Search

CN-121999000-A - Digestive tract polyp semi-supervised segmentation method for improving pseudo-label reliability

CN121999000ACN 121999000 ACN121999000 ACN 121999000ACN-121999000-A

Abstract

The invention belongs to the technical field of image processing, and provides a digestive tract polyp semi-supervised segmentation method for improving the reliability of pseudo labels, the invention is based on a semi-supervised segmentation framework formed by a teacher network and a student network, and mainly comprises a pseudo tag refining module for reliability perception and a consistency learning strategy based on progressive data enhancement. The method is used for improving the reliability of the pseudo tag in the label-free data and enhancing the segmentation performance of the model under the complex endoscope imaging condition by the cooperative work of the modules.

Inventors

  • YUE GUANGHUI
  • MENG XIANGFEI
  • LIN DONGMEI
  • Feng Juerong
  • ZHANG XIANGYU

Assignees

  • 深圳大学

Dates

Publication Date
20260508
Application Date
20260202

Claims (7)

  1. 1. The utility model provides a digestive tract polyp semi-supervised segmentation method for improving the reliability of pseudo labels, which is characterized by comprising the following steps: Step 1, selecting an endoscope image data set of polyps of the digestive tract, and dividing the image data set into labeled data And non-labeling data ; Step 2, constructing a student network and a teacher network based on a mean-teacher architecture; Step 3, for the marked data Weak enhancement treatment is carried out to the mixture to obtain Then input into student network for supervision training, and for non-marked data Weak enhancement treatment is carried out to obtain And input it into the teacher's network to generate initial predictions ; Step 4, predicting result of teacher network by utilizing reliability sensing pseudo tag refining module Processing to obtain refined pseudo tag ; Step 5, adopting a progressive enhancement consistency learning strategy to apply enhancement processing with different intensities to the same non-labeled data respectively to generate corresponding enhanced non-labeled data , , ; Step 6, respectively inputting the unmarked data into a student network to obtain corresponding segmentation prediction results , , And pseudo tag after refining Consistency constraint learning is performed under supervision.
  2. 2. The method for semi-supervised segmentation of polyps of the digestive tract to improve pseudo label reliability of claim 1, wherein step 4 comprises: For initial prediction result Performing maximum value selection operation along the channel direction to obtain initial pseudo tag ; Based on the initial prediction result Applying softmax operation to calculate the prediction probability of each pixel belonging to polyp category to obtain prediction confidence ; According to the initial pseudo tag From the initial prediction result Is characterized by (a) Extracting all pixel characteristics judged to be polyp category, averaging the extracted pixel characteristics, and constructing prototype characteristics of the polyp category ; Then, for each pixel determined to be a polyp category Calculating the characteristics through cosine similarity function And polyp prototype features Similarity between the two to obtain a similarity value ; Then, based on the similarity value Confidence with prediction The difference between them, a similarity-confidence alignment score is obtained The calculation method is as follows: (3) constructing unified reliability scores The calculation mode is as follows: (4) Wherein, the Is a weight parameter; screening the pseudo tags by adopting a dynamic threshold strategy, wherein the dynamic threshold is used for Calculated as follows: (5) Wherein, the And Representing the initial threshold value and the final threshold value respectively, Representing the current training round of the present training, Representing a maximum training round; In the training process, the unified reliability scoring corresponding to the pixels is only performed Greater than the current threshold If the pseudo tag corresponding to the pixel is reserved for training, otherwise, the pixel is set as a background category, and the screening rule is as follows: (6) performing connected region analysis on the screened pseudo tags, and calculating the area of each independent region when the area of each region is smaller than a preset minimum threshold value When the area is re-determined as the background type, the refined pseudo tag is finally obtained 。
  3. 3. The method for semi-supervised segmentation of polyps of the digestive tract to improve pseudo-label reliability of claim 2, wherein the initial pseudo-label is The calculation mode of (2) is as follows: (1)。
  4. 4. The method for semi-supervised segmentation of polyps of the digestive tract to improve pseudo label reliability of claim 2, wherein the confidence of the prediction is The calculation mode of (a) is as follows: (2) Wherein, the Representative of Is a prediction of polyp channels.
  5. 5. The method for semi-supervised segmentation of polyps of the digestive tract for improving the reliability of pseudo labels according to claim 2, wherein in step 5, medium enhancement and strong enhancement are respectively applied to the same non-labeled data after weak enhancement treatment, so as to generate corresponding medium enhancement non-labeled data and strong enhancement non-labeled data.
  6. 6. The method for semi-supervised segmentation of polyps of the digestive tract with improved pseudo-label reliability as recited in claim 5, wherein refined pseudo-labels are utilized for each enhanced unlabeled data Combining supervision loss functions And unsupervised loss function And (5) optimizing.
  7. 7. The method for semi-supervised segmentation of polyps of the digestive tract to improve pseudo label reliability as recited in claim 6, wherein the unsupervised loss function comprises a cross entropy loss function And the Dice loss function The method is used for constraining the consistency of the prediction results of the same unlabeled data under different enhancement conditions in category judgment, and the formula is as follows: (7)。

Description

Digestive tract polyp semi-supervised segmentation method for improving pseudo-label reliability Technical Field The invention belongs to the technical field of image processing, and particularly relates to a digestive tract polyp semi-supervised segmentation method for improving the reliability of pseudo labels. Background In clinical practice, endoscopy has become a major technical tool for digestive tract polyp detection and excision. However, due to physician experience differences, imaging condition variations, and polyps' complexity in morphology and spatial location, missed or erroneous determinations may still occur during existing endoscopy procedures. Therefore, developing an automated and high-precision segmentation method for polyps of the digestive tract has important significance for improving the diagnosis reliability. In the prior art, an image segmentation method based on depth learning has been introduced into the field of lesion segmentation and applied to polyp segmentation tasks. However, such methods typically rely on large-scale, pixel-level precisely labeled datasets, while the acquisition costs of the relevant data are high, and highly dependent on manual labeling by medical professionals, limiting their further application in practical scenarios. In order to reduce the labeling cost, the semi-supervised segmentation method is an alternative scheme in current research and application by performing model training by using a small amount of labeling data and a large amount of unlabeled data. However, existing semi-supervised segmentation techniques still suffer from the disadvantage that, on the one hand, part of the method evaluates the reliability of pseudo tags mainly based on semantic prediction probabilities or on entropy-based uncertainty estimates. Because the evaluation mode mainly depends on model output distribution, the problem of excessive confidence is easy to occur under the complex endoscope imaging condition, thereby introducing a noise pseudo tag and affecting the model training effect. On the other hand, the existing consistency regularization method mostly adopts a data enhancement strategy of copying, splicing or mixing. In an endoscope image, since polyps often have the characteristics of low contrast, complex textures, blurred boundaries and the like, the enhancement mode is easy to cause inconsistent semantics or local structural distortion, and particularly the enhancement mode is more obvious in a target boundary area, so that the learning ability of a model on a real polyp structure is weakened. Therefore, it is still necessary to propose a polyp semi-supervised segmentation method capable of improving the reliability of pseudo labels and enhancing the robustness of a model under complex endoscopic imaging conditions while reducing the labeling cost. Disclosure of Invention In order to solve the technical problems, the invention provides a digestive tract polyp semi-supervised segmentation method for improving the reliability of pseudo labels, which solves the problems in the prior art, and adopts the following technical scheme: a digestive tract polyp semi-supervised segmentation method for improving pseudo-label reliability, comprising: Step 1, selecting an endoscope image data set of polyps of the digestive tract, and dividing the image data set into labeled data And non-labeling data; Step 2, constructing a student network and a teacher network based on a mean-teacher architecture; Step 3, for the marked data Performing weak enhancement treatment on the data, inputting the data into a student network for supervision training, and performing non-labeling dataWeak enhancement processing is carried out and input into a teacher network to generate initial prediction results; Step 4, processing the prediction result of the teacher network by using a reliability sensing pseudo tag refining module to obtain a refined pseudo tag; Step 5, adopting a progressive enhancement consistency learning strategy to respectively apply enhancement treatments with different intensities to the same non-labeled data after the weak enhancement treatment to generate corresponding enhanced non-labeled data; And 6, respectively inputting the enhanced non-labeling data into a student network to obtain a corresponding segmentation prediction result, and performing consistency constraint learning under the supervision of the refined pseudo tag. Further, step 4 includes: For initial prediction result Performing maximum value selection operation along the channel direction to obtain initial pseudo tag; Based on the initial prediction resultApplying softmax operation to calculate the prediction probability of each pixel belonging to polyp category to obtain prediction confidence; According to the initial pseudo tagFrom the initial prediction resultIs characterized by (a)Extracting all pixel characteristics judged to be polyp category, averaging the extracted pixel characteristics, and con