CN-121982025-A - Uterus image evaluation system combining U-shaped network and modal missing segmentation method

CN121982025ACN 121982025 ACN121982025 ACN 121982025ACN-121982025-A

Abstract

The invention discloses a uterus image evaluation system combining a U-shaped network and a modal missing segmentation method in the technical field of medical image recognition, which comprises the steps of acquiring a T1w-FS modal image and a T2w-FS modal image as bimodal input data, respectively inputting the bimodal image into a dual-branch residual error encoder to extract multi-scale modal characteristics, carrying out unified characteristic channel and intensity distribution alignment treatment, combining a cross-modal residual error alignment module to realize spatial characteristic calibration, combining modal existence marks to realize self-adaptive fusion under the condition of modal missing, reinforcing key area characteristics through a modal specific texture attention block, then carrying out layer-by-layer up sampling and restoring spatial resolution through a decoder, and constructing a joint loss function for optimization in a training stage.

Inventors

XING WEN
SUN BAIYUN

Assignees

核工业总医院

Dates

Publication Date: 20260505
Application Date: 20260403

Claims (10)

1. A uterus image evaluation system combining a U-shaped network and a modal missing segmentation method is characterized by comprising the following steps: inputting the acquired T1w-FS mode image and T2w-FS mode image as dual-mode input data into a dual-branch residual error encoder, respectively extracting mode scale features, and carrying out unified feature channel and alignment intensity distribution treatment; The processed modal scale features are input into a cross-modal residual alignment module together, a first alignment path from the T1w-FS modal feature to the T2w-FS modal feature and a second alignment path from the T2w-FS modal feature to the T1w-FS modal feature are respectively constructed through bidirectional residual mapping, spatial calibration of the bimodal scale features is achieved through the first alignment path and the second alignment path, and adaptive fusion is carried out on the bimodal scale features by combining with modal existence marks, so that fusion features are obtained; Inputting the fusion features into a modal specific texture attention block, extracting high-frequency texture information of each mode through Gaussian filtering difference, generating an attention weight graph according to the obtained high-frequency texture information, and carrying out weighted enhancement on the fusion features according to the attention weight graph to obtain an enhanced feature graph; inputting the enhanced feature map into a decoder, carrying out layer-by-layer up sampling through transpose convolution and splicing with the corresponding scale features, recovering the feature map, and outputting a class probability map corresponding to each pixel through a segmentation head; Constructing a joint loss function according to the class probability map, and training each module parameter in the uterine image segmentation method by utilizing the joint loss function so as to obtain a final uterine image segmentation result.
2. The uterine image assessment system combining a U-shaped network and a modality deficiency segmentation method according to claim 1, wherein the dual-branch residual encoder comprises a first branch encoder and a second branch encoder, both of which are provided with 5-layer residual blocks, The first branch encoder and the second branch encoder are respectively provided with a first layer convolution layer in front of the first layer residual error block and are used for mapping an original input image into initial characteristics; Characterised by the initial feature As input feature, the first branch encoder The output characteristics of the layer residual block are: ; Characterised by the initial feature As input feature, the second branch encoder The output characteristics of the layer residual block are: ; Wherein, the Representing the first branch encoder The output characteristics of the layer residual block; Representing the first branch encoder The output characteristics of the layer residual block, Representing a second constituent encoder The output characteristics of the layer residual block, Representing a second constituent encoder The output characteristics of the layer residual block, Representation of The operation of the convolution is performed, For the purpose of batch normalization, To activate the function.
3. The uterine image assessment system combining the U-shaped network and the mode missing segmentation method according to claim 2, wherein a mode calibration layer is arranged behind each layer of residual block, and the mode calibration layer is expressed as: ; Wherein, the Representation of The convolution operation maps the bimodal feature to the same channel dimension.
4. A uterine image assessment system in combination with a U-network and a modality-missing segmentation method according to claim 3, characterized in that the first alignment path of the cross-modality residual alignment module And a second pair Ji Lujing of Expressed as: ; Wherein, the Representation of Is performed by the convolution operation of (a).
5. The uterine image assessment system combining a U-shaped network and a modality absence segmentation method of claim 4, wherein the modality presence flag is set to @ , ); Construction of fusion features Expressed as: 。
6. The uterine image assessment system combining a U-shaped network and a modality miss segmentation method of claim 5, wherein when bi-directional residual fusion alignment features By aligning functions in the case of containing only valid information of one modality or in the case of containing complete modality information Feature alignment is achieved: ; Wherein, the The feature stitching is represented and is performed, Is that Norms.
7. The uterine image evaluation system combining a U-shaped network and a mode missing segmentation method according to claim 6, wherein the extracting of each mode high frequency texture information by gaussian filtering difference is expressed as: = ; = ; Wherein, the Corresponding to the high frequency characteristics of the T1w-FS mode image, Corresponding to the high frequency characteristics of the T2w-FS mode image, Is the mode characteristic of the T1w-FS mode image, Is a modal feature of the T2w-FS modal image, For a two-dimensional gaussian filtering operation, And Parameters of two-dimensional Gaussian filtering respectively; The generation of the attention weight graph according to the acquired high-frequency texture information is respectively expressed as follows: ; Wherein, the An attention weighting map corresponding to the T1w-FS mode image, An attention weight map corresponding to the T2w-FS mode image; representation downsampling to Scale, meaning that the first branch encoder and the second branch encoder are at the first The layer residual block output feature map size is Wherein Representation of Is provided with a height of (1), Representation of Through a width of step 2 The maximum pooling enables downsampling to be performed, Is a Sigmoid function; The weighted enhancement of the fusion features according to the attention weight graph comprises the steps of multiplying the attention weight graph with the aligned fusion features element by element to strengthen the key region features, wherein the formula is as follows: ; Wherein, the In order to enhance the characteristics of the fusion after it has been enhanced, Is an element-by-element multiplication operation.
8. The uterine image evaluation system combining the U-shaped network and the modality absence segmentation method according to claim 7, wherein when only one piece of acquired high-frequency texture information is obtained, an attention weight map is generated from the acquired high-frequency texture information, expressed as: ; Wherein, the Indicating KL divergence loss.
9. The uterine image assessment system combining a U-shaped network and a modality deficiency segmentation method of claim 8, wherein the transposed convolution upsamples layer by layer at a first level The layer up-sampling formula is: ; Wherein, the Is that Transpose convolution, will Layer scale of Is up-sampled to the feature map of (1) Layer scale of ; The sum corresponding scale feature splice is expressed as: ; the outputting the class probability map P corresponding to each pixel by the segmentation head comprises the following steps of Feeding a lightweight dividing head: ; Wherein, the Is that The convolution is performed with the result that, Representation of Is used in the convolution operation of (1), For the last layer of output features of the decoder, To output various probabilities.
10. The uterine image assessment system combining a U-shaped network and a modality deficiency segmentation method of claim 9, wherein the joint loss function Expressed as: ; Wherein, the 、、 In order to lose the weight of the weight, As the number of residual layers, For the loss of the spatial domain segmentation, Is a frequency domain boundary loss; The spatial domain segmentation loss Expressed as: ; Wherein, the In order to smooth the term(s), Is a pixel Belongs to the category of Is a function of the probability of (1), In order to correspond to the actual tag(s), Representing a segmentation class; the frequency domain boundary loss Expressed as: ; ; ; wherein P is a class probability map corresponding to each pixel output by the segmentation head, G is an input image, In the form of a two-dimensional fast fourier transform, In the form of a gaussian low pass filter, Is that Norms.

Description

Uterus image evaluation system combining U-shaped network and modal missing segmentation method Technical Field The invention relates to a uterus image evaluation system combining a U-shaped network and a modal missing segmentation method, and belongs to the technical field of medical image recognition. Background The uterus is used as an important organ in female pelvis, the morphological structure and the anatomical relation between the uterus and surrounding tissues are important basis for gynecological disease diagnosis, treatment scheme formulation and curative effect evaluation, and clinically common diseases such as hysteromyoma, endometriosis, adenomyosis, malignant tumor and the like all need to accurately evaluate the focus range and the overall outline of the uterus through medical images. Magnetic Resonance Imaging (MRI) has become the main means for evaluating uterine disease images because of the advantages of high resolution of soft tissues, no ionizing radiation, excellent contrast and the like, wherein a T1 weighted fat suppression sequence (T1 w-FS) can effectively suppress fat signal interference and highlight the characteristics of signals of blood-rich lesions and hemorrhagic lesions, and a T2 weighted fat suppression sequence (T2 w-FS) can clearly show layered structures such as myometrium, endometrium, cervical and the like and better show boundary relations between uterus, bladder, rectum and other adjacent organs. The bimodal image has obvious complementarity in structure and texture information, and is helpful for improving the accuracy of the uterine region segmentation. In the prior art, a U-shaped network and an improved structure thereof are adopted for end-to-end training, a better segmentation effect can be obtained under the condition of dual-mode complete input, however, in an actual clinical scene, due to the reasons of equipment configuration difference, non-uniformity of a scanning protocol, insufficient patient coordination degree, limited inspection flow and the like, the condition of missing of a certain mode often occurs, the performance of a model which only depends on complete dual-mode data training is obviously reduced when the model is applied, the segmentation result is unstable and even distorted, the inherent problems of intensity distribution difference and spatial characteristic offset exist between different modes, and the direct characteristic fusion is easy to cause information dilution or boundary blurring, so that the complementary advantages of dual-mode images are difficult to fully develop. The existing partial mode missing processing method realizes compatibility by a simple weighting or feature zero filling mode, but cannot effectively solve the problems of cross-mode feature alignment and key texture information reinforcement, and particularly under the conditions of lower contrast between uterus and surrounding soft tissues and fuzzy boundary, the segmentation precision still cannot meet the clinical application requirements. Disclosure of Invention The invention aims to overcome the defects in the prior art, the performance of the existing automatic uterine MRI segmentation method is obviously reduced under the condition of single-mode deletion, and the problems of intensity distribution difference and spatial characteristic deviation among different modes and easy key information dilution and boundary blurring caused by direct fusion are solved. In order to solve the technical problems, the invention is realized by adopting the following technical scheme: the uterus image evaluation system combining the U-shaped network and the modal absence segmentation method comprises the following steps: inputting the acquired T1w-FS mode image and T2w-FS mode image as dual-mode input data into a dual-branch residual error encoder, respectively extracting mode scale features, and carrying out unified feature channel and alignment intensity distribution treatment; The processed modal scale features are input into a cross-modal residual alignment module together, a first alignment path from the T1w-FS modal feature to the T2w-FS modal feature and a second alignment path from the T2w-FS modal feature to the T1w-FS modal feature are respectively constructed through bidirectional residual mapping, spatial calibration of the bimodal scale features is achieved through the first alignment path and the second alignment path, and adaptive fusion is carried out on the bimodal scale features by combining with modal existence marks, so that fusion features are obtained; Inputting the fusion features into a modal specific texture attention block, extracting high-frequency texture information of each mode through Gaussian filtering difference, generating an attention weight graph according to the obtained high-frequency texture information, and carrying out weighted enhancement on the fusion features according to the attention weight graph to obtain an enhanced feature g