Lightweight medical image segmentation method based on re-parameterization and related device

Abstract

The invention discloses a lightweight medical image segmentation method based on re-parameterization and a related device. The method comprises the following steps: acquiring a medical image to be segmented; the to-be-segmented medical image is input into a trained lightweight image segmentation network RepStarNet, a segmentation result of the to-be-segmented medical image is obtained, the lightweight image segmentation network RepStarNet is constructed based on a wavelet module and a re-parameterized Hadamard product module, and the wavelet module and the re-parameterized Hadamard product module are constructed based on the wavelet module and the re-parameterized Hadamard product module. According to the method and the related device, the precision and robustness of medical image segmentation can be improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image processing, and relates to a lightweight medical image segmentation method based on reparameterization and a related device. BACKGROUND

[0002] Most existing medical image segmentation models rely on large convolutional neural networks or transformer structures, which can achieve high accuracy but consume a lot of computing resources, and are not suitable for edge devices and portable medical imaging devices.

[0003] Medical image segmentation, as a core part of computer-aided diagnosis systems, has great significance for improving the efficiency and accuracy of disease diagnosis. In recent years, methods dominated by deep learning, especially the U-shaped encoder-decoder architecture represented by U-Net structure [Ronneberger, O., Fischer, P., & Brox, T. (2015). U-Net: Convolutional Networks for Biomedical Image Segmentation. ArXiv, abs / 1505.04597.], have achieved remarkable success in this field and are widely used in various biomedical imaging modalities, including optical coherence tomography (OCT) image analysis, ultrasound image processing, skin lesion recognition and segmentation, etc. These networks effectively realize pixel-level semantic segmentation through mutually symmetrical encoding (downsampling-feature extraction) and decoding (upsampling-precise positioning) paths.

[0004] However, with the growth of actual clinical application requirements, the computational efficiency, storage consumption and real-time processing capability of the model become key factors[ Z., & Ayd1n, M. (2024). MedSegBench: A comprehensive benchmark for medical image segmentation in diverse data modalities. Scientific Data, 11. To achieve the goal of lightweight deployment, researchers have proposed various improvement strategies to enable the algorithm to run on portable devices, edge computing platforms, or real-time diagnostic systems with limited computing power.Representative works include embedding lightweight backbone networks (e.g., MobileNet [Sinha, D., and El-Sharkawy, M., 2019, October. Thin mobilenet: An enhanced mobilenet architecture. In 2019 IEEE 10th annual ubiquitous computing, electronics & mobile communication conference (UEMCON) (pp. 0280-0285). IEEE.] family) into U-Net [Howard, A. G., Zhu, M., Chen, B., Kalenichenko, D., Wang, W., Weyand, T., Andreetto, M., & Adam, H. (2017). MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications. ArXiv, abs / 1704.04861.], or designing new lightweight structures such as UNeXt [Valanarasu, J. M. J., and Patel, V. M., 2022, September. Unext: Mlp-based rapid medical image segmentation network. In International conference on medical image computing and computer-assisted intervention (pp. 23-33). Cham: Springer Nature Switzerland.] and Rolling-Unet [Liu, Y., Zhu, H., Liu, M., Yu, H., Chen, Z., and Gao, J., 2024, March. Rolling-unet: Revitalizing mlp’s ability to efficiently extract long-distance dependencies for medical image segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 38, No. 4, pp. 3819-3827).].The core innovation of these schemes is to ingeniously use deep separable convolution to greatly reduce the parameter quantity and computational quantity, or introduce attention compression mechanism to reduce the channel redundancy of feature maps, thereby significantly reducing the complexity and inference overhead of the model.

[0005] Although these lightweight efforts have made positive progress, there are still two important limitations in existing research:

[0006] The lack of synergy between frequency domain feature enhancement and model structure optimization: current methods mainly focus on optimizing network structure in the spatial domain (such as reducing the number of layers, channels, or using lightweight operators), and have not fully explored how to purposefully integrate the intrinsic spectral characteristics of images revealed by frequency domain analysis techniques (such as Fourier transform, wavelet transform, etc.) into the design process of the model. Since the frequency domain information of images (such as specific spectral components corresponding to edge, texture features) is crucial for improving the robustness, noise resistance, and sensitivity to details of segmentation, the lack of systematic and end-to-end collaborative design of efficient frequency domain feature extraction, enhancement, or guidance mechanism with lightweight network structure optimization may limit the further improvement space of model performance.

[0007] Consistency problem of structure between training and inference stage: Some lightweight techniques (especially some methods that rely on dynamic mechanisms or complex training tricks) have differences in network structure between training stage and deployment / inference stage. For example, some dynamic pruning [Cheng, H., Zhang, M. and Shi, J. Q., 2024. A survey on deep neural network pruning: Taxonomy, comparison, analysis, and recommendations. IEEE Transactions on Pattern Analysis and Machine Intelligence.] applied during training, channel random dropout [Salehin, I. and Kang, D. K., 2023. A review on dropout regularization approaches for deep neural networks within the scholarly domain. Electronics, 12(14), p. 3106.] or special regularization [Santos, C. F. G. D. and Papa, J. P., 2022. Avoiding overfitting: A survey on regularization methods for convolutional neural networks. ACM Computing Surveys (Csur), 54(10s), pp. 1-25.] methods may be disabled or simplified during inference. This inconsistency in structure / behavior may lead to unpredictable gaps between the performance of the final deployed model and the performance observed on the training / validation set, reducing the reliability and predictability of the model when applied in practice, and bringing additional challenges to the optimization and debugging of the model.

[0008] In summary, although there have been many effective attempts in the research of lightweight models for medical image segmentation, how to more deeply integrate image frequency domain characteristics to guide network design and ensure strict structural consistency between training and inference stages remains a key scientific problem and technical challenge that needs to be solved. Solving these problems is crucial for developing the next generation of medical image AI tools that are both efficient and accurate, and truly have clinical practical value.

[0009] Based on the foregoing analysis, the existing medical image lightweight segmentation model technology has the following core shortcomings:

[0010] Cooperative utilization of video domain information: existing lightweight methods mainly focus on spatial domain structure pruning (such as reducing depth, width, and using lightweight operators), and have not effectively integrated key frequency domain information of images (especially high-frequency details). High-frequency features (such as tissue boundaries and lesion textures) are crucial for accurate segmentation, especially in low-contrast and noisy medical images (such as OCT and ultrasound). Simply relying on spatial domain information limits the model's ability to capture subtle structures and resist noise interference, affecting the accuracy and robustness of segmentation. SUMMARY

[0011] The present application aims to overcome the above-mentioned shortcomings of the prior art and provides a lightweight medical image segmentation method based on reparameterization and related devices, which can improve the accuracy and robustness of medical image segmentation.

[0012] To achieve the above-mentioned purpose, the present application discloses a lightweight medical image segmentation method based on reparameterization, comprising:

[0013] Obtaining a medical image to be segmented;

[0014] Inputting the medical image to be segmented into a trained lightweight image segmentation network RepStarNet to obtain the segmentation result of the medical image to be segmented, wherein the lightweight image segmentation network RepStarNet is constructed based on a wavelet module and a reparameterization Hadamard product module.

[0015] The lightweight medical image segmentation method based on reparameterization of the present application is further improved in that:

[0016] Further, the process of inputting the medical image to be segmented into the trained lightweight image segmentation network RepStarNet is:

[0017] Obtaining a training data set;

[0018] Constructing a lightweight image segmentation network RepStarNet;

[0019] Training the lightweight image segmentation network RepStarNet based on the training data set to obtain a trained lightweight image segmentation network RepStarNet.

[0020] Further, the wavelet module includes a plurality of branches of 1x1 convolution and batch normalization layers, and during the training of the lightweight image segmentation network RepStarNet based on the training data set, the 1x1 convolution and batch normalization layers of each branch are trained respectively.

[0021] Further, the inputting the medical image to be segmented into the trained lightweight image segmentation network RepStarNet further comprises:

[0022] The 1x1 convolution and batch normalization layer of each branch in the trained lightweight image segmentation network RepStarNet are fused.

[0023] The application discloses a lightweight medical image segmentation system based on reparameterization, comprising:

[0024] The acquisition module is configured to acquire a medical image to be segmented.

[0025] The segmentation module is configured to input the medical image to be segmented into a trained lightweight image segmentation network RepStarNet to obtain a segmentation result of the medical image to be segmented.

[0026] The application further improves the lightweight medical image segmentation system based on reparameterization, and the improvement lies in that:

[0027] Further, the application further comprises:

[0028] The second acquisition module is configured to acquire a training data set.

[0029] The construction module is configured to construct a lightweight image segmentation network RepStarNet.

[0030] The training module is configured to train the lightweight image segmentation network RepStarNet based on the training data set to obtain a trained lightweight image segmentation network RepStarNet.

[0031] Further, the wavelet module comprises 1x1 convolution and batch normalization layers of several branches, and the 1x1 convolution and batch normalization layers of each branch are trained during the training of the lightweight image segmentation network RepStarNet based on the training data set.

[0032] Further, the inputting the medical image to be segmented into the trained lightweight image segmentation network RepStarNet further comprises:

[0033] The 1x1 convolution and batch normalization layer of each branch in the trained lightweight image segmentation network RepStarNet are fused.

[0034] The application discloses a computer device, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the light-weight medical image segmentation method based on reparameterization when executing the computer program.

[0035] The application discloses a computer readable storage medium, which stores a computer program, wherein the computer program implements the steps of the light-weight medical image segmentation method based on reparameterization when executed by a processor.

[0036] The application has the following beneficial effects:

[0037] The light-weight medical image segmentation method based on reparameterization and the related device have the following beneficial effects: the medical image to be segmented is input into the trained light-weight image segmentation network RepStarNet to obtain the segmentation result of the medical image to be segmented, the light-weight image segmentation network RepStarNet is constructed based on a wavelet module and a reparameterization Hadamard product module, the reparameterization is introduced to improve the recognition ability of edge structure, thereby enhancing the expression ability of the network to complex lesion texture, and the accuracy and robustness of medical image segmentation are improved. BRIEF DESCRIPTION OF DRAWINGS

[0038] The drawings constituting a part of the specification of the application are used to provide further understanding of the application, the illustrative embodiments of the application and the description thereof are used to explain the application, and do not constitute improper limitation on the application. In the drawings:

[0039] Figure 1 It is a network architecture diagram of the application;

[0040] Figure 2a It is a structure diagram of a down-sampling reparameterization Hadamard product module;

[0041] Figure 2b It is a structure diagram of an up-sampling reparameterization Hadamard product module;

[0042] Figure 3 It is a combined schematic diagram of the reparameterization Hadamard product module after training;

[0043] Figure 4 It is a structure diagram of a light-weight segmentation head;

[0044] Figure 5 It is a structure diagram of a light-weight restoration module;

[0045] Figure 6 It is a system structure diagram of the application. DETAILED DESCRIPTION

[0046] Clearly, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments. Based on the embodiments of the present application, all the other embodiments obtained by a person of ordinary skill in the art without creative effort are within the protection scope of the present application.

[0047] In the description of the present application, it should be understood that the terms "comprising" and "including" indicate the presence of the described features, integers, steps, operations, elements, and / or components, but do not exclude one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0048] It should also be understood that the terms used in the present application specification are only for the purpose of describing particular embodiments and are not intended to limit the present application. As used in the present application specification and the appended claims, the singular forms "a", "an" and "the" are intended to include the plural forms unless the context clearly indicates otherwise.

[0049] It should be further understood that the term "and / or" used in the present application specification is intended to mean one or more of any combination of the associated listed items and all possible combinations thereof, and includes these combinations, for example, A and / or B can mean the existence of A alone, the existence of B alone, or the existence of both A and B. In addition, the character " / " in the present application generally represents an "or" relationship between the front and rear associated objects.

[0050] It should be understood that although the terms first, second, third, etc. can be used in the embodiments of the present application to describe preset ranges, etc., these preset ranges should not be limited to these terms. These terms are only used to distinguish the preset ranges from each other. For example, the first preset range can also be referred to as the second preset range, and similarly, the second preset range can also be referred to as the first preset range without departing from the scope of the embodiments of the present application.

[0051] Depending on the context, the word "if" as used herein can be interpreted as meaning "when" or "while" or "in response to determining" or "in response to detecting". Similarly, depending on the context, the phrase "if determined" or "if detecting (a stated condition or event)" can be interpreted as meaning "when determined" or "in response to determining" or "when detecting (a stated condition or event)" or "in response to detecting (a stated condition or event)".

[0052] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present application provided in the drawings is not intended to limit the scope of the claimed present application, but only represents selected embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work are within the scope of protection of the present application.

[0053] Various structural schematic diagrams according to the disclosed embodiments of the present application are shown in the drawings. These diagrams are not drawn to scale, in which some details are exaggerated for the purpose of clear expression, and some details can be omitted. The shapes of various regions, layers and their relative size and positional relationship shown in the drawings are only exemplary, and in actuality, there can be deviations due to manufacturing tolerances or technical limitations, and regions / layers with different shapes, sizes and relative positions can be additionally designed by those skilled in the art according to actual needs.

[0054] Embodiment one

[0055] The light-weight medical image segmentation method based on reparameterization according to the present application comprises the following steps:

[0056] 1) Obtain a medical image to be segmented;

[0057] 2) Construct a light-weight image segmentation network RepStarNet, wherein the light-weight image segmentation network is constructed based on a reparameterized Hadamard product module.

[0058] Specifically, referring to Figure 1 , Figure 2a and Figure 2b , the light-weight image segmentation network comprises an input layer, a wavelet module, a convolution layer, a first down-sampling reparameterized Hadamard product module, a second down-sampling reparameterized Hadamard product module, a third down-sampling reparameterized Hadamard product module, a fourth down-sampling reparameterized Hadamard product module, a spatial pyramid pooling layer, a first up-sampling reparameterized Hadamard product module, a second up-sampling reparameterized Hadamard product module, a third up-sampling reparameterized Hadamard product module, a fourth up-sampling reparameterized Hadamard product module, a fifth up-sampling reparameterized Hadamard product module, a light-weight segmentation head and an output layer.

[0059] The input layer is connected with the input end of the wavelet module and the input end of the convolution layer, the output end of the wavelet module is connected with the first input end of the first down-sampling re-parameterized Hadamard product module, the output end of the convolution layer is connected with the second input end of the first down-sampling re-parameterized Hadamard product module and the jump information input end of the fifth up-sampling re-parameterized Hadamard product module, the jump output end of the first down-sampling re-parameterized Hadamard product module is connected with the jump information input end of the fourth up-sampling re-parameterized Hadamard product module, the first output end and the second output end of the first down-sampling re-parameterized Hadamard product module are connected with the first input end and the second input end of the second down-sampling re-parameterized Hadamard product module, the jump output end of the second down-sampling re-parameterized Hadamard product module is connected with the jump information input end of the third up-sampling re-parameterized Hadamard product module, the first output end and the second output end of the second down-sampling re-parameterized Hadamard product module are connected with the first input end and the second input end of the third down-sampling re-parameterized Hadamard product module, the jump output end of the third down-sampling re-parameterized Hadamard product module is connected with the jump information input end of the second up-sampling re-parameterized Hadamard product module, the first output end and the second output end of the third down-sampling re-parameterized Hadamard product module are connected with the first input end and the second input end of the fourth down-sampling re-parameterized Hadamard product module, the jump output end of the fourth down-sampling re-parameterized Hadamard product module is connected with the jump information input end of the first up-sampling re-parameterized Hadamard product module, the output end of the fourth down-sampling re-parameterized Hadamard product module is connected with the input end of the spatial pyramid pooling layer, the output end of the intermediate pyramid pooling layer is connected with the input end of the first up-sampling re-parameterized Hadamard product module, the output end of the first up-sampling re-parameterized Hadamard product module is connected with the input end of the second up-sampling re-parameterized Hadamard product module, the output end of the second up-sampling re-parameterized Hadamard product module is connected with the input end of the third up-sampling re-parameterized Hadamard product module, the output end of the third up-sampling re-parameterized Hadamard product module is connected with the input end of the fourth up-sampling re-parameterized Hadamard product module, the output end of the fourth up-sampling re-parameterized Hadamard product module is connected with the input end of the fifth up-sampling re-parameterized Hadamard product module, the output end of the fifth up-sampling re-parameterized Hadamard product module is connected with the input end of the lightweight segmentation head, and the output end of the lightweight segmentation head is connected with the output layer.

[0060] It should be noted that the down-sampling re-parameterized Hadamard product module down-samples the input image to a low resolution and extracts information, the spatial pyramid pooling layer is used to enhance the information transmitted by the encoder to the decoder, and then the up-sampling re-parameterized Hadamard product module is used to restore the information, and finally the lightweight segmentation head is used to decode the information and obtain the segmentation result.

[0061] For the wavelet module WFM:

[0062] Because the lesion region in medical images often has fuzzy boundaries or low contrast, it is difficult for traditional convolutional neural networks to extract high-frequency detailed features. WFM decomposes the image into multiple parts through two-dimensional DWT, thereby introducing frequency domain information to enhance the structural expression capability. The implementation process of the wavelet module is as follows:

[0063] DWT decomposition: after wavelet transformation of the input image I, the following is obtained:

[0064] I LL ,I LH ,I HL ,I HH = DWT(I)

[0065] where I LL is the low-frequency part for retaining structural information; I LH ,I HL ,I HH are high-frequency details, representing horizontal, vertical and diagonal texture information respectively.

[0066] Frequency enhancement: to enhance edge information, the high-frequency part is weighted using the coefficient γ to obtain the enhanced image I wavelet :

[0067] I wavelet = I LL + γ(I LH + I HL + I HH )

[0068] Set γ = 3 to improve edge contrast.

[0069] Dual-input encoder: the main encoder inputs the original image, and the secondary encoder inputs the enhanced image I wavelet , and the features of the two paths are fused in the middle and later layers.

[0070] For the RepStarBlock reparameterization Hadamard product module, refer to Figure 3 To balance the expression ability in the training stage and the efficiency in the inference stage, the reparameterization Hadamard product module introduces a reparameterization mechanism to realize the free switching and structural consistency of training and inference. The specific steps of reparameterization are as follows:

[0071] The convolution kernels and BN parameters of each branch are fused into a set of equivalent weights and

[0072]

[0073] where γ i is the scaling factor, β i is the bias offset, μ i is the mean, is a small constant to avoid the denominator being zero.

[0074] Fusing each branch into:

[0075]

[0076] During training, 1x1 convolution and batch normalization (BN) operation of multiple branches are adopted; due to the introduction of reparameterization, during inference, based on the principle of equivalent combination of convolution, the 1x1 convolution of multiple branches is combined and added to obtain single branch 1x1 convolution. In both stages, the results of different branches are enhanced through Hadamard product, and finally the effect of feature enhancement is achieved, and the purpose of lightweight inference of the end side device is realized.

[0077] In the above manner, the frequency domain features can be introduced as auxiliary information, the edge structure recognition capability is significantly improved, the expression capability of the lightweight network on complex lesion texture is enhanced, and the module does not introduce additional parameters, the calculation cost is small, and is suitable for edge devices.

[0078] For the lightweight segmentation head AdaSeg Head, reference Figure 4 , the traditional segmentation head generally uses 3x3 regular convolution to compress the information of the last layer. The lightweight segmentation head described in the application extracts information by using channel separation convolution (each channel is processed by 3x3 convolution to obtain enhanced spatial information) and point-by-point convolution (all values at a position are enhanced), and finally obtains a segmentation mask.

[0079] When there is input data on the left side, via two branches, one is the input information itself; the other sequentially passes through channel separation convolution, point-by-point convolution and an activation function to obtain enhanced information; then, the enhanced information is multiplied by the original information to further enhance the input information; finally, via point-by-point convolution, the channel is compressed to a single channel and via an activation function (Sigmoid) to obtain a mask between 0 and 1.

[0080] For the lightweight restoration module Lite-Deconv, reference Figure 5 , the traditional deconvolution has problems such as complex calculation and easy to produce chessboard effect, therefore, the application uses bilinear interpolation to abolish deconvolution, and efficiently restores information without additional parameter quantity.

[0081] 3) input the medical image to be segmented into the lightweight image segmentation network RepStarNet to obtain a segmentation result.

[0082] Simulation experiment

[0083] The application has been tested on five medical image datasets, including OCT, ultrasound and skin lesion images. Experimental results show that the application can achieve high-precision segmentation performance under the condition of only 3.78M parameters, and the actual measurement frame rate is 18.6 on devices such as RK3588, the real-time inference of iphone15 reaches 289.01 per second, and the frame rate reaches 289.01 per second under the input of 320 size.

[0084] Embodiment two

[0085] Reference Figure 6 The light-weight medical image segmentation system based on reparameterization provided by the application comprises:

[0086] The acquisition module is configured to acquire a medical image to be segmented.

[0087] The segmentation module is configured to input the medical image to be segmented into the trained light-weight image segmentation network RepStarNet to obtain a segmentation result of the medical image to be segmented.

[0088] In this embodiment, the light-weight medical image segmentation system based on reparameterization further comprises:

[0089] The second acquisition module is configured to acquire a training data set.

[0090] The construction module is configured to construct a light-weight image segmentation network RepStarNet.

[0091] The training module is configured to train the light-weight image segmentation network RepStarNet based on the training data set to obtain a trained light-weight image segmentation network RepStarNet.

[0092] In this embodiment, the wavelet module comprises a plurality of branches of 1x1 convolution and batch normalization layer, and during the training of the light-weight image segmentation network RepStarNet based on the training data set, the 1x1 convolution and batch normalization layer of each branch are trained respectively.

[0093] In this embodiment, before the medical image to be segmented is input into the trained light-weight image segmentation network RepStarNet, the method further comprises:

[0094] The 1x1 convolution and batch normalization layer of each branch in the trained light-weight image segmentation network RepStarNet are fused.

[0095] The division of the modules in the embodiments of the present application is illustrative, and is only a logical function division. In actual implementation, another division manner can be used. In addition, the function modules in each of the embodiments of the present application can be integrated in one processor, or can be physically separated, or two or more modules can be integrated in one module. The integrated module can be implemented in the form of hardware or in the form of a software function module.

[0096] Embodiment three

[0097] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the steps of the light-weight medical image segmentation method based on reparameterization when executing the computer program, for example, including: obtaining a to-be-segmented medical image; inputting the to-be-segmented medical image into a trained light-weight image segmentation network RepStarNet to obtain a segmentation result of the to-be-segmented medical image, wherein the light-weight image segmentation network RepStarNet is constructed based on a wavelet module and a reparameterization Hadamard product module. The memory can include a memory, for example, a high-speed random memory, and can also include a non-volatile memory, for example, at least one disk memory, etc.; the processor, the network interface, and the memory are connected to each other through an internal bus, which can be an industry standard architecture bus, a peripheral component interconnect standard bus, an extended industry standard structure bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. The memory is used to store programs, specifically, the programs can include program codes, and the program codes include computer operation instructions. The memory can include a memory and a non-volatile memory, and provide instructions and data for the processor.

[0098] Embodiment four

[0099] A computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps of the light-weight medical image segmentation method based on reparameterization, for example, including: obtaining a to-be-segmented medical image; inputting the to-be-segmented medical image into a trained light-weight image segmentation network RepStarNet to obtain a segmentation result of the to-be-segmented medical image, wherein the light-weight image segmentation network RepStarNet is constructed based on a wavelet module and a reparameterization Hadamard product module. Specifically, the computer readable storage medium includes but is not limited to, for example, a volatile memory and / or a non-volatile memory. The volatile memory can include a random access memory (RAM) and / or a cache memory, etc. The non-volatile memory can include a read-only memory (ROM), a hard disk, a flash memory, an optical disk, a magnetic disk, etc.

[0100] Those skilled in the art will appreciate that embodiments of the present application can be readily used as a method, a system or a computer program product. Accordingly, the present application can take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) embodying computer readable program code thereon for use by or in connection with an instruction execution system. For the purposes of this description, a computer-usable or computer readable storage medium can be any apparatus that can contain, store, communicate, propagate, or transport the program for use by or in connection with the instruction execution system, apparatus, or device. The computer-usable storage medium can be a computer- readable storage medium that can be any media that can be accessed by the computer. By way of example, and not limitation, such computer-readable media can include RAM, ROM, EEPROM, CD-ROM or other optical storage, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and that can be accessed by a computer. Also, any connection is properly termed a computer- readable medium. For example, if the software is transmitted from a website, server, or other remote source using a coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or

[0101] The present application is described herein with reference to the flowchart and / or block diagram illustrations that are used to describe the methods, apparatus (systems) and computer program products according to the embodiments of the present application. It will be understood that each block of the flowchart and / or block diagrams, and combinations of blocks in the flowchart and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processing system or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart and / or block diagram block(s) or a combination of flowchart and / or block diagram block(s). These computer program instructions can also be stored in a computer readable storage medium that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer readable storage medium produce an article of manufacture including instructions which implement the function specified in the flowchart and / or block diagram block(s) or a combination of flowchart and / or block diagram block(s). Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the described functions of one or more flowcharts and / or blocks

[0102] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart and / or block diagram block(s) or a combination of flowchart and / or block diagram block(s). Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the described functions of one or more flowcharts and / or blocks

[0103] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart and / or block diagram block(s) or a combination of flowchart and / or block diagram block(s). Figure 1 one or more flowcharts and / or blocks Figure 1 means for functionally implementing the described functions of one or more flowcharts and / or blocks

[0104] Other embodiments of the present application will be apparent to those skilled in the art from consideration of the specification and practice of the application disclosed herein. It is intended that the present application cover any and all variations of the present application that come within the scope of the claims and their equivalents. It is intended that the specification and examples be considered exemplary only, with the true scope and spirit of the application being indicated by the following claims.

[0105] It should be understood that the application is not limited to the precise construction which has been described above and illustrated in the accompanying drawings, and that various modifications and changes can be made by those skilled in the art without departing from the scope of the application. The scope of the application should only be limited by the appended claims.

[0106] The above description is only the preferred embodiment of the present application, not any limitation to the present application, any simple modification, change and equivalent structure change of the above embodiment according to the technical essence of the present application are still within the protection scope of the technical solution of the present application.

Claims

1. A light-weight medical image segmentation method based on reparameterization, characterized in that, The method comprises the following steps: obtaining a medical image to be segmented; inputting the medical image to be segmented into a trained lightweight image segmentation network RepStarNet to obtain a segmentation result of the medical image to be segmented, wherein the lightweight image segmentation network RepStarNet is constructed based on a wavelet module and a reparameterization Hadamard product module. 2.The light-weight medical image segmentation method based on reparameterization according to claim 1, wherein, Before the medical image to be segmented is inputted into the trained lightweight image segmentation network RepStarNet, the following steps are further included: obtaining a training data set; constructing a lightweight image segmentation network RepStarNet; training the lightweight image segmentation network RepStarNet based on the training data set to obtain a trained lightweight image segmentation network RepStarNet.

3. The reparameterization-based light-weight medical image segmentation method according to claim 2, characterized in that, The wavelet module includes a plurality of branches of 1x1 convolution and batch normalization layers, and each branch of 1x1 convolution and batch normalization layer is trained during the training of the lightweight image segmentation network RepStarNet based on the training data set.

4. The light-weight medical image segmentation method based on reparameterization according to claim 3, characterized in that, Before the medical image to be segmented is inputted into the trained lightweight image segmentation network RepStarNet, the following steps are further included: fusing each branch of 1x1 convolution and batch normalization layer in the trained lightweight image segmentation network RepStarNet.

5. A reparameterization-based lightweight medical image segmentation system, characterized by, The method comprises the following steps: a first obtaining module for obtaining a medical image to be segmented; a segmentation module for inputting the medical image to be segmented into a trained lightweight image segmentation network RepStarNet to obtain a segmentation result of the medical image to be segmented, wherein the lightweight image segmentation network RepStarNet is constructed based on a wavelet module and a reparameterization Hadamard product module.

6. The light-weight medical image segmentation system based on reparameterization according to claim 5, wherein, Further comprising: a second obtaining module for obtaining a training data set; a construction module for constructing a lightweight image segmentation network RepStarNet; a training module for training the lightweight image segmentation network RepStarNet based on the training data set to obtain a trained lightweight image segmentation network RepStarNet.

7. The light-weight medical image segmentation system based on reparameterization according to claim 6, wherein, The wavelet module includes a plurality of branches of 1x1 convolution and batch normalization layers, and each branch of 1x1 convolution and batch normalization layer is trained during the training of the lightweight image segmentation network RepStarNet based on the training data set.

8. The light-weight medical image segmentation system based on reparameterization according to claim 7, wherein, Before the medical image to be segmented is inputted into the trained lightweight image segmentation network RepStarNet, the following steps are further included: fusing each branch of 1x1 convolution and batch normalization layer in the trained lightweight image segmentation network RepStarNet.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor executes the computer program to realize the steps of the reparameterization-based lightweight medical image segmentation method according to any one of claims 1-4.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. The computer program is executed by the processor to realize the steps of the reparameterization-based lightweight medical image segmentation method according to any one of claims 1-4.

Images