CN-115239950-B - Medical image segmentation method, system, equipment and computer readable medium

CN115239950BCN 115239950 BCN115239950 BCN 115239950BCN-115239950-B

Abstract

The invention discloses a medical image segmentation method which comprises the steps of preprocessing an original medical image, inputting the preprocessed medical image into a coding-decoding network structure to extract image features, processing the image features output by each decoder stage in the coding-decoding network structure, enabling the image features of any decoder stage except the deepest decoder stage to perform self-distillation learning on the image features of all the decoder stages in the deeper layers, performing computationally intensive imitation loss, combining corresponding activation functions according to the image features of an output layer in the coding-decoding network structure to obtain a prediction result, combining target masks according to the prediction result to calculate loss, and updating network parameters through a back propagation algorithm. The invention also discloses a medical image segmentation system, a device and a computer readable medium. The invention can effectively improve the performance of the medical image segmentation network under the condition of not increasing or obviously increasing parameters.

Inventors

GAO YING
XIE LINSEN
CAI WENTIAN

Assignees

华南理工大学

Dates

Publication Date: 20260512
Application Date: 20220331

Claims (12)

1. A medical image segmentation method, comprising: preprocessing an original medical image; inputting the preprocessed medical image into an encoding-decoding network structure to extract image features; The method comprises the steps of obtaining image characteristics output by each decoder stage, obtaining a corresponding characteristic diagram according to the image characteristics output by each decoder stage, taking the characteristic diagram corresponding to the deeper decoder stage as a label between any two characteristic diagrams of all decoder stages, calculating the imitation loss according to a loss function, adding all imitation losses to obtain the imitation loss, wherein the decoder stage comprises a multi-scale characteristic extraction flow consisting of a plurality of continuous multi-scale characteristic extraction methods, and the multi-scale characteristic extraction method comprises the steps of splitting the input image characteristics into a plurality of groups, inputting the image characteristics of each group into different branches, carrying out branch processing, namely respectively carrying out different scale convolution characteristic extraction on the image characteristics input into each branch, and keeping the number of the characteristics to the different branches, and splicing the image characteristics to the image with the different channels, and obtaining the image characteristics by the same channel, and splicing the image characteristics by the channel, wherein the channel is provided with the different channel characteristics; Combining corresponding activation functions according to image features of an output layer in the encoding-decoding network structure to obtain a prediction result; calculating loss according to the prediction result and combining with a target mask; network parameters of the encoding-decoding network structure are updated by a back propagation algorithm.
2. The medical image segmentation method as set forth in claim 1, further comprising: And selecting a characteristic diagram corresponding to at least part of decoder stages, respectively carrying out loss calculation with the input real labels of the medical images, and adding all the results to obtain target perception loss.
3. A medical image segmentation method according to claim 2, wherein: the feature map for the target perceptual loss employs a feature map corresponding to a deeper partial decoder stage.
4. A medical image segmentation method according to claim 1, wherein: The preprocessing includes at least one of data enhancement, control input range, control input data distribution.
5. A medical image segmentation method according to claim 1, wherein: after said and passing it back, it further comprises: channel aliasing, namely rearranging the output features in a manner that disrupts the channel order.
6. The medical image segmentation device is characterized by comprising a data preprocessing module, a data segmentation module and a data segmentation module, wherein the data preprocessing module is used for preprocessing an original medical image; The feature extraction module is used for inputting the preprocessed medical image into the encoding-decoding network structure to extract image features; The self-distillation learning module is used for processing the image characteristics output by each decoder stage in the coding-decoding network structure, so that the image characteristics of any decoder stage except the deepest decoder stage are self-distilled learned to the image characteristics of all the decoder stages in the deeper layers, and the loss is simulated in a computationally intensive manner; the method comprises the steps of obtaining image features output by each decoder stage, obtaining corresponding feature images according to the image features output by each decoder stage, calculating imitation losses according to a loss function by taking the feature images corresponding to the deeper decoder stages as labels between any two feature images of all decoder stages, adding all imitation losses to obtain dense imitation losses, wherein the decoder stage comprises a multi-scale feature extraction flow consisting of a plurality of continuous multi-scale feature extraction methods, the multi-scale feature extraction method comprises the following steps of splitting channels, dividing the input image features into a plurality of groups, inputting the image features of each group into different branches, carrying out branch processing, namely respectively carrying out convolution feature extraction processing on the image features input into each branch, keeping the number of feature channels unchanged in the process, and carrying out channel splicing on the image features output by different branches according to channels to obtain spliced image features, wherein the spliced image features are equal to the number of channels possessed by the input image features, and carrying out channel-by-channel addition on the spliced image features, and gradually transmitting the spliced image features to the input image features to the image features; the result prediction module is used for obtaining a prediction result according to the image characteristics of the output layer in the encoding-decoding network structure and the corresponding activation function; the loss calculation module is used for calculating loss according to the prediction result and combining with a target mask; and the back propagation module is used for updating the network parameters of the encoding-decoding network structure through a back propagation algorithm.
7. The medical image segmentation apparatus as set forth in claim 6, wherein: The self-distillation learning module includes: the feature acquisition module is used for acquiring the image features output by each decoder stage; The feature map acquisition module is used for acquiring a corresponding feature map according to the image features output by each decoder stage; And the intensive imitation loss module is used for calculating imitation losses according to the loss function by taking the feature graphs corresponding to the deeper decoder stages as labels between any two feature graphs of all decoder stages, and adding all imitation losses to obtain the intensive imitation losses.
8. The medical image segmentation apparatus as set forth in claim 7, wherein: The self-distillation learning module further comprises: And the target perception loss module is used for selecting at least part of characteristic diagrams corresponding to the decoder stage, respectively carrying out loss calculation with the input real labels of the medical images, and adding all the results to obtain target perception loss.
9. The medical image segmentation apparatus as set forth in claim 6, wherein: the decoder comprises a multi-scale feature extraction module, wherein the multi-scale feature extraction module comprises a plurality of multi-scale feature extraction units which are connected in sequence, and the multi-scale feature extraction unit comprises: the channel splitting module is used for dividing the input image features into a plurality of groups, wherein the image features of each group are input into different branches; The branch processing module is used for respectively carrying out convolution characteristic extraction processing on the image characteristics input into each branch and keeping the characteristic channel number unchanged in the process, and different branches have different pooling sizes; The channel stitching module is used for stitching the image features output by different branches according to channels to obtain stitched image features, and the stitched image features and the input image features have the same channel number; And the output module is used for adding the spliced image features and the input image features channel by channel and pixel by pixel to obtain output features and transmitting the output features backwards.
10. A medical image segmentation apparatus as set forth in claim 9 wherein: the decoder further comprises a channel confusion module located between adjacent multi-scale feature extraction units; The channel confusion module is used for rearranging the output characteristics in a mode of disturbing the channel sequence.
11. A medical image segmentation apparatus comprising: one or more processors; a storage device having one or more programs stored thereon; the one or more programs, when executed by the one or more processors, cause the one or more processors to implement a medical image segmentation method as set forth in any one of claims 1-5.
12. A computer readable medium having stored thereon a computer program, wherein the program when executed by a processor implements a medical image segmentation method according to any of claims 1-5.

Description

Medical image segmentation method, system, equipment and computer readable medium Technical Field The invention belongs to the technical field of medical image processing, and particularly relates to a medical image segmentation method, a system, equipment and a computer readable medium. Background Medical images are important bases for doctors to judge the illness state of patients and to assign diagnosis and treatment schemes. When observing medical images, focus and medical targets are focused, but medical images are difficult to observe, time and effort are consumed, and individual differences exist in the professional experience of doctors depending on professional medical knowledge. The automatic medical image segmentation can objectively and rapidly extract the required targets in the medical image, can relieve the diagnosis pressure of doctors to a certain extent, and reduces missed diagnosis and misdiagnosis caused by subjective factors of the doctors. The traditional medical image segmentation method has poor effect and low generalization. The application of deep learning models in recent years has greatly improved the accuracy of medical image segmentation. Among them, the encoding-decoding structure represented by U-Net has become a mainstream scheme for medical image segmentation. However, although the U-Net method has great universality, the method is limited by the problems of low precision, low contrast, large image difference, low network depth, variable target scale and the like of the medical image, and the precision is difficult to further improve. Aiming at the problems of shallower network depth and changeable target scale in medical images existing in a medical image segmentation method based on deep learning, a plurality of attempts are made to solve the problems of the two problems in the existing method, but the method often brings a huge amount of additional parameters, and a certain improvement space still exists for the performance of the method. For example, the use of deeper and more powerful base networks can improve the segmentation performance to some extent, but the parameters are often huge, the training difficulty is high, and the equipment burden is high. Existing multi-scale methods also tend to extract multi-scale features through additional modules, which typically have multiple branches, bringing about a large number of additional parameters. For example, the pyramid pooling module in PSPNet is used for pooling input features into a plurality of scales, extracting features respectively, and finally fusing the features together to improve the adaptability to a multi-scale target, and the DeepLab series network is used for extracting information of the plurality of scales on a fixed input scale in a cavity convolution mode, and the cavity convolution is characterized in that the intervals among convolution kernel parameters are enlarged, so that the effects of enlarging the convolution kernel size and covering a larger data range are achieved. The above operations are performed on the complete input features and similar structures typically have multiple branches, the number of parameters for one such multi-scale module will be several times the number of parameters required for a common convolutional layer. Disclosure of Invention In order to overcome the defects in the prior art, the invention provides a medical image segmentation method, a system, equipment and a computer readable medium, which can effectively improve the performance of a medical image segmentation network and ensure that parameters are not increased or are not obviously increased. In a first aspect, a medical image segmentation method includes: preprocessing an original medical image; inputting the preprocessed medical image into an encoding-decoding network structure to extract image features; Processing the image characteristics output by each decoder stage in the coding-decoding network structure, so that the image characteristics of any decoder stage except the deepest decoder stage are subjected to self-distillation learning to the image characteristics of all the deeper decoder stages, and the loss is simulated in a computationally intensive manner; Combining corresponding activation functions according to image features of an output layer in the encoding-decoding network structure to obtain a prediction result; calculating loss according to the prediction result and combining with a target mask; network parameters of the encoding-decoding network structure are updated by a back propagation algorithm. Preferably, the processing the image feature output by each decoder stage in the encoding-decoding network structure makes the image feature of any decoder stage except the deepest decoder stage perform self-distillation learning on the image features of all the decoder stages in the deeper layers, and the method includes the following steps: acquiring image characteristics output by each decod