuous prior and L1 sparse prior constraints, and achieves deblurring and noise suppression through iterative solution within a limited parameter range; the relevant parameters are fixed for the test set after being determined on the training / validation set, and test labels are not used to participate in parameter adjustment during the testing phase.
[0052] (4) Segmentation Model and Training Settings: U-net and MKU-net were selected as segmentation networks. Two models were trained for each training set size: the Raw process directly used the original images after basic preprocessing; the weak sparse deconvolution augmentation process used the images after sparse deconvolution augmentation. The two processes maintained the same network structure, input size, data augmentation strategy, loss function, and optimization strategy to ensure that the performance difference mainly came from the input changes brought about by the augmentation preprocessing. The model training was implemented using PyTorch, with BCEDiceLoss as the loss function, Adam optimizer, and an initial learning rate of 0.001. The learning rate was decayed by a factor of 10 every 10 epochs, and a total of 100 epochs were trained with an early stopping mechanism enabled. The model weights were saved when the Dice on the validation set reached the optimal value.
[0053] (5) Reasoning and evaluation: Output the prediction mask on the test set, calculate Dice, IoU, Sensitivity and Specificity, and record the Dice curve of the validation set during training to compare the convergence speed; at the same time, provide a visual comparison of "original image / enhanced image / local magnification / prediction mask" for typical samples.
[0054] (6) Results: The enhanced image showed clearer tumor boundaries and the predicted contours were closer to the annotations. Figure 3The gains are more significant with lower sample sizes: when training samples are 400 and 800, U-Net Dice improves by approximately 10.4% and 6.2%, respectively, while MKU-Net Dice improves by approximately 9.0% and 3.6%, respectively; when training samples are 400, reaching validation set Dice=0.5 can be achieved approximately 6 epochs earlier (speed improvement of approximately 13%); while maintaining high specificity, IoU and Sensitivity are further improved (U-Net approximately +10.2% and +20.5%, MKU-Net approximately +11.1% and +18.8%).
[0055] Example 2 Please refer to Figure 5 This embodiment 2 provides an MRI brain tumor image segmentation system based on sparse deconvolution enhancement preprocessing, including: The dataset preprocessing unit is used to prepare MRI image datasets containing brain tumor annotation masks, complete dataset partitioning, and perform standardized preprocessing on the MRI images of each dataset to obtain basic preprocessed images. The sparse deconvolution enhancement unit is used to construct a convolution-noise degradation model based on the k-space band-limited sampling characteristics of MRI imaging, and derive the equivalent point spread function corresponding to the MRI image. Taking the basic preprocessed image as input, a sparse deconvolution reconstruction optimization model is constructed that integrates data fidelity term, Hessian continuity prior regularization term and L1 sparse prior regularization term. The parameter strategy adapted to MRI two-dimensional slices is configured and the model is solved iteratively. Sparse deconvolution enhancement processing is uniformly performed on the basic preprocessed images of the training set, validation set and test set to obtain the enhanced MRI image. The segmentation model training unit is used to select a medical image segmentation neural network as a brain tumor segmentation model, build the training framework of the segmentation model and configure a standardized training optimization strategy; the model is trained on training sets with different sample sizes under the same training parameters and data augmentation strategy, using basic preprocessed images and sparse deconvolution enhanced MRI images as model inputs respectively; the optimal weights of the model under each training mode are saved based on the core evaluation indicators of medical image segmentation to obtain the optimal segmentation model. The inference and performance evaluation unit performs standardized basic preprocessing on the test set MRI images in a manner consistent with the training phase. It employs both the original input inference process and the sparse deconvolution enhanced inference process to input the processed images into the corresponding optimal segmentation model, thereby obtaining a brain tumor prediction mask. The image reconstruction effect is evaluated using a contour consistency quantification index, and the model segmentation effect is evaluated using a pixel-level segmentation performance quantification index. The final MRI brain tumor image segmentation result and comprehensive performance evaluation index are then output.
[0056] Example 3 This embodiment 3 also provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement any step of an MRI brain tumor image segmentation method based on sparse deconvolution enhancement preprocessing.
[0057] The computer-readable storage medium may include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0058] For a description of the computer-readable storage medium provided in this application, please refer to the above method embodiments; further details will not be repeated here.
[0059] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for segmenting MRI brain tumor images based on sparse deconvolution enhancement preprocessing, characterized in that, include: S1. Prepare an MRI image dataset containing brain tumor annotation masks, complete the dataset division, and perform standardized preprocessing on the MRI images of each dataset to obtain basic preprocessed images; S2. Construct a convolution-noise degradation model based on the k-space band-limited sampling characteristics of MRI imaging, and derive the equivalent point spread function corresponding to the MRI image; Using preprocessed images as input, a sparse deconvolution reconstruction optimization model is constructed that integrates data fidelity terms, Hessian continuity prior regularization terms, and L1 sparse prior regularization terms. A parameter strategy adapted to MRI two-dimensional slices is configured and the model is solved iteratively. Sparse deconvolution enhancement processing is uniformly performed on the preprocessed images of the training set, validation set, and test set to obtain enhanced MRI images. S3. Select a medical image segmentation neural network as the brain tumor segmentation model, build the training framework of the segmentation model and configure a standardized training optimization strategy; Using preprocessed images and sparse deconvolution-enhanced MRI images as model inputs, respectively, and with unified training parameters and data augmentation strategies, model training was performed on training sets with different sample sizes. The optimal weights of the model under each training mode were saved based on the core evaluation indicators of medical image segmentation to obtain the optimal segmentation model. S4. Perform standardized basic preprocessing on the test set MRI images in the same way as in the training phase. Use the original input inference process and the sparse deconvolution enhanced inference process respectively. Input the processed images into the corresponding optimal segmentation model to obtain the brain tumor prediction mask. Evaluate the image reconstruction effect through contour consistency quantification index and evaluate the model segmentation effect through pixel-level segmentation performance quantification index. Output the final MRI brain tumor image segmentation result and comprehensive performance evaluation index.
2. The MRI brain tumor image segmentation method based on sparse deconvolution enhancement preprocessing according to claim 1, characterized in that, The preprocessing process in S1 includes, but is not limited to: size unification, cropping, and offset correction of the input image; And convert the single-channel image into a 3-channel input according to the network input requirements.
3. The MRI brain tumor image segmentation method based on sparse deconvolution enhancement preprocessing according to claim 1, characterized in that, The equivalent point spread function in S2 is specifically: The MRI spatial domain image is considered as the inverse Fourier transform result after k-space finite bandwidth sampling. Let M(k) be the bandwidth sampling window in k-space. The process of band-limited sampling of the ideal image x and superimposing noise to obtain the observed image is as follows: ; in X(k) is the inverse Fourier transform from k-space to spatial domain, where X(k) is the k-space representation of the ideal image x, and n is the noise term. The process is transformed by an equivalent method, and the equivalent point spread function h is defined as follows: This transforms the bandwidth-limited effect of MRI imaging into a spatial domain convolutional form. , This represents the convolution operation.
4. The MRI brain tumor image segmentation method based on sparse deconvolution enhancement preprocessing according to claim 1, characterized in that, The sparse deconvolution reconstruction optimization model in S2 is specifically as follows: Using the preprocessed image as the image to be reconstructed, a sparse deconvolution reconstruction optimization model is constructed, jointly constrained by a data fidelity term, a Hessian continuity prior regularization term, and an L1 sparse prior regularization term. The expression of this model is: in, For the reconstructed enhanced MRI image, The weighting coefficients for data fidelity items. The pixel value matrix of the preprocessed image as the basis. The convolution operator defined for the equivalent point spread function, To smooth the background item, For the squaring operation of the L2 norm, This is a Hessian-based continuity regularization term used to suppress artifacts caused by noise amplification and improve image structural continuity. The weighting coefficients are for L1 sparse priors. This is an L1 norm operation used to balance the extraction of high-frequency information from images with the preservation of weak structures.
5. The MRI brain tumor image segmentation method based on sparse deconvolution enhancement preprocessing according to claim 1, characterized in that, The iterative solution process in S2 is as follows: First, a parameter strategy adapted to MRI two-dimensional slices is configured for the sparse deconvolution reconstruction optimization model. The parameter strategy includes turning off the background estimation and upsampling steps, setting the axial continuity term to 0, setting the number of sparse reconstruction iterations to no less than 100, and using the RL form for the iterative deconvolution step and configuring a preset number of iterations. Then, based on the parameter strategy, the sparse deconvolution reconstruction optimization model is iteratively solved. During the iteration process, the content-aware parameters of image fidelity, sparsity, and number of deconvolution iterations are determined only on the training set and validation set through grid search and visual inspection. All determined parameters are fixed and reused in the sparse deconvolution enhancement process of the test set. The solution is terminated after completing a preset number of iterations to obtain the enhanced MRI image.
6. The MRI brain tumor image segmentation method based on sparse deconvolution enhancement preprocessing according to claim 1, characterized in that, The original input inference process in S4 uses the basic preprocessed image as input; the sparse deconvolution enhanced inference process uses the enhanced image as input.
7. The MRI brain tumor image segmentation method based on sparse deconvolution enhancement preprocessing according to claim 1, characterized in that, The process of contour consistency quantification in S4 is as follows: The pixel set of the predicted contour after MRI image reconstruction and the pixel set of the real contour are extracted. The 95% Hausdorff distance and the mean surface distance between the two point sets are calculated respectively. The contour consistency is quantitatively evaluated by the two indicators. The 95% Hausdorff distance is the 95th percentile of all unidirectional distances between the predicted contour point set and the true contour point set, used to measure the maximum deviation level of the contour boundary; the average surface distance is the average Euclidean distance between the two sets of contour point sets, used to reflect the overall fitting accuracy of the contour.
8. The MRI brain tumor image segmentation method based on sparse deconvolution enhancement preprocessing according to claim 1, characterized in that, The process of evaluating the model's segmentation effect using pixel-level segmentation performance quantification metrics in S4 is as follows: Based on the pixel-level comparison between the brain tumor prediction mask and the real labeled mask, the number of true positive pixels (TP), the number of false negative pixels (FN), the number of false positive pixels (FP), and the number of true negative pixels (TN) were statistically obtained. The performance of the segmentation model was quantified by four core indicators: Intersection over Union (IoU), Dice similarity coefficient, Sensitivity, and Specificity. The comprehensive quantitative evaluation of the model's segmentation effect was completed through the calculation results of the four indicators. The Intersection over Union (IoU) ratio reflects the degree of overlap between the predicted region and the actual region, and its calculation formula is as follows: ; The Dice similarity coefficient is used to characterize the overlap consistency of segmentation results and has good adaptability to class imbalance scenarios. The calculation formula is as follows: ; Sensitivity measures a model's ability to identify real positive samples, and is calculated using the following formula: ; Specificity measures a model's ability to identify true negative samples, and is calculated using the following formula: ; The values of IoU, Dice similarity coefficient, sensitivity, and specificity are all in the range of [0,1]. The higher the value, the better the model segmentation performance.
9. An MRI brain tumor image segmentation system based on sparse deconvolution enhancement preprocessing, characterized in that, include: The dataset preprocessing unit is used to prepare MRI image datasets containing brain tumor annotation masks, complete dataset partitioning, and perform standardized preprocessing on the MRI images of each dataset to obtain basic preprocessed images. The sparse deconvolution enhancement unit is used to construct a convolution-noise degradation model based on the k-space band-limited sampling characteristics of MRI imaging, and derive the equivalent point spread function corresponding to the MRI image. Using preprocessed images as input, a sparse deconvolution reconstruction optimization model is constructed that integrates data fidelity terms, Hessian continuity prior regularization terms, and L1 sparse prior regularization terms. A parameter strategy adapted to MRI two-dimensional slices is configured and the model is solved iteratively. Sparse deconvolution enhancement processing is uniformly performed on the preprocessed images of the training set, validation set, and test set to obtain enhanced MRI images. The segmentation model training unit is used to select a medical image segmentation neural network as a brain tumor segmentation model, build the training framework of the segmentation model, and configure a standardized training optimization strategy. Using preprocessed images and sparse deconvolution-enhanced MRI images as model inputs, respectively, and with unified training parameters and data augmentation strategies, model training was performed on training sets with different sample sizes. The optimal weights of the model under each training mode were saved based on the core evaluation indicators of medical image segmentation to obtain the optimal segmentation model. The inference and performance evaluation unit performs standardized basic preprocessing on the test set MRI images in a manner consistent with the training phase. It employs both the original input inference process and the sparse deconvolution enhanced inference process to input the processed images into the corresponding optimal segmentation model, thereby obtaining a brain tumor prediction mask. The image reconstruction effect is evaluated using a contour consistency quantification index, and the model segmentation effect is evaluated using a pixel-level segmentation performance quantification index. The final MRI brain tumor image segmentation result and comprehensive performance evaluation index are then output.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is executed by a processor as described in any one of claims 1-8: a method for segmenting MRI brain tumor images based on sparse deconvolution enhancement preprocessing.