CN-121982307-A - Uncertainty guidance-based three-dimensional tumor image semi-supervised segmentation method

CN121982307ACN 121982307 ACN121982307 ACN 121982307ACN-121982307-A

Abstract

The invention provides a three-dimensional tumor image semi-supervised segmentation method based on uncertainty guidance. The method comprises the steps of preprocessing unlabeled three-dimensional tumor images, inputting the unlabeled three-dimensional tumor images into a three-dimensional medical image segmentation basic model by combining prompt information, utilizing enhancement to obtain a plurality of tumor segmentation masks during testing, calculating case-level uncertainty scores and generating voxel-level uncertainty thermodynamic diagrams according to the obtained result, dividing samples into high-uncertainty samples and low-uncertainty samples according to the case-level uncertainty scores, manually labeling the high-uncertainty samples, adopting a prediction result of the basic model as a pseudo tag for the low-uncertainty samples, combining the two types of samples to construct a mixed training set, and adopting a weighted sampling strategy to train a target model. The invention can effectively screen samples based on uncertainty guidance, and can fully utilize the general segmentation capability of the basic model while obviously reducing the labeling cost, thereby improving the precision of the model on specific tumor segmentation tasks.

Inventors

SHI ZHENYANG
CAI JINYAN
YIN SHI

Assignees

南京工业大学

Dates

Publication Date: 20260505
Application Date: 20260119

Claims (10)

1. The uncertainty guidance-based three-dimensional tumor image semi-supervised segmentation method is characterized by comprising the following steps of: Step S1, preprocessing unlabeled three-dimensional tumor images, and inputting the unlabeled three-dimensional tumor images into a three-dimensional medical image segmentation basic model by combining prompt information; S2, enhancing and acquiring a plurality of tumor segmentation masks by utilizing test time, calculating case-level uncertainty scores according to the tumor segmentation masks and generating a voxel-level uncertainty thermodynamic diagram; s3, dividing the sample into a high uncertainty sample set and a low uncertainty sample set according to the case-level uncertainty score, manually marking the high uncertainty sample, and adopting a prediction result of the basic model as a pseudo tag for the low uncertainty sample; and S4, combining the two types of samples to construct a mixed training set, and training a target model by adopting a weighted sampling strategy.
2. The uncertainty guidance-based three-dimensional tumor image semi-supervised segmentation method as set forth in claim 1, wherein the preprocessing in the step S1 comprises resampling, normalizing and normalizing the tumor image, the prompt information comprises a text prompt and a spatial prompt, the text prompt is a text description of a tumor category, and the spatial prompt can be one of a point and a frame.
3. The method of semi-supervised segmentation of three dimensional tumor images based on uncertainty guidance of claim 1, wherein in step S1, the three dimensional medical image segmentation base model adopts an encoder-decoder architecture, and comprises an image encoder, a space encoder, a text encoder and a mask decoder, wherein the image encoder is used for extracting global and local features of the three dimensional medical image, the space encoder is used for receiving and encoding space prompt information, the space prompt information comprises points and boxes, the text encoder is used for performing text encoding on text descriptions or feature vectors of tumor categories, and the mask decoder is used for fusing the image features, the space encoding features and the text encoding features and outputting a final segmentation mask.
4. The method of uncertainty-guided three-dimensional tumor image semi-supervised segmentation as set forth in claim 1, wherein the test-time enhancement comprises one or more of geometric transformation, intensity transformation, noise injection, and blurring in step S2.
5. The uncertainty guidance-based three-dimensional tumor image semi-supervised segmentation method as set forth in claim 1, wherein the case-level uncertainty score calculation process in step S2 specifically includes calculating an average pair-wise Dice coefficient for all possible combinations (P i ,P j ) for N tumor segmentation masks { P 1 ,P 2 ,...,P N } based on enhancement-while-test for unlabeled tumor image samples X Then the case-level uncertainty score U case (X) =1-MPD (X) for that sample.
6. The uncertainty guidance-based three-dimensional tumor image semi-supervised segmentation method as set forth in claim 1, wherein the voxel level uncertainty thermodynamic diagram visualizes the uncertainty score of each voxel of the three-dimensional tumor image by Jet color mapping in step S2, calculates the average prediction probability of enhancement at N tests for voxels at spatial location v Wherein P n (v) represents the predicted probability value of voxel v under the enhanced policy at the nth test, then the voxel-level uncertainty score
7. The three-dimensional tumor image semi-supervised segmentation method based on uncertainty guidance of claim 1, wherein in the step S3, the high uncertainty samples and the low uncertainty samples are classified according to a case-level uncertainty score from high to low, all samples are classified into the high uncertainty samples and the low uncertainty samples according to a preset fixed threshold, the fixed threshold can be set to any one of 5%, 10%, 15% or 20%, the high uncertainty samples are manually marked to obtain manual marked samples, and the low uncertainty samples are pseudo-labeled by using a prediction result of a basic model as a pseudo-label.
8. The uncertainty guidance-based three-dimensional tumor image semi-supervised segmentation method as set forth in claims 1 and 7, wherein the weighted sampling strategy assigns sampling weights according to the case-level uncertainty fraction of the samples in step S4, and the method specifically comprises the steps of assigning a fixed weight α (α≥1) to the artificially labeled samples, calculating weights for pseudo-labeled samples according to an exponential decay function w=exp (-u/τ) based on the case-level uncertainty fraction u thereof, wherein τ is a temperature coefficient for adjusting the decay rate, and calculating the sum W sum = Σwof the sampling weights of all samples to obtain a normalized sampling probability of each sample And randomly extracting samples in each iteration step of model training according to the sampling probability to form a training batch.
9. The uncertainty-guided three-dimensional tumor image semi-supervised segmentation method according to claim 1, wherein in step S4, the target model is a deep neural network based on a three-dimensional U-Net structure.
10. A computer-readable storage medium, on which a computer program is stored, which, when being executed by a processor, is capable of carrying out the method steps of any one of claims 1 to 9.

Description

Uncertainty guidance-based three-dimensional tumor image semi-supervised segmentation method Technical Field The invention belongs to the technical field of medical image segmentation, and particularly relates to a three-dimensional tumor image semi-supervised segmentation method based on uncertainty guidance. Background Medical imaging plays a critical role in cancer diagnosis, treatment planning and prognosis evaluation. Accurate and reliable tumor segmentation not only can help clinicians to accurately identify tumor boundaries and evaluate tumor volumes, but also can remarkably improve prognosis of patients and promote development of accurate medical treatment. Although imaging techniques such as Computed Tomography (CT) and Magnetic Resonance Imaging (MRI) continue to advance, tumors exhibit a high degree of heterogeneity in shape, size, and location, and tend to be similar in density to surrounding soft tissues (e.g., blood vessels, muscles, or adjacent organs), with blurred boundaries, resulting in accurate tumor segmentation facing serious challenges. In recent years, deep learning technology has made remarkable progress in the field of medical image segmentation, wherein the introduction of uncertainty quantization provides a new idea for solving the above-mentioned problems. The uncertainty reflects the confidence level of the model for the self-predicted result and characterizes the ambiguity of the model in outputting correct segmentation under a given input. By uncertainty evaluation of the segmentation result, the generated uncertainty thermodynamic diagram can show areas with large segmentation ambiguity, so that more reliable references are provided for clinical diagnosis and treatment planning, and the scientificity and safety of decision making are enhanced. Currently, most advanced tumor automatic segmentation models (such as nnU-Net) have reached higher segmentation accuracy on the public dataset, but still have the limitation that firstly, these models usually only output deterministic segmentation results, lack quantification of prediction reliability, and are difficult to identify and prompt areas where erroneous segmentation is likely. Secondly, the performance of the model is highly dependent on a large amount of accurate pixel-level labeling data, and the acquisition of the data requires a great deal of time for a doctor of the deep imaging department to manually sketch, so that the cost is high and the efficiency is low. In addition, the existing model lacks an efficient interaction and correction mechanism, and doctors either fully accept the result with possible errors or have to make full manual modification, so that the popularization of the model in clinical work is limited. To break through data bottlenecks and interaction dilemma, a medical image segmentation base model (e.g., segVol) has been developed. The basic models are pre-trained on massive medical data, learn general anatomic features, and have strong feature extraction and generalization capabilities. More importantly, the method supports multi-mode interaction modes such as text prompt and space prompt, so that semantic ambiguity of liver and liver tumor in a segmentation task can be effectively reduced, and the method is more suitable for clinical actual demands. While the base model exhibits excellent general performance, when dealing with certain specific types of tumors (e.g., soft tissue tumors with blurred boundaries, low contrast), the accuracy of its direct reasoning may be insufficient to meet clinical requirements, often requiring further migration training for the target lesion. Therefore, under the condition of lacking large-scale labeling data, how to efficiently utilize the general capability of a medical image segmentation basic model, realize high-precision tumor segmentation through a low-cost labeling strategy, and simultaneously provide reliable uncertainty measurement to assist clinical decision is a technical problem to be solved in the current medical image segmentation field. Disclosure of Invention In view of the above, the invention provides a three-dimensional tumor image semi-supervised segmentation method based on uncertainty guidance, which can effectively screen samples based on uncertainty guidance, and can fully utilize the general segmentation capability of a basic model while remarkably reducing the labeling cost and improve the precision of the model on a specific tumor segmentation task. In order to achieve the above purpose, the present invention adopts the following technical scheme: A three-dimensional tumor image semi-supervised segmentation method based on uncertainty guidance comprises the following steps: Step S1, preprocessing unlabeled three-dimensional tumor images, and inputting the unlabeled three-dimensional tumor images into a three-dimensional medical image segmentation basic model by combining prompt information; S2, enhancing and acquiring a plurality of tumor s