Search

CN-121999226-A - Image segmentation model training method, image segmentation equipment, medium and product

CN121999226ACN 121999226 ACN121999226 ACN 121999226ACN-121999226-A

Abstract

The invention discloses an image segmentation model training method, an image segmentation device, a medium and a product, wherein the image segmentation model training method comprises the steps of inputting a first image sample into a first image segmentation model to obtain a first image segmentation prediction result; the method comprises the steps of converting a first image segmentation prediction result into a prediction result slice sequence to obtain segmentation prompt information, inputting the prediction result slice sequence and the segmentation prompt information into a second image segmentation model to obtain a slice segmentation prediction result based on each prediction result slice, obtaining a second image segmentation prediction result based on the slice segmentation prediction result, using the second image segmentation prediction result as a pseudo tag of a first image sample, determining learning loss of the first image segmentation model based on the pseudo tag, updating the first image segmentation model based on a learning loss reverse gradient, and completing model training. According to the technical scheme provided by the embodiment of the invention, the lightweight model with strong image segmentation capability can be obtained based on a semi-supervised learning mode.

Inventors

  • HOU WENQIANG
  • LONG XIAOJING
  • Niu Donghao

Assignees

  • 中国科学院深圳先进技术研究院

Dates

Publication Date
20260508
Application Date
20260130

Claims (10)

  1. 1. An image segmentation model training method, comprising the steps of: Inputting a first image sample into a first image segmentation model to obtain a first image segmentation prediction result, wherein the first image sample is a three-dimensional medical image, and the first image segmentation model is a student model to be trained; converting the first image segmentation prediction result into a prediction result slice sequence, and obtaining segmentation prompt information corresponding to each prediction result slice in the prediction result slice sequence; Inputting the predicted result slice sequence and the segmentation prompt information into a second image segmentation model to obtain a slice segmentation predicted result based on each predicted result slice, and obtaining a second image segmentation predicted result based on the slice segmentation predicted result, wherein the second image segmentation model is a teacher model with training completed; And taking the second image segmentation prediction result as a pseudo tag of the first image sample, determining learning loss of the first image segmentation model based on the pseudo tag, and updating the first image segmentation model based on the learning loss reverse gradient to complete model training.
  2. 2. The method of claim 1, wherein determining a learning penalty for the first image segmentation model based on the pseudo tag comprises: evaluating the segmentation quality of the second image segmentation prediction result, and obtaining the loss weight of the pseudo tag based on the segmentation quality; And obtaining learning loss of the first image segmentation model based on the second image segmentation result and the loss weight.
  3. 3. The method of claim 2, wherein said evaluating the segmentation quality of the second image segmentation prediction comprises: Obtaining a 3D connectivity assessment result of the second image segmentation prediction result based on a ratio of a maximum connected region to a total segmentation volume in the second image segmentation prediction result, and/or, Obtaining a volume consistency assessment result of the second image segmentation prediction result based on the segmentation volume difference value of the second image segmentation prediction result and the first image segmentation prediction result, and/or, Obtaining a boundary quality assessment result of the second image segmentation prediction result based on smoothness and definition of a segmentation boundary of the second image segmentation prediction result, and/or, And analyzing the uncertainty of the second image segmentation prediction result based on the 3D Monte Carlo Dropout to obtain an uncertainty evaluation result of the second image segmentation prediction result.
  4. 4. The method according to claim 1, wherein obtaining the partition hint information corresponding to each predictor slice in the sequence of predictor slices comprises: Extracting pixels with confidence degree larger than a first confidence degree threshold value in each predicted result slice in the predicted result slice sequence as key points, and taking the key points as corresponding segmentation prompt information, and/or, Based on the segmentation result in each prediction result slice in the prediction result slice sequence, obtaining the minimum circumscribed rectangle of the segmentation result, smoothing the minimum circumscribed rectangle corresponding to the adjacent prediction result slice to obtain a smoothed minimum circumscribed rectangle, taking the smoothed minimum circumscribed rectangle as the corresponding segmentation prompt information, and/or, And taking each prediction result slice in the prediction result slice sequence as a first mask image, performing morphological processing and/or noise filtering on the first mask image to obtain a second mask image, and taking the second mask image as corresponding segmentation prompt information.
  5. 5. The method of claim 1, wherein the deriving a second image segmentation prediction result based on the slice segmentation prediction result comprises: combining each slice segmentation prediction result into a three-dimensional prediction result according to a slice sequence; and 3D space optimization is carried out on the three-dimensional prediction result through at least one processing mode of 3D morphological processing, connected component analysis and volume constraint, so that the second image segmentation prediction result is obtained.
  6. 6. The method according to any one of claims 1-5, wherein the first image segmentation model is trained based on three-dimensional image samples in a first set of image samples, the first image segmentation model being a U-shaped network comprising a 3D convolution layer and an attention mechanism, three of the first set of image samples being associated with a real sample label, and the second image segmentation model being a medical generic segmentation model 2.
  7. 7. An image segmentation method, comprising: Acquiring a first three-dimensional medical image; inputting the first three-dimensional medical image into a first image segmentation model to obtain a first three-dimensional image segmentation result; wherein the first image segmentation model is a model trained based on the image segmentation model training method according to any one of claims 1-6.
  8. 8. An electronic device, the electronic device comprising: One or more processors; Storage means for storing one or more programs, The one or more programs, when executed by the one or more processors, cause the one or more processors to implement the image segmentation model training method or the image segmentation method as set forth in any one of claims 1-7.
  9. 9. A computer readable storage medium having stored thereon a computer program, which when executed by a processor implements an image segmentation model training method or an image segmentation method as claimed in any one of claims 1-7.
  10. 10. A computer program product comprising a computer program, characterized in that the computer program, when being executed by a processor, implements the image segmentation model training method or the image segmentation method as claimed in any one of claims 1-7.

Description

Image segmentation model training method, image segmentation equipment, medium and product Technical Field The present invention relates to the field of medical image processing technologies, and in particular, to an image segmentation model training method, an image segmentation apparatus, an image segmentation device, a medium, and a product. Background In the field of 3D medical image segmentation based on deep learning, 3D data is directly processed through 3D convolution by 3D U-Net, V-Net and the like, so that 3D spatial relations can be effectively captured. However, these methods rely heavily on large amounts of labeling data, and labeling of 3D medical images is extremely costly, limiting its application in practical clinics. However, the quality of the pseudo tag is insufficient in a semi-supervised learning mode, and the image segmentation capability and reliability of the 3D model are also affected. Disclosure of Invention The embodiment of the invention provides an image segmentation model training method, an image segmentation method, image segmentation equipment, a medium and a product, which can transfer medical image segmentation knowledge of a teacher network in a 2D sequence to a 3D network, fully utilize spatial structure information of 3D data and obtain a light model with strong image segmentation capability based on a semi-supervised learning mode. In a first aspect, an embodiment of the present invention provides an image segmentation model training method, including: Inputting a first image sample into a first image segmentation model to obtain a first image segmentation prediction result, wherein the first image sample is a three-dimensional medical image, and the first image segmentation model is a student model to be trained; Converting the first image segmentation prediction result into a prediction result slice sequence, and obtaining segmentation prompt information corresponding to each prediction result slice in the prediction result slice sequence; Inputting the predicted result slice sequence and the segmentation prompt information into a second image segmentation model to obtain a slice segmentation predicted result based on each predicted result slice, and obtaining a second image segmentation predicted result based on the slice segmentation predicted result, wherein the second image segmentation model is a teacher model with training completed; and taking the second image segmentation prediction result as a pseudo tag of the first image sample, determining learning loss of the first image segmentation model based on the pseudo tag, updating the first image segmentation model based on the learning loss reverse gradient, and completing model training. In a second aspect, an embodiment of the present invention further provides an image segmentation method, where the method includes: Acquiring a first three-dimensional medical image; Inputting the first three-dimensional medical image into a first image segmentation model to obtain a first three-dimensional image segmentation result; The first image segmentation model is a model obtained by training based on the image segmentation model training method of any embodiment. In a third aspect, an embodiment of the present invention provides an image segmentation model training apparatus, including: the sample input module is used for inputting a first image sample into a first image segmentation model to obtain a first image segmentation prediction result, wherein the first image sample is a three-dimensional medical image, and the first image segmentation model is a student model to be trained; The cross-dimension prompt information generation module is used for converting the first image segmentation prediction result into a prediction result slice sequence and obtaining segmentation prompt information corresponding to each prediction result slice in the prediction result slice sequence; The pseudo tag generation module is used for inputting the predicted result slice sequence and the segmentation prompt information into a second image segmentation model to obtain a slice segmentation predicted result based on each predicted result slice, and obtaining a second image segmentation predicted result based on the slice segmentation predicted result, wherein the second image segmentation model is a teacher model which has completed training; and the model training updating module is used for taking the second image segmentation prediction result as a pseudo tag of the first image sample, determining the learning loss of the first image segmentation model based on the pseudo tag, updating the first image segmentation model based on the learning loss reverse gradient and finishing model training. In a fourth aspect, an embodiment of the present invention further provides an image segmentation apparatus, including: The image acquisition module is used for acquiring a first three-dimensional medical image; the image segmentation module