A semi-supervised tumor segmentation method of iterative pseudo-label self-adaptive copy-paste

Abstract

The application discloses a semi-supervised tumor segmentation method based on iterative pseudo-label adaptive copy-paste, which integrates bidirectional uncertainty into an adaptive enhancement strategy to highlight the prediction difference between teacher and student models, and gradually aligns the pseudo-label generated by the student model with the prediction of the teacher model in an iterative manner to generate pseudo-labels with more information for unlabeled samples. The method combines a bidirectional copy-paste strategy and can effectively utilize labeled and unlabeled data.

Description

Technical Field

[0001] This invention belongs to the field of pattern recognition technology, specifically relating to a semi-supervised tumor segmentation method. Background Technology

[0002] Recent advances in deep learning have had a significant impact on medical image segmentation, particularly in tumor segmentation tasks. For example, Wang et al. proposed a U-Net and ResBlock-based architecture for fully automated brain tumor segmentation in MRI images. Christ et al. used a cascaded fully convolutional neural network (FCNN) for automatic segmentation of liver and tumors. However, current medical tumor segmentation methods mainly rely on large amounts of labeled data, and the scarcity of high-quality annotated datasets and the difficulty of annotation severely hinder the development and optimization of segmentation models. In addition, many variants of U-Net and non-U-Net architectures, especially in the field of 3D medical imaging, often have a large number of parameters, which not only increases the computational complexity but also hinders the application of these models in reality. Semi-supervised learning (SSL) utilizes both labeled and unlabeled data, providing an effective solution to the challenges posed by the annotation bottleneck, and has attracted considerable attention in medical image processing. Semi-supervised learning techniques can be broadly classified into three categories: (1) Methods based on consistency regularization: For example, URPC improves the reliability of predictions by minimizing the difference between each pyramid prediction and its collective mean. (2) Self-training and pseudo-label-based methods: For example, Chaitanya et al. improved pixel-level feature learning by integrating local contrast loss and utilizing pseudo-labels of unlabeled images. (3) Adversarial learning-based methods: ACTS combines adversarial learning with consistency learning to explore a new adversarial network based on dynamic convolution and liver and skin lesion segmentation.

[0003] Despite significant progress in recent years, current semi-supervised methods still have limitations. Most semi-supervised methods, such as UA-MT and bidirectional copy-paste (BCP), are primarily designed for organs with larger shapes and fewer lesion areas, limiting their applicability to patient studies with multiple tumor lesion areas or small-volume tumors. Secondly, some methods, such as MCF and MCNet+, focus primarily on embedding additional components into the backbone network and designing hybrid loss functions, neglecting the exploration of data augmentation strategies between labeled and unlabeled data. Some semi-supervised methods train labeled and unlabeled data separately, ignoring the importance of consistency learning. Furthermore, due to differences between different tumors and between teacher and student models, current methods for improving pseudo-labels inevitably still suffer from inaccurate predictions. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention provides a semi-supervised tumor segmentation method based on iterative pseudo-label adaptive copy-paste. It incorporates bidirectional uncertainty into an adaptive enhancement strategy to highlight the prediction differences between teacher and student models. Simultaneously, it iteratively aligns the pseudo-labels generated by the student model with the predictions of the teacher model, generating more informative pseudo-labels for unlabeled samples. This method, combining a bidirectional copy-paste strategy, can effectively utilize both labeled and unlabeled data.

[0005] The technical solution adopted by this invention to solve its technical problem is as follows:

[0006] Step 1: Training set creation;

[0007] Step 1-1: Creation of the FSD (esophageal fistula tumor) dataset;

[0008] The dataset used for diagnosing esophageal fistula through images contains CT scan images of multiple patients with esophageal fistula cancer; esophageal tumors in the images are labeled.

[0009] The entire dataset is divided into a training set and a test set; in the training set, N samples are randomly selected for validation.

[0010] Steps 1-2: MSD (Medical Segmentation Decathlon) dataset creation;

[0011] The MSD dataset was created using the liver tumor segmentation challenge dataset. The test cases in the liver tumor segmentation challenge dataset were randomly divided into three subsets: training set, validation set, and test set.

[0012] Step 2: Construct the IPA-CP network model training framework;

[0013] Step 2-1: Adaptive Enhancement TUAA Based on Two-Way Uncertainty;

[0014] A Mean Teacher architecture is used to model prediction discrepancies; both weak and strong augmentation strategies are applied to unlabeled samples; for each unlabeled data sample u i Generate a random region mask M with different numbers and sizes of holes. i An adaptive augmentation strategy is used to refine the augmentation of unlabeled samples, and a two-way uncertainty estimation is applied to each weakly and strongly augmented sample.

[0015] Step 2-2: Iterative pseudo-label transformation IPT;

[0016] Fine-tuning pseudo-labels from student and teacher models using the IPT strategy. and

[0017] Steps 2-3: Two-way copy and paste supervision of BCP;

[0018] The BCP method is integrated to obfuscate the final image composition and pseudo-labels; binary masks distinguish voxels from foreground or background; the "select the largest connected component" strategy is omitted;

[0019] Step 3: Construct the training loss function;

[0020] This invention employs both cross-entropy loss (CE) and Dice loss.

[0021] Step 4: Medical image classification quality score assessment.

[0022] Preferably, step 2-1 specifically comprises:

[0023] For each unlabeled instance u i Using A w (·) Generate a weak augmented view and use A s (·) Generate a strongly augmented view; for each unlabeled data sample u i Generate a random region mask M with different numbers and sizes of holes. i Subsequently, in the case of strongly enhanced samples and mask M i A blending step is applied between the two steps; the prediction results are calculated as follows:

[0024]

[0025]

[0026] In the formula, This represents the segmentation prediction generated by the student model. Corresponding to the segmentation prediction generated by the teacher model;

[0027] Inconsistent predictions of P i dif The statement is as follows:

[0028]

[0029] P i dif =||P i s -P i t ||

[0030] Where P i s and P i t It's a fake tag;

[0031] The difference between predictions was quantified using the KL (Kullback-Leibler) divergence measure;

[0032] Two-way uncertainty estimation is applied to each weakly enhanced sample and each strongly enhanced sample;

[0033] The calculation process for the uncertainty score is as follows:

[0034]

[0035] Where Ⅱ is the index function, and τ represents a predefined threshold. This represents the divergence of the probability distribution of the student model relative to the teacher model. Represents the divergence in the opposite direction;

[0036] The weak-strong augmentation was adaptively refined, with each unlabeled sample u i Adaptive Enhanced Version The statement is as follows:

[0037]

[0038] Introducing a random binary mask M i Changes caused by unlabeled samples The following expansions have been made:

[0039]

[0040] Preferably, step 2-2 specifically includes:

[0041] The pseudo-label target is updated as follows:

[0042]

[0043] Where e is the e-th iteration of the total E training iterations; E is the total number of training rounds.

[0044] Preferably, steps 2-3 specifically include:

[0045] For labeled small batch data and unlabeled mini-batch data Randomly generate a binary mask M, where (B x |=|B u |); Binary masks distinguish voxels from foreground or background;

[0046] A two-way approach was used:

[0047] First, create a representation X by copying the background of the i-th labeled sample and pasting it onto the i-th unlabeled sample. l->uThe mixed samples; then copy the background of the j-th unlabeled sample and paste it onto the j-th labeled sample to obtain an X. u->l The mixed samples; these processes are represented by the following equations:

[0048]

[0049]

[0050] Accordingly, in order to train the student network, the BCP method is used to generate supervision signals;

[0051] Using IPT, P i mix The finely adjusted labels, which are treated as unlabeled data, are processed similarly to the true labels of labeled images:

[0052] Y l->u =M i y i +(1-M i )P i mix (8)

[0053]

[0054] in,

[0055] Preferably, step 3 specifically comprises:

[0056] The calculation formula is as follows:

[0057]

[0058]

[0059] in, Let X represent the student network model. l->u Y represents the mixed sample obtained by pasting the background of the i-th labeled sample onto the i-th unlabeled sample. l->u X represents the mixed sample obtained by pasting the background of the label corresponding to the i-th labeled sample onto the finely adjusted label of the i-th unlabeled sample. u->l Y represents the mixed sample obtained by pasting the background of the j-th unlabeled sample onto the j-th labeled sample. u->l This represents the mixed sample obtained by pasting the background of the finely adjusted label of the j-th unlabeled sample onto the corresponding label of the j-th labeled sample.

[0060] Preferably, the segmentation quantitative evaluation indicators used for the quality score assessment include: Dice coefficient (Dice), Jaccard index (JA), 95% Hausdorff distance (HD), root mean square error (RMSE), and average surface distance (ASD).

[0061] The beneficial effects of this invention are as follows:

[0062] This invention proposes a semi-supervised tumor segmentation method based on iterative pseudo-labels and adaptive copy-paste supervision. Comprehensive experiments on an internal esophageal tumor dataset and a public liver tumor dataset demonstrate the effectiveness of this method. The model framework of this invention outperforms state-of-the-art semi-supervised methods in medical image segmentation, and ablation experiments prove its structural effectiveness. Attached Figure Description

[0063] Figure 1 This is a framework diagram of the IPA-CP model of the present invention; wherein, Figure 1 (a) is a weak-strong enhancement module (WSA). Figure 1 (b) is the Adaptive Enhancement Module (TUAA) based on bidirectional uncertainty. Figure 1 (c) is the iterative pseudo-label conversion module (IPT). Figure 1 (d) is the bidirectional copy and paste supervision module (BCP).

[0064] Figure 2 This is a schematic diagram illustrating the qualitative analysis results of three cases from the FSD dataset and three cases from the MSD dataset in an embodiment of the present invention. Detailed Implementation

[0065] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0066] This invention proposes a simple yet effective method—Adaptive Copy-Paste Supervision Based on Iterative Pseudo-Labels (IPA-CP)—for tumor segmentation. Bidirectional uncertainty is incorporated into the adaptive enhancement strategy to highlight the predictive differences between the teacher and student models. Simultaneously, an iterative approach is used to progressively align the pseudo-labels generated by the student model with the predictions of the teacher model, generating more informative pseudo-labels for unlabeled samples. This method, combining a bidirectional copy-paste strategy, can effectively utilize both labeled and unlabeled data.

[0067] A semi-supervised tumor segmentation method based on iterative pseudo-label adaptive copy-paste (IPA-CP) is described below:

[0068] Step 1: Training set creation;

[0069] (1) Creation of the FSD (Esophageal Fistula Tumor) Dataset: This invention diagnoses esophageal fistula through images. The dataset used contains CT scan images of 558 patients with esophageal fistula cancer. For accurate subsequent analysis, the esophageal tumors in the images were manually labeled and verified by two radiation therapists. The entire dataset was divided into a training set and a test set, with the training set containing 390 samples and the test set containing 168 samples. In the training set, 30 samples were randomly selected for verification.

[0070] (2) MSD (Medical Segmentation Decathlon) Dataset Creation: The Liver Tumor Segmentation Challenge (LiTS) dataset used in this invention is a CT dataset focusing on the segmentation of the liver and its tumors. This dataset collects data from 7 different medical centers. It is one of the most popular liver tumor segmentation datasets, containing 200 CT scan images. These images are divided into 130 training set cases and 70 test set cases. All scans have an in-plane resolution of 512×512, but the number of axial slices varies. Since the corresponding labels for the test set have not yet been published, after removing some scan images without tumors, 116 cases from the training set were retained for analysis. These cases were randomly divided into three subsets: the training set contains 100 samples, the validation set contains 8 samples, and the test set contains 8 samples.

[0071] The second step is to construct the IPA-CP network model training framework.

[0072] (1) Adaptive Augmentation Based on Bidirectional Uncertainty (TUAA): This invention uses a Mean Teacher architecture to model prediction discrepancies. To better model prediction discrepancies, this invention employs both weak and strong augmentation strategies for unlabeled samples. To further enhance inconsistency and introduce variability into small tumor data, this invention provides a method for each unlabeled data sample u... i A random region mask M with different numbers and sizes of holes was generated. i To mitigate the potential adverse effects of strong augmentation on data distribution and segmentation performance, this invention employs an adaptive augmentation strategy to refine the augmentation of unlabeled samples. This method uses bidirectional uncertainty estimation for each weakly and strongly augmented sample, aiming to highlight uncertain regions through prediction inconsistencies. Figure 1 (a) and Figure 1 (b) presents the overall structure of the adaptive reinforcement strategy based on bidirectional uncertainty.

[0073] (2) Iterative Pseudo-Label Transformation (IPT): Pseudo-labeling is a simple and effective semi-supervised technique in the medical field. Most advanced pseudo-labeling methods emphasize the study of sampling strategies, prioritizing pseudo-labels with high confidence. However, these methods may neglect the refinement of pseudo-labels for unlabeled data in the Mean Teacher model, leading to student models tending to produce predictions with high confidence. Distinguishing between foreground and background in tumor segmentation is very challenging, making it difficult to select reliable pseudo-labels. To address this issue, the IPT strategy is proposed to fine-tune pseudo-labels from both the student and teacher models (i.e., ...). and This process makes the predictions of the student model closer to the more stable predictions of the teacher model. Figure 1 (c) provides a flow representation of this strategy.

[0074] (3) Bidirectional Copy-Paste Supervision (BCP): To simultaneously utilize knowledge from labeled and unlabeled data, the BCP method is integrated, facilitating the final image composition and pseudo-label obfuscation. Binary masks distinguish voxels from foreground (0) or background (1) (e.g., ...). Figure 1 (d) shows that the present invention omits the "select the largest connectivity component" strategy, which is a standard component of BCP. This modification is because the original strategy was primarily optimized for segmenting large organs, making it less effective when segmenting many smaller tumors.

[0075] Step 3: Construct the training loss function;

[0076] Constructing the training loss function: Since the student model is trained in a supervised manner, the choice of loss function can be flexible. To align with common practices in medical image segmentation, this invention employs both cross-entropy loss (CE) and Dice loss.

[0077] Step 4: Medical Image Classification Quality Score Assessment

[0078] Quality score indicators: Quantitative evaluation indicators for esophageal fistula segmentation include: Dice coefficient (Dice), Jaccard index (JA), 95% Hausdorff distance (HD), root mean square error (RMSE), and average surface distance (ASD).

[0079] Example:

[0080] A semi-supervised tumor segmentation method based on iterative pseudo-label adaptive copy-paste is described below:

[0081] Step 1: Training set creation;

[0082] (1) FSD (Esophageal Fistula Tumor) Dataset Creation: This invention collected chest CT scan images from 558 patients with esophageal cancer. These scan images are characterized by Huntsfield units (HU) ranging from -1024 to 3071, and voxel sizes of 0.5429 × 0.5429 × 3 mm. 3 and 1.2519×1.2519×5mm 3 The resolution of chest CT scans ranged from 512×512×71 to 512×512×114. Esophageal tumors in the images were manually labeled and validated by two radiation therapists for subsequent analysis. The entire dataset was divided into training and test sets, with the training set containing 390 samples and the test set containing 168 samples. Thirty samples were randomly selected from the training set for validation.

[0083] (2) MSD (Medical Segmentation Decathlon) Dataset Creation: The Liver Tumor Segmentation Challenge (LiTS) dataset used in this invention is a CT dataset focusing on the segmentation of the liver and its tumors. This dataset collects data from 7 different medical centers. This dataset is one of the most popular liver tumor segmentation datasets, including 200 CT scan images. These images are divided into 130 training set cases and 70 test set cases. All scans have an in-plane resolution of 512×512, but the number of axial slices varies. Since the corresponding labels for the test set have not yet been published, after removing some scan images without tumors, 116 cases from the training set were retained for analysis. These cases were randomly divided into three subsets: the training set contains 100 samples, the validation set contains 8 samples, and the test set contains 8 samples.

[0084] The second step is to construct the IPA-CP network model training framework.

[0085] (1) Adaptive Augmentation Based on Bidirectional Uncertainty (TUAA): Following existing methods, this invention uses a MeanTeacher architecture to model prediction discrepancies. This architecture has two main objectives: (1) to encourage student models to acquire more robust learning capabilities; and (2) to amplify prediction inconsistencies and introduce more variability, especially in the context of small tumor data. To better model prediction discrepancies, this invention employs both weak and strong augmentation strategies for unlabeled samples. For each unlabeled instance u... i Using A w (·) Generate a weak augmented view and use A s (·) Generate a strongly augmented view. To further enhance inconsistency and introduce variability in small tumor data, this invention generates a strongly augmented view for each unlabeled data sample u. i A random region mask M with different numbers and sizes of holes was generated. iSubsequently, in the case of strongly enhanced samples and mask M i A blending step is applied between the two steps. The prediction results are calculated as follows:

[0086]

[0087]

[0088] here, This represents the segmentation prediction generated by the student model, while This corresponds to the segmentation predictions generated by the teacher model. Finally, the P-values ​​with inconsistent predictions... i dif This can be expressed as follows:

[0089]

[0090] P i dif =||P i s -P i t ||

[0091] Where P i s and P i t It's a fake tag.

[0092] In semi-supervised tumor segmentation, accurately assessing the uncertainty of unlabeled data faces two challenges: the lack of precise labels and the need to effectively capture subtle differences between teacher and student model predictions. To address these challenges, a bidirectional uncertainty estimation strategy is designed to evaluate the inconsistency between segmentation results generated by the two models. This invention uses KL (Kullback-Leibler) divergence to quantify the differences between predictions. To mitigate the potential adverse effects of strong augmentation on data distribution and segmentation performance, this invention develops an adaptive augmentation strategy to refine the augmentation of unlabeled samples. This method employs bidirectional uncertainty estimation for each weakly and strongly augmented sample, aiming to highlight uncertain regions through prediction inconsistency. Initially, a symmetric uncertainty thresholding method is implemented, focusing on prediction regions with high uncertainty. The uncertainty score is calculated as follows:

[0093]

[0094] Where Ⅱ is the index function, and τ represents a predefined threshold used to filter out voxels with high uncertainty levels. This represents the divergence of the probability distribution of the student model relative to the teacher model, while This represents the divergence in the opposite direction. Thus, samples exhibiting greater uncertainty in both the teacher and student models are assigned larger μ values. This invention further adaptively refines the weak-strong enhancement. Each unlabeled sample u... i Adaptive Enhanced Version It is described as follows:

[0095]

[0096] To simulate discrepancies in predictions, this invention introduces a random binary mask M. i The resulting changes. Unlabeled samples The following expansions have been made:

[0097]

[0098] In this way, the diversity of tumor characteristics is continuously enhanced.

[0099] (2) Iterative Pseudo-Label Transformation (IPT): Pseudo-labeling is a simple and effective semi-supervised technique in the medical field. Most advanced pseudo-labeling methods emphasize the study of sampling strategies and prioritize high-confidence pseudo-labels. However, these methods may neglect the refinement of pseudo-labels for unlabeled data in the Mean Teacher model, leading to student models tending to produce predictions with high confidence. Distinguishing between foreground and background in tumor segmentation is very challenging, making it difficult to select reliable pseudo-labels. To address this issue, the IPT strategy is proposed to fine-tune pseudo-labels from both the student and teacher models (i.e., P...). i s and P i t This process makes the student model's predictions closer to the more stable predictions of the teacher model. The pseudo-label target is updated as follows:

[0100]

[0101] Where e represents the e-th iteration in a total of E training iterations. In practical implementation, the number of training rounds is denoted as e, and E is the total number of training rounds. Therefore, this invention generates more robust and diverse pseudo-labels for unlabeled samples. Figure 1 (c) provides a flow representation of this strategy.

[0102] (3) Bidirectional Copy-Paste Supervision (BCP): To simultaneously utilize knowledge from both labeled and unlabeled data, the BCP method described in [reference needed] is integrated, facilitating the final image composition and pseudo-label obfuscation. This is for labeled mini-batch data. and unlabeled mini-batch data Randomly generate a binary mask M, where (B x |=|B u|). A binary mask distinguishes a voxel from the foreground (0) or background (1) (e.g. Figure 1 (d) shows the method used in this invention: First, by copying the background of the i-th labeled sample and pasting it onto the i-th unlabeled sample, a representation X is created. l->u The mixed samples. Then copy the background of the j-th unlabeled sample and paste it onto the j-th labeled sample to obtain an X. u->l The mixed samples. These processes can be represented by the following equations:

[0103]

[0104]

[0105] Accordingly, to train the student network, the BCP method is used to generate supervisory signals. Using IPT technology, P... i mix The labels are treated as finely adjusted labels for unlabeled data. The true labels for labeled images are processed in a similar way:

[0106] Y l->u =M i y i +(1-M i )P i mix (8)

[0107]

[0108] in, This invention omits the "select the largest connectivity component" strategy, a standard component of BCP. This modification is because the original strategy was primarily optimized for segmenting large organs, making it less effective when segmenting many smaller tumors.

[0109] Step 3: Construct the training loss function;

[0110] Constructing the training loss function: Since the student model is trained in a supervised manner, the choice of loss function can be flexible. To align with common practices in medical image segmentation, this invention employs both cross-entropy loss (CE) and Dice loss. The calculation formulas are as follows:

[0111]

[0112]

[0113] in, Let X represent the student network model. l->uY represents the mixed sample obtained by pasting the background of the i-th labeled sample onto the i-th unlabeled sample. l->u X represents the mixed sample obtained by pasting the background of the label corresponding to the i-th labeled sample onto the finely adjusted label of the i-th unlabeled sample. u->l Y represents the mixed sample obtained by pasting the background of the j-th unlabeled sample onto the j-th labeled sample. u->l This represents the mixed sample obtained by pasting the background of the finely adjusted label of the j-th unlabeled sample onto the corresponding label of the j-th labeled sample.

[0114] Step 4: Implementation details and evaluation measures;

[0115] (1) Implementation Details: The method of this invention was implemented using an NVIDIA RTX 3090 graphics card within the PyTorch and MONAI frameworks. The batch size was set to 2, the esophageal tumor patch size was 160×160×80, and the liver tumor patch size was 112×112×64. The Adam optimizer employed a multinomial learning rate strategy, where the initial learning rate was 2.5×10⁻⁴ multiplied by... The power was set to 0.9. The total number of training iterations was set to 6000 for esophageal tumors and 42000 for liver tumors. The EMA decay weight factor α was set to 0.99. A 3D U-Net was used as the base network. All performance data were obtained as the average of three runs.

[0116] (2) Assessment measures: The quantitative evaluation indicators for esophageal fistula segmentation include: Dice coefficient (Dice), Jaccard index (JA), 95% Hausdorff distance (HD), root mean square error (RMSE), and average surface distance (ASD).

[0117] This invention proposes a semi-supervised tumor segmentation method based on iterative pseudo-labels and adaptive copy-paste supervision, such as... Figure 1This method segmentes and diagnoses esophageal fistulas using CT scans. It employs a simple yet effective adaptive enhancement technique based on bidirectional inconsistency to increase the diversity of tumor representations. An iterative pseudo-label transformation strategy guides predictions towards more robust pseudo-labels, and a bidirectional copy-paste (BCP) mechanism facilitates model learning from both labeled and unlabeled data. Furthermore, the method was thoroughly tested on a self-constructed internal esophageal tumor dataset and a publicly available liver tumor dataset to validate its effectiveness. On the FSD dataset, it demonstrated superior performance across all five evaluation metrics at 10% and 20% labeled data ratios, outperforming other semi-supervised methods such as ICT, CCT, URPC, CPS, RD, ACL, BCP, and MCF. As shown in Table 1, with only 10% labeled data, IPA-CP significantly outperformed the second-best model, increasing the Dice score by 8% compared to the baseline model. Moreover, it is noteworthy that the overall performance of this method is close to that of fully supervised learning using a 100% labeled dataset. These results validate the effectiveness of IPA-CP in esophageal tumor segmentation.

[0118] To visually demonstrate the small tumor segmentation capabilities of IPA-CP, qualitative experiments were conducted on the FSD and MSD datasets. Visualizations of several typical segmentation results are shown below. Figure 2 As shown. For larger tumors, such as the esophageal tumor case in the second row, the prediction results of IPA-CP are basically consistent with the true label, while other methods either produce a large number of false positives or miss key tumor areas. For small tumors, such as Figure 2 In the fourth line, IPA-CP successfully captured and identified small tumors on the MSD dataset, while other semi-supervised methods failed to do so. The IPA-CP method designed in this invention exhibits more accurate tumor segmentation results, while other methods often missegment non-target areas or fail to identify all tumors. This is mainly attributed to the presence of a large number of blurry tumor voxels with significant size differences in CT images, and models initially designed for larger organs often perform poorly in segmenting these voxels.

[0119] This invention conducted ablation studies to validate the effectiveness of each proposed component, including WSA, BCP, TUE, IPT, and PDI under the Mean Teacher architecture. The results are shown in Table 2. Ablation experiments were conducted using 10% / 90% (labeled / unlabeled) samples from the FSD / MSD datasets. As the analysis in Table 2 shows, on the FSD dataset, introducing WSA alone improved Dice by approximately 2%. BCP significantly improved the performance of all metrics by combining labeled and augmented unlabeled tumor data to enhance data diversity. Furthermore, combining TUE with WSA improved Dice by over 3%, highlighting the impact of flexible data augmentation on model effectiveness. On the MSD dataset, the introduction of BCP and TUE significantly improved the performance of the proposed model compared to the baseline model, increasing Dice from 33.73% to 43.98%. This is likely due to the effective use of TUE in refining over-augmented data. The performance improvement brought by IPT (37.23% vs. 38.22%) demonstrates the importance of improving pseudo-labels; IPT effectively brings predictions closer to the more stable predictions of the self-teacher model. The improvements resulting from the introduction of PDI validate that PDI has a more accurate segmentation effect when dealing with relatively small tumors.

[0120] Figure 2 Qualitative analysis results are presented for three cases from the FSD dataset and three cases from the MSD dataset. From left to right, the results show the 3D tumor segmentation generated by different state-of-the-art SSL methods on 10% labeled data and 90% unlabeled data, with the last column displaying the segmentation labels.

[0121] Table 1

[0122]

[0123] Table 2 shows the segmentation performance in ablation experiments on the FSD / MSD dataset using 10% (36 / 10) labeled data.

[0124]

[0125] Table 1 shows the experimental results using the IPA-CP method of this invention and the state-of-the-art SSL method on the FSD and MSD datasets. Note that 10% (36 / 10) indicates that 10% of the data is labeled, and the corresponding number of labeled data in the MSD and FSD datasets is 36 / 10.

Claims

1. A semi-supervised tumor segmentation method based on iterative pseudo-label adaptive copy-paste, characterized in that, Includes the following steps: Step 1: Training set creation; Step 1-1: Creating the FSD dataset; The dataset used for diagnosing esophageal fistula through images contains CT scan images of multiple patients with esophageal fistula cancer; esophageal tumors in the images are labeled. The entire dataset is divided into a training set and a test set; in the training set, N samples are randomly selected for validation. Steps 1-2: Creating the MSD dataset; The MSD dataset was created using the liver tumor segmentation challenge dataset. The test cases in the liver tumor segmentation challenge dataset were randomly divided into three subsets: training set, validation set, and test set. Step 2: Construct the IPA-CP network model training framework; Step 2-1: Adaptive Enhancement TUAA Based on Two-Way Uncertainty; A Mean Teacher architecture is used to model the prediction discrepancies; both weak and strong augmentation strategies are applied to unlabeled samples. For each unlabeled data sample Generate a random region mask with different numbers and sizes of holes. An adaptive augmentation strategy is used to refine the augmentation of unlabeled samples, and a two-way uncertainty estimation is applied to each weakly and strongly augmented sample. Step 2-1 specifically involves: For each unlabeled instance ,use Generate a weak augmented view and use Generate a strongly augmented view; For each unlabeled data sample Generate a random region mask with different numbers and sizes of holes. ; Subsequently, in the strongly enhanced samples and masks A blending step is applied between the two steps; the prediction results are calculated as follows: ，（1） In the formula, This represents the segmentation prediction generated by the student model. Corresponding to the segmentation prediction generated by the teacher model; Inconsistent predictions The statement is as follows: ， （2） in and It's a fake tag; Use KL divergence to measure the differences between predictions; Two-way uncertainty estimation is applied to each weakly enhanced sample and each strongly enhanced sample; The calculation process for the uncertainty score is as follows: （3） in, It is an index function. This represents a predefined threshold. This represents the divergence of the probability distribution of the student model relative to the teacher model. Represents the divergence in the opposite direction; The weak-strong augmentation was adaptively refined for each unlabeled sample. Adaptive Enhanced Version The statement is as follows: （4） Introducing a random binary mask Changes caused by unlabeled samples The following expansions have been made: （5） Step 2-2: Iterative pseudo-label transformation IPT; Fine-tuning pseudo-labels from student and teacher models using the IPT strategy. and ; Step 2-2 specifically involves: The pseudo-label target is updated as follows: （6） in, For the total The training iteration of the ... The next iteration; This represents the total number of training rounds. Steps 2-3: Two-way copy and paste supervision of BCP; The BCP method is integrated to obfuscate the final image composition and pseudo-labels; binary masks distinguish voxels from foreground or background; the "select the largest connected component" strategy is omitted; Step 3: Construct the training loss function; Simultaneously employing cross-entropy loss (CE) and Dice loss; Step 4: Medical image classification quality score assessment.

2. The semi-supervised tumor segmentation method with iterative pseudo-label adaptive copy-paste according to claim 1, characterized in that, Steps 2-3 specifically refer to: Mini-batch data for labeling and unlabeled mini-batch data Randomly generate binary mask ,in Binary masks distinguish voxels from the foreground or background. A two-way approach was used: First, by copying the first The background of the first marked sample is pasted into the first... On each unlabeled sample, create a representation as Mixed samples; then copy the first The background of the unlabeled sample, and pasted into the first On each labeled sample, a The mixed samples; these processes are represented by the following equations: ，（7） Accordingly, in order to train the student network, the BCP method is used to generate supervision signals; Using IPT, The finely adjusted labels, which are treated as unlabeled data, are processed in a similar way to the true labels of labeled images: ，（8） in, .

3. The semi-supervised tumor segmentation method with iterative pseudo-label adaptive copy-paste according to claim 2, characterized in that, Step 3 specifically involves: The calculation formula is as follows: ，（9） in, Representing the student network model, Indicates the first The background of the marked sample is pasted to the first... The mixed sample obtained from the unlabeled sample Indicates the first Paste the background of the label corresponding to the first marked sample to the first... A mixed sample obtained by finely adjusting the labels on unlabeled samples. Indicates the first The background of the unlabeled sample is pasted to the first... The mixed sample obtained from the labeled samples Indicates the first The background of the finely adjusted label for the unlabeled sample was pasted onto the first... Each labeled sample corresponds to a mixed sample obtained from the labels.

4. The semi-supervised tumor segmentation method with iterative pseudo-label adaptive copy-paste according to claim 2, characterized in that, The quantitative evaluation indicators used for the quality score assessment include: Dice coefficient (Dice), Jaccard index (JA), 95% Hausdorff distance (HD), root mean square error (RMSE), and average surface distance (ASD).

Images