Method for segmenting liver tumor on CT (Computed Tomography) slice image by applying deep learning

A deep learning framework combining a noise-robust adapter and semi-supervised learning addresses noisy label issues in liver tumor segmentation, enhancing model robustness and accuracy for improved clinical applications.

CN119067992APending Publication Date: 2024-12-03YANGTZE DELTA REGION INST (QUZHOU) UNIV OF ELECTRONIC SCI & TECH OF CHINA +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411174674.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-26
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

Existing deep learning models for liver tumor segmentation in CT images face challenges due to noisy labels, which affect performance, generalization, and clinical reliability, necessitating improved strategies to handle label noise and enhance model robustness.

Method used

A deep learning framework combining a noise-robust adapter model with a semi-supervised learning method, utilizing a noise adapter and a multi-stage perturbation and variable teacher model to mitigate the impact of noisy labels, enhancing model robustness and generalization.

Benefits of technology

The proposed framework effectively reduces the adverse effects of noisy labels, improving the accuracy and reliability of liver tumor segmentation, enabling better clinical decision-making through non-invasive CT image analysis.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119067992A_ABST
    Figure CN119067992A_ABST
Patent Text Reader

Abstract

The invention discloses a method for segmenting a liver tumor on a CT (Computed Tomography) slice image by applying deep learning. According to the method, a neural network framework named MPVT + is introduced and is used for training a liver tumor segmentation model on a data set with a noise tag. By using an adapter model method having robustness to noise, the MPVT + model can effectively adapt and screen noise tags in a training data set, thereby reducing interference of noise to a neural network. In addition, the generalization ability of the MPVT + model is enhanced through a semi-supervised learning method, so that the MPVT + model can efficiently learn complex CT slice image features. On a test data set, the Sorenson-Dess coefficient of the model is 80.29%, the Jacokan index of the model is 68.68%, the symmetric volume difference of the model is 19.71%, and the volume overlapping error of the model is 31.32%. The model can noninvasively use the CT image to segment and locate the tumor in the liver, so that the model becomes a potential tool for helping doctors to formulate various liver tumor treatment strategies.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the fields of deep learning and electronic computer tomography image diagnosis, and in particular to a method for segmenting liver tumors on CT slice images using deep learning. Background Art

[0002] Liver cancer is a common malignancy and ranked third among the causes of cancer-related deaths worldwide in 2020 (Sung, Ferlay, Siegel, Laversanne, Soerjomataram, Jemal and Bray, 2021). The number of liver cancer cases has increased year by year, making it the only cancer among the top five deadly cancers with an increasing number of cases (Siegel, Miller, Wagle and Jemal, 2023). Diagnosis of liver cancer usually involves various medical imaging techniques such as ultrasound, magnetic resonance imaging (MRI) and computed tomography (CT). Among these methods, CT is the most commonly used method because its multi-phase contrast enhancement capability provides a more advanced imaging mode (Bialecki and Di Bisceglie, 2005). However, accurate diagnosis and subsequent treatment planning rely heavily on the precise delineation of tumor boundaries within the liver. Physical characteristics of tumors, such as size, shape, and location, are important biomarkers for accurate diagnosis and treatment (Christ, Elshaer, Ettlinger, Tatavarty, Bickel, Bilic, Rempfler, Armbruster, Hofmann, D'Anastasi, Sommer, Ahmadi, and Menze, 2016).

[0003] For tumor segmentation tasks, the quality of the labels used for training is particularly important because it directly affects the performance of the model in correctly identifying and segmenting tumors, thereby affecting clinicians' diagnosis and treatment decisions for tumors. Accurate, consistent, and clear labels not only improve the performance and generalization ability of the model, but also ensure that the segmentation results are clinically interpretable and reliable in application. First, incorrectly labeled data may affect the performance of deep learning models, resulting in learning incorrect features or segmentation boundaries. Second, noisy labels can affect the generalization ability of the model, leading to poor performance on unseen data. In addition, overfitting to noisy labels may cause the model to perform poorly on test data because it fails to accurately capture the true distribution of the data. Finally, the presence of low-quality labels increases the difficulty of training, and corresponding strategies are needed to mitigate their impact in order to improve the performance and generalization ability of the model. Therefore, ensuring high-quality labeled data is crucial to the success of tumor segmentation tasks.

[0004] To address the challenges posed by noisy labels, the present invention explores a variety of strategies. Some studies use consistent labels from many domain experts to improve data quality (Gulshan, Peng, Coram, Stumpe, Wu, Narayanaswamy, Venugopalan, Widner, Madams, Cuadros, Kim, Raman, Nelson, Mega and Webster, 2016; Esteva, Kuprel, Novoa, Ko, Swetter, Blau and Thrun, 2017). However, this approach consumes a lot of financial and logistical resources and is difficult to popularize. In addition, inter-observer variability further exacerbates this problem, as different radiologists may interpret ambiguous features differently based on their experience and training (Warfield, Zou and Wells, 2004; Rohlfing, Russakoff and Maurer, 2004). Other studies aim to reduce the workload of experts in medical image annotation through computer assistance (Deng, Sun, Zhu, Xu, Yang, Zhang, Wang, Sun, Zhao, Zhou and Yuan, 2019). However, when using this approach, differences between different observers and within the same observer cannot be ignored, because a single radiologist may provide different annotations at different times, affected by factors such as fatigue or changes in perceptual standards (Joskowicz, Cohen, Caplan and Sosna, 2019). Some studies use data automatically extracted from medical image databases such as hospital picture archiving and communication systems to create training datasets (Yan, Wang, Lu and Summers, 2018; Irvin, Rajpurkar, Ko, Yu, Ciurea-Ilcus, Chute, Marklund, Haghgoo, Ball, Shpanskaya et al, 2019). Historical data may not be available for every study and are often noisy. Some studies use crowdsourcing to collect labeled data from non-experts (Kaur and Kaur, 2014; Gurari, Theriault, Sameki, Isenberg, Pham, Purwada, Solski, Walker, Zhang, Wong, and Betke, 2015). However, they usually fail to produce labels of sufficient quality for general research and have high noise rates (Gurarie et al, 2015). Some researchers have proposed using noisy labels to improve model performance.These methods focus on creating a robust architecture to model latent parameters, enhancing generalization, tuning loss functions, and selecting appropriate samples to reduce the impact of noisy labels on the model (Song, Kim, Park, Shin, and Lee, 2023). For example, some methods utilize robust optimization techniques that specifically account for label noise (Frénay and Verleysen, 2013), while others propose noise-resistant loss functions that aim to reduce the impact of mislabeled data during training (Zhang and Sabuncu, 2018). In addition, there is growing interest in the implementation of transfer learning, where models pre-trained on large, clean datasets are fine-tuned on smaller, noisier datasets to leverage learned representations that are less susceptible to label noise (Shin, Roth, Gao, Lu, Xu, Nogues, Yao, Mollura, and Summers, 2016). Techniques such as label smoothing and uncertainty modeling can also help models not overfit to noisy labels, but instead focus on underlying patterns in the data (Müller, Kornblith, and Hinton, 2019). Some methods ignore noisy labels entirely and use semi-supervised methods to train the network. These evolving methods emphasize the dynamic nature of deep learning in medical images, and ongoing research continues to refine the balance between data quality and model reliability.

[0005] Probabilistic noise models are the basic concept of robust architectures. These methods are based on the concept that noise can be represented by latent variables of neural networks (Tong Xiao, Tian Xia, Yi Yang, Chang Huang and Xiaogang Wang, 2015; Song et al, 2023). The method of the present invention is inspired by robust architectures and semi-supervised learning. In order to train deep neural networks with noisy labels, it is key to identify noisy labels and eliminate their negative effects as much as possible. Based on the concept of noise adaptation layer (Goldberger and BenReuven, 2016), a noise adapter is introduced to change the convolutional neural network by adding fully connected layers and softmax layers to predict hidden variables and simulate the generation of label noise. This module is dynamically adjusted during training and repairs unreliable pixels in the label to produce an unbiased supervisory signal. In addition, in order to prevent the model from overfitting due to potential changes in data distribution, a semi-supervised framework called MPVT is implemented. The framework includes a parameter-variable mean teacher model whose update is inspired by dropout (Srivastava, Hinton, Krizhevsky, Sutskever and Salakhutdinov, 2014) to generate more robust pseudo-label guidance, thereby allowing more perturbations to be introduced during training, enhancing the robustness and stability of the student model. This process effectively improves the model's resistance to overfitting by providing a variety of training scenarios, ensuring that it generalizes from the core features of the data rather than its noisy aspects. The framework aims to help the model learn from reliable data and extract valuable information from noisy labels without being affected by noise. Compared with previous methods, the method of the present invention not only identifies and reduces the impact of noisy labels, but also strengthens the robustness of the network by exploiting the inherent uncertainty in the data. This dual strategy effectively creates a noise-resistant learning process, thereby improving the quality of model predictions and ensuring higher reliability in clinical applications.

[0006] In this invention, a method for segmenting liver tumors on computed tomography images is introduced by introducing a deep learning framework based on CT images. Compared with the traditional model, the results of the framework proposed in this invention show excellent performance. Summary of the invention

[0007] The purpose of the present invention is to address the deficiencies of the prior art and provide a method for segmenting liver tumors on CT slice images using deep learning, introducing a framework for training a neural network for segmenting liver tumors on a data set with noisy labels by combining a noise-robust adapter model method with a semi-supervised learning method.

[0008] To achieve the above object, the present invention provides a method for segmenting liver tumors on CT slice images using deep learning, comprising the following steps:

[0009] (1) Divide the original data set into data with noise labels and data without noise labels;

[0010] (2) After weak enhancement processing, the two types of training data obtained in step (1) are input into an encoder of a U-Net-based image segmentation network for feature extraction to obtain a feature vector;

[0011] (3) sending the feature vector of the training data without noise labels obtained in step (2) into the decoder of the U-Net-based image segmentation network to obtain the segmentation result of the network;

[0012] (4) performing loss calculation on the segmentation result obtained in step (3) and the pre-given label, wherein the loss function is the sum of the cross-entropy loss and the Dice loss;

[0013] (5) sending the feature vector of the training data with noise labels obtained in step (2) into the decoder and noise adapter of the U-Net-based image segmentation network to obtain the network segmentation result and the noise adaptation mask respectively;

[0014] (6) The network segmentation result and the noise adaptation mask obtained in step (5) are used to obtain a predicted segmentation result with noise according to the full probability formula;

[0015] (7) performing loss calculation on the predicted noisy segmentation result obtained in step (6) and the noisy label, wherein the loss function is the sum of the cross-entropy loss and the Dice loss;

[0016] (8) performing weak enhancement processing and strong enhancement processing on the training data obtained in step (1), and then performing processing according to step (2) to obtain respective feature vectors;

[0017] (9) inputting the two types of training data after weak enhancement processing obtained in step (8) into the teacher model to obtain their respective feature vector noise and pseudo labels;

[0018] (10) adding the feature vector noise obtained in step (9) to the respective feature vectors obtained in step (8);

[0019] (11) Processing the feature vector of the data without noise labels obtained in step (10) after adding noise according to steps (5) to (7) to obtain the corresponding loss;

[0020] (12) Processing the feature vector after adding noise to the data with noise labels obtained in step (10) according to steps (3) to (4) to obtain corresponding losses;

[0021] (13) The loss results obtained in step (4), step (7), step (11), and step (12) are weighted and then back-propagated to optimize network parameters and improve the accuracy of model prediction results.

[0022] Furthermore, the step (1) includes the following sub-steps:

[0023] (1.1) Screening cases that are obviously misaligned with the anatomical structures they should represent as data with noisy labels;

[0024] (1.2) Screening cases where labels infiltrate adjacent structures due to segmentation errors as data with noisy labels;

[0025] (1.3) Screen cases with inconsistent labels or boundary conflicts between slices as data with noisy labels;

[0026] (1.4) Screen cases where labels contain random noise or do not represent the true anatomical structure as data with noisy labels;

[0027] (1.5) Screen cases with incomplete or cut-off labeled regions as data with noisy labels.

[0028] Furthermore, the weak enhancement processing is specifically: randomly flipping the image upside down with a probability of 50%, randomly flipping the image left to right with a probability of 50%, and randomly performing an affine transformation on the image.

[0029] Furthermore, the strong enhancement processing is specifically: randomly performing elastic transformation on the image, randomly performing CutMix processing on the image, randomly performing MaskMix processing on the image, randomly erasing part of the image content, randomly generating black squares in the image, randomly performing Gaussian blur processing on the image, randomly adding Rice noise to the image, randomly adding Gaussian noise to the image, randomly sharpening the image and randomly performing color transformation on the image.

[0030] Furthermore, in step (6), the total probability formula is expressed as:

[0031]

[0032] in represents the prediction after adaptation, is the prediction of U-Net, p ada (j=1|i=0) is the probability of mislabeling the background as foreground, p ada(j=1|i=1) is the probability of correctly labeling the foreground.

[0033] Furthermore, the encoder is used to extract features of the image and gradually reduce the spatial resolution of the image to capture high-level semantic information.

[0034] Furthermore, the decoder is used to restore the feature map compressed by the encoder to the resolution of the original image and generate a segmentation result of the same size as the input image.

[0035] The beneficial effects of the present invention are: the present invention proposes a novel deep learning framework that combines a robust noise adapter and consistency regularization technique to mitigate the adverse effects of noisy labels on deep neural network training. By combining clean and noisy labeled data, the method of the present invention enhances the robustness and generalization ability of the model to noise, outperforming traditional methods and achieving better results. The method of the present invention shows good prospects in training segmentation convolutional neural networks suitable for detecting lesions in data sets affected by noisy labels. It effectively reduces the need to construct perfectly labeled data sets for deep neural network training and becomes a powerful tool to assist clinical decision-making. The model can non-invasively use computed tomography image segmentation to locate tumors in the liver, thus becoming a potential tool to help doctors formulate various liver tumor treatment strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 It is a supervised training framework diagram of the MPVT+ model provided by the present invention;

[0037] Figure 2 It is a semi-supervised training framework diagram of the MPVT+ model provided by the present invention;

[0038] Figure 3 Comparison results between the MPVT+ provided by the present invention and the traditional model;

[0039] Figure 4 It is the visualization of the segmentation results of the MPVT+ model provided by the present invention and other methods. DETAILED DESCRIPTION

[0040] The core technology of the present invention is to combine the noise adapter module with the consistency regularization framework of multi-stage perturbation and variable teachers to build a deep learning training framework to solve the harmful effects of noise labels on neural network training. The structural framework of the MPVT+ model of the present invention is as follows Figure 1 and Figure 2 It contains two modules, namely the noise adapter module and the multi-stage perturbation and variable teacher module.

[0041] The present invention provides a method for segmenting liver tumors on CT slice images using deep learning, comprising the following steps:

[0042] (1) The original dataset is divided into data with noise labels and data without noise labels.

[0043] (2) The two types of training data obtained in step (1) are weakly enhanced and then input into an encoder of an image segmentation network based on U-Net for feature extraction to obtain a feature vector; wherein the basic architecture of U-Net and the functions of the encoder and decoder are common knowledge in the field of deep learning and computer vision. The encoder consists of a series of convolutional layers (usually with ReLU activation function) and pooling layers; it mainly functions to extract image features and gradually reduce the spatial resolution of the image to capture high-level semantic information.

[0044] (3) The feature vector of the training data without noise labels obtained in step (2) is sent to the decoder of the U-Net-based image segmentation network to obtain the segmentation result of the network; wherein the decoder is usually composed of an upsampling layer (such as transposed convolution or bilinear interpolation) and a convolution layer; it mainly functions to restore the feature map compressed by the encoder to the resolution of the original image and generate a segmentation result of the same size as the input image.

[0045] (4) The segmentation result obtained in step (3) and the pre-given label are used to calculate the loss, where the loss function is the sum of the cross-entropy loss and the Dice loss.

[0046] (5) The feature vector of the training data with noise labels obtained in the step (2) is sent to the decoder of the U-Net-based image segmentation network and the noise adapter modified based on the U-Net to obtain the network segmentation result and the noise adaptation mask respectively; the segmentation model is more similar to the pixel classification task of the image. The present invention believes that in the segmentation model, different CT slice image pixels should have different classification probabilities. In view of the above assumptions, the present invention proposes a noise adapter, whose architecture is exactly the same as the decoder of the unet segmentation network, the difference is that the noise adapter has a different interpretation and loss function. Generally, the edge signal intensity of CT slice images is low, while the middle signal is relatively strong. The present invention believes that the area with opposite signal intensity from the surrounding pixels is more biased towards the lesion area. This concept can be called position label noise, which can be expressed as

[0047]

[0048] Wherein, w represents a specific pixel point in the CT slice image.

[0049] like Figure 1As shown in Figure 2, after the noise adapter, the noise adaptation mask of the CT slice image is obtained. The noise adaptation mask is divided into two parts: the probability that the background is incorrectly marked as the foreground and the probability that the foreground is correctly marked as the foreground, which can be expressed as

[0050]

[0051] in, represents the predicted probability of the noise adapter, represents the prediction probability of the unet decoder. ada (j=1|i=0) represents the probability that the background is mistakenly marked as the foreground, p ada (j=1|i=1) represents the probability that the foreground is correctly labeled as the foreground. In this way, the present invention can train the segmentation model with data containing noise labels in a supervised learning manner.

[0052] Then, the training loss function of the unet segmentation network with the noise adapter is defined as:

[0053] L=L sup +β ada (L ada +β reg σ)

[0054]

[0055] Among them, β ada and β reg is a hyperparameter. n represents a dataset with label noise, l is the loss function of supervised learning (the cross-entropy loss function and Dice loss function are used in this invention), τ weak Indicates weak data enhancement. θ Indicates a unet segmented network. For noise adapter.

[0056] (6) The network segmentation result and the noise adaptation mask obtained in step (5) are used according to the full probability formula to obtain a predicted segmentation result with noise.

[0057] (7) Performing loss calculation on the predicted noisy segmentation result obtained in step (6) and the noisy label, wherein the loss function is the sum of the cross-entropy loss and the Dice loss.

[0058] (8) The training data obtained in step (1) is subjected to weak enhancement processing and strong enhancement processing, and then processed according to step (2) to obtain respective feature vectors.

[0059] (9) The two types of training data after weak enhancement processing obtained in step (8) are input into the teacher model to obtain their respective feature vector noise and pseudo labels.

[0060] The teacher model is usually the same neural network as the student model, and its parameters are updated by the exponential moving average (EMA) of the student model parameters; the purpose is to make the teacher model smoother and more stable and reduce the impact of noise. One of the main functions of the teacher model is to generate pseudo labels; since the parameters of the teacher model are updated by the exponential moving average (EMA), it is more stable than the student model and can provide more accurate predictions; therefore, when the weakly enhanced training data is input to the teacher model, the teacher model will generate more reliable pseudo labels, which are used to guide the training of the student model. The teacher model can also improve the robustness of the student model through consistency regularization; specifically, the weakly enhanced data is passed through the teacher model to obtain feature vectors and pseudo labels; then, when the student model faces different perturbations of the same input data (different enhancements), its output should be consistent with the output of the teacher model; this consistency regularization can prevent the student model from overfitting the noisy labels, thereby improving its generalization ability.

[0061] Semi-supervised learning techniques are valuable in extracting features from unlabeled data, which aids in the learning of neural networks and thus enhances their overall generalization. However, when training a model with noisy labels, the model may become biased and suffer from poor generalization. By discarding the noisy labels, the training process can be converted to semi-supervised learning, making full use of data with corrupted labels. Common techniques include pseudo-labeling and perturbed consistency regularization.

[0062] To realize this concept, this paper proposes the MPVT framework, a self-training and consistency regularization framework with multi-stage perturbation and variable teachers. By introducing a multi-stage perturbation strategy and a variable teacher model, the framework can better handle the noisy label problem and improve the robustness and generalization ability of the model.

[0063] During training, the loss function is defined as:

[0064] L=L sup +β semi L semi

[0065]

[0066] Among them, τ strong It indicates strong enhancement processing. Denoted as the ensemble of teacher networks, ρ τIt is a function that estimates pseudo labels from predicted probabilities, and the thresholds are set to ε1,ε2. In order to solve the possible sample selection bias that affects the training process, this patent proposes a customized pseudo label generation method, which is expressed as:

[0067]

[0068] (10) Add the feature vector noise obtained in step (9) to the respective feature vectors obtained in step (8).

[0069] (11) The feature vector after adding noise to the data without noise labels obtained in step (10) is processed according to steps (5) to (7) to obtain the corresponding loss.

[0070] (12) The feature vector after adding noise to the data with noise labels obtained in step (10) is processed according to steps (3) to (4) to obtain the corresponding loss.

[0071] (13) The loss results obtained in step (4), step (7), step (11), and step (12) are weighted and then back-propagated to optimize network parameters and improve the accuracy of model prediction results.

[0072] Furthermore, the step (1) includes the following sub-steps:

[0073] (1.1) Screening cases that are obviously misaligned with the anatomical structures they should represent as data with noisy labels;

[0074] (1.2) Screening cases where labels infiltrate adjacent structures due to segmentation errors as data with noisy labels;

[0075] (1.3) Screen cases with inconsistent labels or boundary conflicts between slices as data with noisy labels;

[0076] (1.4) Screen cases where labels contain random noise or do not represent the true anatomical structure as data with noisy labels;

[0077] (1.5) Screen cases with incomplete or cut-off labeled regions as data with noisy labels.

[0078] Furthermore, the weak enhancement processing is specifically as follows: randomly flipping the image upside down with a probability of 50%, randomly flipping the image left to right with a probability of 50%, and randomly performing affine transformation on the image; the strong enhancement processing is specifically as follows: randomly performing elastic transformation on the image, randomly performing CutMix processing on the image, randomly performing MaskMix processing on the image, randomly erasing part of the image content, randomly generating black squares in the image, randomly performing Gaussian blur processing on the image, randomly adding Rice noise to the image, randomly adding Gaussian noise to the image, randomly sharpening the image, and randomly performing color transformation on the image.

[0079] Furthermore, the expression of the total probability formula in step (6) is:

[0080]

[0081] in represents the prediction after adaptation, is the prediction of U-Net, p ada (j=1|i=0) is the probability of mislabeling the background as foreground, p ada (j=1|i=1) is the probability of correctly labeling the foreground. This process can be regarded as a corruption process from accurate labels to noisy labels. By adopting this approach, the model can be trained in a normal supervised manner using data containing noisy labels.

[0082] Implementation Examples

[0083] The training and testing data included 739 patients diagnosed with hepatocellular carcinoma, intrahepatic cholangiocarcinoma, or liver metastases. The inclusion criteria were as follows: (1) patients had to be 18 years of age or older to ensure that the subjects had adult physiological characteristics, (2) patients with a history of liver resection, transarterial chemoembolization, or radiofrequency ablation before CT examination were excluded, and (3) all included patients had a diagnosis of hepatocellular carcinoma, intrahepatic cholangiocarcinoma, or liver metastases confirmed by pathology reports or at least two independent imaging methods, and follow-up imaging studies lasted for at least 6 months. Finally, this study included a dataset of 739 patients.

[0084] During the training process, the present invention built a U-Net with a noise adapter, using Rectified Adam (RAdam) as the optimizer, with an initial learning rate of 0.001 and a weight decay of 0.0001. Gradients were accumulated every 2 batches, and the batch size of each subprocess was 24. The learning rate was dynamically reduced by 10 times according to the current loss trend. The present invention implemented the proposed model using PyTorch 1.2.0. The experiments were conducted on a server running Ubuntu 20.04.6LTS, which was equipped with two NVIDIA GeForce RTX 3090 graphics cards.

[0085] like Figure 3 As shown, the MPVT+ model of the present invention exhibits the most effective segmentation performance compared to the baseline model. The Sorenson-Dyess coefficient, Jacquard index, symmetric volume difference and product overlap error are used as evaluation indicators. Compared with the most effective model in the baseline model, the MPVT+ model of the present invention outperforms all other methods in various indicators including the Sorenson-Dyess coefficient, Jacquard index, symmetric volume difference and product overlap error. It is particularly noteworthy that in terms of the improvements in the Sorenson-Dyess coefficient and Jacquard index, which are key indicators of segmentation performance, the method of the present invention is 1.79% and 1.93% higher than the best performing alternative method, respectively.

[0086] In order to conduct a more intuitive qualitative assessment, Figure 4 Eight liver tumor segmentation cases are shown, with each column representing a different stage of the segmentation process. The first column shows the original CT image, and the second column is the doctor's manual annotation. The third column shows the U-Net segmentation results using ResNet34 as the backbone network. Subsequent columns show the impact of different techniques on the segmentation results, including pseudo-labeling, hard starting, mean teacher ensemble, the ELR method for preventing noisy label memory, and finally the method proposed in this invention. A detailed inspection of these columns reveals the effectiveness of the method proposed in this invention in dealing with false positives, under-segmentation, and rough boundary problems in various situations. Compared with other alternative methods, the method of this invention has significant improvements in liver segmentation accuracy.

Claims

1. A method for segmenting liver tumors on CT slice images using deep learning, characterized in that: The steps include: (1) Divide the original data set into data with noise labels and data without noise labels; (2) After weak enhancement processing, the two types of training data obtained in step (1) are input into an encoder of a U-Net-based image segmentation network for feature extraction to obtain a feature vector; (3) sending the feature vector of the training data without noise labels obtained in step (2) into the decoder of the U-Net-based image segmentation network to obtain the segmentation result of the network; (4) performing loss calculation on the segmentation result obtained in step (3) and the pre-given label, wherein the loss function is the sum of the cross-entropy loss and the Dice loss; (5) sending the feature vector of the training data with noise labels obtained in step (2) into the decoder and noise adapter of the U-Net-based image segmentation network to obtain the network segmentation result and the noise adaptation mask respectively; (6) The network segmentation result and the noise adaptation mask obtained in step (5) are used to obtain a predicted segmentation result with noise according to the full probability formula; (7) performing loss calculation on the predicted noisy segmentation result obtained in step (6) and the noisy label, wherein the loss function is the sum of the cross-entropy loss and the Dice loss; (8) performing weak enhancement processing and strong enhancement processing on the training data obtained in step (1), and then performing processing according to step (2) to obtain respective feature vectors; (9) inputting the two types of training data after weak enhancement processing obtained in step (8) into the teacher model to obtain their respective feature vector noise and pseudo labels; (10) adding the feature vector noise obtained in step (9) to the respective feature vectors obtained in step (8); (11) Processing the feature vector of the data without noise labels obtained in step (10) after adding noise according to steps (5) to (7) to obtain the corresponding loss; (12) Processing the feature vector after adding noise to the data with noise labels obtained in step (10) according to steps (3) to (4) to obtain corresponding losses; (13) The loss results obtained in step (4), step (7), step (11), and step (12) are weighted and then back-propagated to optimize network parameters and improve the accuracy of model prediction results.

2. The method for segmenting liver tumors on CT slice images using deep learning according to claim 1, characterized in that: The step (1) comprises the following sub-steps: (1.1) Screening cases that are obviously misaligned with the anatomical structures they should represent as data with noisy labels; (1.2) Screening cases where labels infiltrate adjacent structures due to segmentation errors as data with noisy labels; (1.3) Screen cases with inconsistent labels or boundary conflicts between slices as data with noisy labels; (1.4) Screen cases where labels contain random noise or do not represent the true anatomical structure as data with noisy labels; (1.5) Screen cases with incomplete or cut-off labeled regions as data with noisy labels.

3. The method for segmenting liver tumors on CT slice images using deep learning according to claim 1, characterized in that: The weak enhancement processing specifically includes: randomly flipping the image upside down with a probability of 50%, randomly flipping the image left to right with a probability of 50%, and randomly performing affine transformation on the image.

4. The method for segmenting liver tumors on CT slice images using deep learning according to claim 1, characterized in that: The strong enhancement processing is specifically: randomly performing elastic transformation on the image, randomly performing CutMix processing on the image, randomly performing MaskMix processing on the image, randomly erasing part of the image content, randomly generating black squares in the image, randomly performing Gaussian blur processing on the image, randomly adding Rice noise to the image, randomly adding Gaussian noise to the image, randomly sharpening the image and randomly performing color transformation on the image.

5. The method for segmenting liver tumors on CT slice images using deep learning according to claim 1, characterized in that: In step (6), the total probability formula is expressed as: in represents the prediction after adaptation, is the prediction of U-Net, p ada (j=1|i=0) is the probability of mislabeling the background as foreground, p ada (j=1|i=1) is the probability of correctly labeling the foreground.

6. The method for segmenting liver tumors on CT slice images using deep learning according to claim 1, characterized in that: The encoder is used to extract features of the image and gradually reduce the spatial resolution of the image to capture high-level semantic information.

7. The method for segmenting liver tumors on CT slice images using deep learning according to claim 1, characterized in that: The decoder is used to restore the feature map compressed by the encoder to the resolution of the original image and generate a segmentation result of the same size as the input image.