Weak supervision lesion segmentation method and device for partial instance labeling

Abstract

The invention provides a weak supervision lesion segmentation method and device for partial instance annotation, and belongs to the technical field of image processing, and the method is used for constructing a meta-learning label correction and feature comparison learning framework based on a segmentation model. In each training iteration of the learning framework, the following steps are executed in sequence: performing false label correction on a part of voxels with the highest foreground probability in an unlabeled region through a selective meta-label correction mechanism; performing semantic consistency verification on a correction result by utilizing a category prototype through a prototype-guided reliability filtering mechanism; and through a joint optimization training mechanism, segmentation model parameters are optimized in combination with segmentation loss and prototype regularization loss, and the prototype regularization loss comprises comparison alignment loss and diversity loss so as to improve category discrimination and feature diversity. According to the method, the potential lesions in the unlabeled region are mined, so that the robustness and generalization ability of the model under the extreme sparse supervision condition are comprehensively improved.

Description

Technical Field

[0001] This invention relates to the field of image processing and vision technology, and in particular to a weakly supervised lesion segmentation method and apparatus for partial instance annotation. Background Technology

[0002] Medical imaging plays a vital role in modern clinical diagnosis and treatment, especially in the detection, diagnosis, and treatment evaluation of diseases such as tumors, inflammation, and hemorrhage. Imaging data provides physicians with rich information on tissue structure and lesions. Precise segmentation of lesion areas not only facilitates quantitative analysis and severity assessment of lesions but also provides crucial support for downstream tasks such as surgical planning and radiotherapy target delineation. Therefore, researching efficient and accurate automated lesion segmentation algorithms has significant clinical implications.

[0003] In recent years, deep learning-based medical image analysis methods have developed rapidly, significantly improving the performance of image segmentation tasks on multiple public datasets. However, the superior performance of these methods largely depends on large-scale, pixel-level fine-annotation training data. In the field of medical imaging, obtaining high-quality annotations still faces many challenges: on the one hand, lesion boundaries are often blurred and have low contrast with surrounding tissues, making the annotation process susceptible to subjective judgment; on the other hand, pixel-by-pixel annotation needs to be completed by professional doctors, a tedious and time-consuming process that severely restricts the construction and application of large-scale, high-quality datasets.

[0004] To alleviate the problem of insufficient labeled data, weakly supervised learning has become an important research direction in the field of medical image segmentation. This type of method trains segmentation models using only weakly labeled information such as image-level labels, points, bounding boxes, or doodles, thereby reducing reliance on full-pixel annotations. Among commonly used weak annotation forms, point annotations provide the approximate location of lesions, bounding boxes define their spatial extent, and doodle annotations typically draw several curves at the center of the lesion or in the skeletal region. These annotation forms achieve a certain balance between annotation cost and spatial cues, but due to the general lack of detailed depiction of lesion boundaries and morphology, the model struggles to fully learn its structural details, and segmentation performance still faces significant bottlenecks.

[0005] To overcome the aforementioned limitations, partial instance annotation represents a more forward-looking weakly supervised annotation strategy. This method provides complete pixel-level annotations only for a few lesion instances in 3D medical images, while leaving the rest unannotated. This approach not only significantly reduces the overall annotation cost, but more importantly, it provides the model with precise morphological and boundary information within the lesions. Compared to sparse annotations such as points and boxes, the complete instance information contained in partial instance annotations can more effectively guide the model to learn the true structural features of lesions. Furthermore, compared to fully supervised annotation, it sacrifices the "completeness" of the annotation quantity in exchange for "high feasibility" in annotation practice, better aligning with the clinical workflow where doctors often only delineate representative lesions.

[0006] Despite the unique advantages of this strategy in clinical practice and performance potential, related research is still in its early exploratory stages, with significant gaps remaining. Most existing methods set high annotation ratios or only perform partial annotation at the 2D slice level, failing to systematically explore extremely sparse supervision scenarios at the level of complete 3D images with very low annotation ratios (e.g., only one instance annotated per 3D image). More importantly, traditional weakly supervised methods are generally based on the assumption that "each lesion possesses some kind of supervisory signal," making it difficult to directly address the challenges of missing labels and class imbalance caused by a large number of unannotated lesions in some instances. This leads to models easily overfitting to the few labeled instances while ignoring a large number of unlabeled potential lesion regions.

[0007] Therefore, for the specific paradigm of partial instance annotation, there is an urgent need to develop new segmentation algorithms that can overcome the problem of large-area missing foreground annotations, so as to discover potential lesions in unlabeled regions and thus comprehensively improve the robustness and generalization ability of the model under extremely sparse supervision conditions. Summary of the Invention

[0008] This invention provides a weakly supervised lesion segmentation method and apparatus for partial instance annotation, thereby overcoming the challenge of severely insufficient supervision signals under partial instance annotation.

[0009] In a first aspect, the present invention provides a weakly supervised lesion segmentation method for partial instance annotation, used to construct a meta-learning label correction and feature comparison learning framework based on a segmentation model; in each training iteration of the learning framework, the following steps are executed sequentially: By using a selective meta-label correction mechanism, pseudo-label correction is performed on a subset of voxels with the highest foreground probability in unlabeled areas; A prototype-guided reliability filtering mechanism is used to verify the semantic consistency of the correction results using category prototypes. By using a joint optimization training mechanism, the segmentation model parameters are optimized by combining segmentation loss and prototype regularization loss. The prototype regularization loss includes contrast alignment loss and diversity loss to improve class discriminativeness and feature diversity.

[0010] The weakly supervised lesion segmentation method for partial instance annotation provided by the present invention performs pseudo-label correction on a subset of voxels with the highest foreground probability in unlabeled regions through a selective meta-label correction mechanism. This includes: given a 3D input image and its foreground probability map, excluding labeled voxels using a binary mask, and selecting the top K voxels with the highest probability from the unlabeled voxels as a correction candidate set; assigning a learnable correction offset to each voxel in the correction candidate set, and calculating the meta-gradient through a meta-optimization step, wherein the meta-gradient is obtained based on backpropagation of the validation set loss; determining the label update direction based on the sign of the meta-gradient, constructing a binary vector, and updating the pseudo-labels using an exponential moving average strategy.

[0011] In the weakly supervised lesion segmentation method for partial instance annotation provided by the present invention, the number K of candidate voxels is calculated by a ratio r.

[0012] The weakly supervised lesion segmentation method for partial instance annotation provided by the present invention uses a prototype-guided reliability filtering mechanism to verify the semantic consistency of the correction results using category prototypes. This includes: learning M category prototypes for each category on the output feature map of the last layer of the segmentation model decoder; optimizing the semantic alignment of the prototype representation with the feature embedding through a cross-attention module; for candidate voxels corrected by meta-learning, extracting the feature vector of the candidate voxel after cross-attention optimization, and calculating the cosine similarity with each category prototype; assigning prototype category labels according to the principle of highest similarity, and constructing a binary filtering mask to remove voxels whose correction direction is inconsistent with their respective prototype labels.

[0013] According to the weakly supervised lesion segmentation method for partial instance annotation provided by the present invention, the segmentation loss in the joint optimization training mechanism is a combination of cross-entropy loss and Dice loss, and the weight of the prototype regularization loss is controlled by hyperparameters.

[0014] According to the weakly supervised lesion segmentation method for partial instance annotation provided by the present invention, the segmentation model adopts a 3D U-Net architecture, wherein the encoder is used to extract multi-scale features and the decoder is used to progressively restore spatial resolution and output segmentation results.

[0015] According to the present invention, a weakly supervised lesion segmentation method based on partial instance annotation is applied to lesion segmentation in three-dimensional medical images.

[0016] Secondly, the present invention also provides a weakly supervised lesion segmentation device for partial instance annotation, used to implement a meta-learning label correction and feature comparison learning framework based on a segmentation model; in each training iteration of the learning framework, the following modules are run sequentially: The meta-label correction module is configured to perform pseudo-label correction on a subset of voxels with the highest foreground probability in the unlabeled region through a selective meta-label correction mechanism. The reliability filtering module is configured to use a prototype-guided reliability filtering mechanism to perform semantic consistency verification of the correction results using category prototypes. The joint optimization module is configured to combine segmentation loss and prototype regularization loss to optimize segmentation model parameters. The prototype regularization loss includes contrast alignment loss and diversity loss to improve class discriminativeness and feature diversity.

[0017] Thirdly, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the weakly supervised lesion segmentation method for partial instance annotation as described above.

[0018] Fourthly, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the weakly supervised lesion segmentation method for partial instance annotation as described above.

[0019] The weakly supervised lesion segmentation method and apparatus for partial instance annotation provided by this invention have the following advantages compared with the prior art: (1) This invention innovatively utilizes the complete lesion morphology information provided by partial instance annotations, and effectively transfers the structural features of the annotated lesions to the unannotated areas through a meta-learning framework, significantly improving the boundary segmentation accuracy.

[0020] (2) The meta-learning label correction mechanism provided by the present invention effectively addresses the challenge of a large number of lesions being completely unlabeled, and prevents the model from overfitting to labeled instances.

[0021] (3) The prototype-guided filtering mechanism provided by the present invention reduces the impact of subjective annotation bias on the model through objective semantic consistency verification. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating the weakly supervised lesion segmentation method for partial instance annotation provided by the present invention. Figure 2 This is a schematic diagram of the model framework provided by the present invention; Figure 3 This is a visual comparison diagram of the segmentation results of the LiTS dataset provided by this invention; Figure 4 This is a visual comparison diagram of the segmentation results of the ISLES22 and MS datasets provided by this invention; Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0025] It should be noted that, in the description of the embodiments of the present invention, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. Those skilled in the art can understand the specific meaning of the above terms in the present invention according to the specific circumstances.

[0026] The following is combined Figures 1-5 This invention describes a weakly supervised lesion segmentation method and apparatus for partial instance annotation provided in embodiments of the present invention.

[0027] This invention provides a weakly supervised lesion segmentation method and apparatus for partially labeled instances. The core of the technical solution of this invention lies in constructing a label correction and feature comparison learning framework based on meta-learning, which aims to effectively utilize the complete morphological information in sparsely labeled instances, while mining potential lesions in unlabeled areas, thereby overcoming the challenge of severely insufficient supervision signals under partially labeled instances.

[0028] Figure 1This is a flowchart illustrating the weakly supervised lesion segmentation method for partial instance annotation provided by the present invention. Figure 2 This is a schematic diagram of the model framework provided by the present invention (showing 3D multi-lesion segmentation based on partial instances and graffiti annotations).

[0029] See Figure 1 and Figure 2 This method executes the following three core mechanisms sequentially in each training iteration: Mechanism 1: Selective Meta-Label Correction Mechanism This mechanism aims to achieve accurate pseudo-label correction for key foreground pixels in unlabeled regions at an affordable computational cost. Based on a model-predicted foreground probability map, it sorts unlabeled pixels after excluding labeled regions, selecting only the highest-probability pixels as a correction candidate set, and assigning a learnable correction offset to each candidate pixel. The meta-gradient is calculated through a meta-optimization step based on validation loss to determine the correction direction, and an exponential moving average strategy is used to achieve smooth pseudo-label updates. This mechanism significantly reduces computational complexity while effectively improving the accuracy and training stability of pseudo-labels.

[0030] Mechanism 2: Prototype-guided reliability filtering mechanism To address the meta-correction bias caused by differences in lesion morphology, this mechanism introduces category prototypes to verify the reliability of the correction results. First, several prototypes are learned for each category, and a lightweight cross-attention module collaboratively optimizes the prototype representation and feature embedding. Then, for candidate pixels corrected by meta-learning, the cosine similarity between their optimized features and the prototypes of each category is calculated, and corresponding prototype category labels are assigned. By constructing a binary filtering mask, candidate pixels whose correction direction is inconsistent with the semantics of their respective prototypes are systematically removed, ensuring that only semantically consistent and reliable corrections are retained for subsequent training, significantly improving the semantic consistency and robustness of pseudo-label correction.

[0031] Mechanism 3: Joint Optimization Training Mechanism To enhance the model's class discrimination ability and semantic coverage, this mechanism employs a joint optimization of segmentation loss and prototype regularization loss during model training. First, two complementary prototype regularization losses are designed: contrastive alignment loss encourages optimized features of each class to cluster towards their corresponding prototypes and distance them from prototypes of other classes; diversity loss penalizes similarity between prototypes of the same class, ensuring that each prototype captures unique semantic patterns and effectively preventing prototype collapse. Building upon this, the filtered reliable pseudo-labels are combined with ground truth annotations to construct a joint objective function incorporating cross-entropy loss and prototype regularization loss. End-to-end training synchronously optimizes the segmentation model parameters and prototype representations, ultimately achieving robust segmentation performance under sparse supervision.

[0032] To make the purpose, technical solution, and advantages of this invention patent clearer, the invention will be further described in detail below with reference to the figures and embodiments.

[0033] Mechanism 1: Selective Meta-Learning Label Correction Optionally, a selective meta-label correction mechanism is used to perform pseudo-label correction on a subset of voxels with the highest foreground probability in the unlabeled region. This includes: given a 3D input image and its foreground probability map, using a binary mask to exclude labeled voxels, and selecting the top K voxels with the highest probability from the unlabeled voxels as a correction candidate set; assigning a learnable correction offset to each voxel in the correction candidate set, and calculating the meta-gradient through a meta-optimization step, wherein the meta-gradient is obtained based on backpropagation of the validation set loss; determining the label update direction based on the sign of the meta-gradient, constructing a binary update vector, and combining it with an exponential moving average strategy to update the pseudo-labels.

[0034] Specifically, performing voxel-level meta-learning label correction in 3D medical images incurs high computational costs, and model training instability and even crashes have been observed in experiments. Furthermore, under the PIS annotation setting adopted in this chapter, manually provided annotations are considered accurate, and therefore label correction is no longer performed on these regions. Based on these considerations, this paper proposes a selective meta-learning label correction strategy, which corrects only the voxels most likely to be lesions in unlabeled regions, thus significantly reducing computational complexity. Simultaneously, since the optimization focuses only on a small number of key voxels, the accuracy of the correction results is also improved. Experiments demonstrate that even when only a very small proportion (e.g., 0.05%) of voxels are selected for the meta-learning process in each round, the model can still maintain good overall segmentation performance while improving lesion recall.

[0035] Given a 3D input image ,Include Individual elements. Let the foreground probability map generated during the virtual training phase be... To simplify the notation, this paper uses three-dimensional voxel coordinates. Mapped to linear index Using binary masks Exclude all voxels that have been manually labeled (including foreground lesions and background graffiti), among which Voxel representation If a voxel is labeled, it is labeled; otherwise, it is unlabeled. For unlabeled voxels, the foreground probability is used as the basis for the labeling. Sort from highest to lowest, and select the top... Individual element composition correction candidate set .make Indicates the sorting result of the first... The selection rule for the foreground probability value corresponding to the voxel is as follows:

[0036] Only for the correction candidate set This Individual morphology is assigned a learnable correction offset, which forms the correction vector. At the same time, a transformation operator is introduced. Used to define in a set any dimensional vector (like This maps back to the complete voxel space (or the label update vector mentioned later). The transformation is defined as follows:

[0037] In each iteration, the current label is used first. and initial correction amount Constructing pseudo tags :

[0038] Subsequently, samples were collected from the training set based on pseudo-labels. For each sample, perform a "virtual" parameter update to obtain temporary model parameters. :

[0039] in, Indicates the learning rate. Let represent the single-sample segmentation loss function, which is a weighted combination of cross-entropy loss and Dice loss. Next, the temporary model is evaluated on the validation set. The correction amount for each candidate voxel is calculated by minimizing the validation loss. The meta-gradient. Validation set samples have complete labels and can be considered as meta-knowledge containing information about the true distribution. The formula for calculating the meta-gradient is:

[0040] According to the metagradient The symbol is used to determine the label update direction for each voxel, and a binary vector is constructed. :

[0041] To ensure the stability and gradual nature of label updates, this paper uses a candidate voxel set... An Exponential Moving Average (EMA) strategy is introduced to smoothly update the label values. The specific update formula is as follows:

[0042] in Controlling the smoothness. Ultimately achieved through a transformation operator. Map the updated label values ​​back to 3D voxel space:

[0043] Although the correction process focuses only on a small number of unlabeled voxels that are most likely lesions, some correction results may still be biased. Therefore, a prototype-based correction screening mechanism will be introduced below to remove potential erroneous labels and improve the reliability of the correction.

[0044] Mechanism 2: Prototype-guided correction label filtering Optionally, a prototype-guided reliability filtering mechanism is used to verify the semantic consistency of the correction results using category prototypes. This includes: learning M category prototypes for each category on the output feature map of the last layer of the segmentation model decoder, and optimizing the semantic alignment between the prototype representation and the feature embedding through a cross-attention module; for candidate voxels corrected by meta-learning, extracting the feature vector of the candidate voxel after cross-attention optimization, and calculating the cosine similarity with each category prototype; assigning prototype category labels according to the principle of highest similarity, and constructing a binary filtering mask to remove voxels whose correction direction is inconsistent with their respective prototype labels.

[0045] Even within the same category, different lesions may exhibit significant differences in morphology, intensity, and other aspects, causing the features of certain voxels to deviate from the typical distribution across all categories. This makes them prone to incorrect updates during meta-learning label correction. To mitigate this issue, this paper constructs a set of class prototypes in the last layer of the 3D U-Net decoder and introduces a lightweight two-stage cross-attention module to optimize the semantic alignment between class prototypes and image features, and to assist in identifying unreliable label correction results.

[0046] Let the output feature map of the last layer of the decoder be... ,in This represents the number of channels. Each voxel feature in 3D space is vectorized as... To better model the feature diversity within the background and lesion classes, this paper constructs a model for each class. The learnable class prototypes are represented as follows: (Background class prototype) (Foreground class prototype). Among them, the first... Class prototype Indicates category This is a feature pattern. In the two-stage cross-attention module, the first stage uses the class prototype as the query vector, maps the image feature map to keys and values, calculates attention weights, and updates the prototypes accordingly to obtain more semantically representative class prototypes. The second stage uses the updated class prototype as the key and value, and the image features as the query vector, to perform reverse attention interaction, resulting in an optimized voxel-level feature representation. This enhances the semantic consistency between the prototype and the class prototype. This module helps improve the prototype's ability to distinguish the semantic categories of voxels, which is then used for subsequent reliability screening. Since this module is not the core innovation of this paper, its specific calculation process will not be elaborated here.

[0047] Obtaining the candidate voxel index set and its corresponding meta-learning correction direction Then, the optimized feature vectors corresponding to these voxels are extracted. For each voxel Calculate its relationship with each class prototype Cosine similarity:

[0048] Subsequently, each voxel is assigned a label corresponding to its class prototype with the highest similarity, calculated as follows:

[0049] Then, based on the consistency between the class prototype label and the meta-learning correction direction, a binary mask is constructed. Correction directions that are inconsistent with semantics:

[0050] set up Represents the set in the previous iteration The label on The meta-learning correction result obtained from formula (7) is represented by the following updated labels after filtering:

[0051] in This represents element-wise multiplication. Finally, it is achieved through the mapping operator. The filtered labels are restored to the complete voxel space to obtain the updated label map:

[0052] The aforementioned prototype-based filtering mechanism effectively filters out incorrect correction labels that are inconsistent with the semantics of the foreground or background, significantly improving the robustness of meta-learning label correction.

[0053] Mechanism 3: Joint Optimization Training Mechanism Optionally, the segmentation loss in the joint optimization training mechanism is a combination of cross-entropy loss and Dice loss, and the weight of the prototype regularization loss is controlled by a hyperparameter.

[0054] Specifically, in order to improve the discriminativeness of class prototypes and the coverage of semantic expression, this invention designs two complementary regularization losses to constrain the feature maps and class prototypes output by the cross-attention module.

[0055] First, a contrastive alignment loss is introduced to encourage the feature distribution of each class to be closer to the prototype mean of its own class, while moving away from prototypes of other classes, thereby enhancing class discriminative power. For any class... ,set up Indicates the class The updated prototype mean. This represents the feature mean of this class of voxels. The similarity of positive samples is defined as... The similarity of negative samples is ,in Let represent cosine similarity. Then, the contrastive loss for each class is defined as:

[0056] in This is a temperature coefficient used to adjust the smoothness of the similarity distribution. The total contrast alignment loss is the sum of the two types of losses. .

[0057] Secondly, to prevent class prototype collapse and encourage each prototype to capture distinct semantic patterns, this paper introduces a diversity loss term, which improves the expressive power of prototypes by penalizing the similarity between prototypes within the same category. The specific definition is as follows:

[0058] in Let be the Euclidean norm. Combining the contrast alignment loss and the diversity loss constitutes the complete prototype regularization loss, expressed as:

[0059] Among them, hyperparameters , Control the relative weights of the two regularization losses.

[0060] After defining the prototype regularization mechanism, this paper integrates it into the actual training process of the model. Based on the pseudo-labels obtained in the previous section... In each small batch On each training sample, the model parameters are optimized by minimizing the sum of the segmentation loss (a combination of cross-entropy and Dice loss) and the prototype regularization term. The updated formula is as follows:

[0061] The experimental evaluation is as follows: Figure 3 This is a visualization comparison diagram of the segmentation results of the LiTS dataset provided by the present invention. Each row shows the segmentation performance of different methods on the same representative slice, and the superimposed orange numbers represent the overall Dice coefficient of the 3D image to which the slice belongs. Figure 4 This is a visual comparison diagram of the segmentation results of the ISLES22 and MS datasets provided by this invention. Each row shows the segmentation performance of different methods on the same representative slice.

[0062] The results show that on both datasets, noise-robust methods (such as RMD and L2B) generally exhibit undersegmentation, capturing only a portion of the lesion area. In contrast, partial instance-based and graffiti-based methods (such as DBDMP and FDDSeg) tend to oversegment, with their predicted lesion regions often intruding into healthy tissue. The undersegmentation in noise-robust methods stems from their treatment of all unlabeled lesions as noise anomalies during training, hindering the network's learning of these lesion features. In contrast, partial instance-based methods typically rely on auxiliary tasks (such as image reconstruction or classification) to compensate for missing background annotations. However, this indirect supervision signal is weak, making accurate lesion-background separation difficult and prone to oversegmentation. Similarly, graffiti-based methods extrapolate sparse annotations to dense pseudo-labels, but this extrapolation often crosses lesion boundaries, reducing segmentation accuracy. In contrast, this invention first recovers missing lesion information using correction directions obtained through meta-learning, and then filters these corrections using learned category prototypes, ultimately obtaining lesion annotations highly consistent with the true annotations, thus achieving more accurate segmentation results.

[0063] Table 1 compares the results of the LiTS and ISLES22 datasets under the single-instance annotation setting. The results are reported under two conditions: partial instance annotation and PIS annotation. The evaluation metrics include Dice (%), IoU (%), and NSD (%). Fully supervised results are also provided for reference. The best results are marked in bold.

[0064] Table 1 compares the results of the LiTS and ISLES22 datasets under the single-instance annotation setting.

[0065] As shown in Table 1, under partial instance annotation conditions only, the present invention (without graffiti annotation) outperforms all comparative methods based on partial instances and noise robustness, achieving a Dice improvement of +7.41% on LiTS and +5.68% on ISLES22, both significantly exceeding the baseline methods. In contrast, partial instance methods that rely on reconstruction pre-training or slice-level classification to learn background priors have limited effectiveness in cases of low-contrast lesions and blurred anatomical boundaries; while noise-resistant methods also perform poorly under this setting because a large number of unannotated lesions will form structured noise, fundamentally violating their methodological assumptions. With the introduction of sparse background graffiti, the present invention achieves the best overall performance: on the LiTS dataset, compared to the best graffiti-based method, it achieves a Dice improvement of +6.45% and an NSD improvement of +12.83%; on ISLES22, it still outperforms the second-best method USTM by +11.73% in Dice. Although graffiti-based methods can benefit from additional graffiti supervision signals, the pseudo-labels they generate vary greatly between different iterations, leading to instability in the training process and ultimately limited performance improvement. In contrast, this invention, by leveraging a meta-label correction mechanism, can progressively recall unlabeled lesion instances, thus consistently outperforming partial instance methods, graffiti methods, and noise-robust methods in segmentation performance, and exhibiting more stable training results. Furthermore, when using random background graffiti, the performance of this invention only shows a slight decrease: a 1.82% decrease in Dice on LiTS and a 0.99% decrease on ISLES22. This result demonstrates that the method is highly robust to incomplete or biased graffiti annotations.

[0066] Table 2 presents a comparison of results on the MS dataset. Results are reported under two conditions: partial instance annotation only and PIS annotation. Evaluation metrics include Dice (%), IoU (%), and NSD (%). Fully supervised results are also provided for reference. The best results are highlighted in bold.

[0067] Table 2 shows three different annotation completeness levels: 60%, 10%, and 5%. At all annotation completeness levels, the present invention outperforms all comparison methods. Compared to the 3D U-Net baseline, the present invention achieves improvements of +4.41%, +8.17%, and +11.23% in the Dice metric, respectively, while also gaining significant improvements in IoU and NSD. Notably, the 10% annotation setting outperforms the 60% setting. This is because 10% of the annotations correspond to the largest lesions in the top 10% of the volume, providing more representative training cues, while the 60% setting uses random sampling, potentially missing important large lesions. More importantly, the performance improvement of the present invention is more significant as the annotations become sparser, while other methods exhibit significant accuracy degradation or boundary instability.

[0068] Table 2 shows the comparison of results on the MS dataset.

[0069] This consistent performance under both extremely sparse and asymptotically sparse annotation conditions demonstrates that the present invention maintains high-quality lesion recall and accurate segmentation results even under extremely limited supervision. Although a performance gap still exists compared to fully supervised methods, the present invention reduces the performance gap of baseline methods on the MS dataset from 23.38%–14.58% to 12.15%–9.73%, and exhibits better stability. Furthermore, since lesions in the MS dataset are smaller and more dispersed, the implementation of fully supervised methods is inherently more challenging, further highlighting the value of the present invention in clinical applications.

[0070] On the other hand, the present invention also provides a weakly supervised lesion segmentation device for partial instance annotation, used to implement a meta-learning label correction and feature comparison learning framework based on a segmentation model; in each training iteration of the learning framework, the following modules are run sequentially: The meta-label correction module is configured to perform pseudo-label correction on a subset of voxels with the highest foreground probability in the unlabeled region through a selective meta-label correction mechanism. The reliability filtering module is configured to use a prototype-guided reliability filtering mechanism to perform semantic consistency verification of the correction results using category prototypes. The joint optimization module is configured to combine segmentation loss and prototype regularization loss to optimize segmentation model parameters. The prototype regularization loss includes contrast alignment loss and diversity loss to improve class discriminativeness and feature diversity.

[0071] It should be noted that the weakly supervised lesion segmentation device for partial instance annotation provided in this embodiment of the invention can execute the weakly supervised lesion segmentation method for partial instance annotation described in any of the above embodiments during specific operation, which will not be elaborated in this embodiment.

[0072] Figure 5 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 5 As shown, the electronic device may include a processor 510, a communications interface 520, a memory 530, and a communications bus 540. The processor 510, communications interface 520, and memory 530 communicate with each other via the communications bus 540. The processor 510 can call logical instructions from the memory 530 to execute a weakly supervised lesion segmentation method based on partial instance annotation.

[0073] Furthermore, the logical instructions in the aforementioned memory 530 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0074] On the other hand, the present invention also provides a computer program product, the computer program product including a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions, and when the program instructions are executed by a computer, the computer is able to execute the weakly supervised lesion segmentation method for partial instance annotation provided in the above embodiments. In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the weakly supervised lesion segmentation method for partial instance annotation provided in the above embodiments.

[0075] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A weakly supervised lesion segmentation method for partial instance annotation, characterized in that, This is used to construct a meta-learning label correction and feature contrast learning framework based on a segmentation model; in each training iteration of the learning framework, the following steps are executed sequentially: By using a selective meta-label correction mechanism, pseudo-label correction is performed on a subset of voxels with the highest foreground probability in unlabeled areas; A prototype-guided reliability filtering mechanism is used to verify the semantic consistency of the correction results using category prototypes. By using a joint optimization training mechanism, the segmentation model parameters are optimized by combining segmentation loss and prototype regularization loss. The prototype regularization loss includes contrast alignment loss and diversity loss to improve class discriminativeness and feature diversity.

2. The weakly supervised lesion segmentation method for partial instance annotation according to claim 1, characterized in that, A selective meta-label correction mechanism is used to perform pseudo-label correction on a subset of voxels with the highest foreground probability in unlabeled regions, including: Given a 3D input image and its foreground probability map, use a binary mask to exclude labeled voxels and select the K voxels with the highest probability from the unlabeled voxels as the correction candidate set. A learnable correction offset is assigned to each voxel in the correction candidate set, and a meta-gradient is computed through a meta-optimization step, which is obtained based on backpropagation of the validation set loss. The label update direction is determined based on the sign of the meta-gradient, a binary vector is constructed, and the pseudo-label is updated by combining the exponential moving average strategy.

3. The weakly supervised lesion segmentation method for partial instance annotation according to claim 2, characterized in that, The number K of candidate voxels is calculated using the ratio r.

4. The weakly supervised lesion segmentation method for partial instance annotation according to claim 1, characterized in that, A prototype-guided reliability filtering mechanism is used to perform semantic consistency verification of the correction results using category prototypes, including: On the output feature map of the last layer of the segmentation model decoder, M class prototypes are learned for each class, and the semantic alignment between the prototype representation and the feature embedding is optimized through the cross-attention module. For candidate voxels corrected by meta-learning, extract the feature vector of the candidate voxel after cross-attention optimization, and calculate the cosine similarity with the prototype of each category. Based on the principle of highest similarity, prototype category labels are assigned, and a binary filter mask is constructed to remove voxels whose correction direction is inconsistent with their respective prototype labels.

5. The weakly supervised lesion segmentation method for partial instance annotation according to claim 1, characterized in that, The segmentation loss in the joint optimization training mechanism is a combination of cross-entropy loss and Dice loss, and the weight of the prototype regularization loss is controlled by hyperparameters.

6. The weakly supervised lesion segmentation method for partial instance annotation according to claim 1, characterized in that, The segmentation model adopts a 3D U-Net architecture, in which the encoder is used to extract multi-scale features and the decoder is used to progressively restore the spatial resolution and output the segmentation results.

7. The weakly supervised lesion segmentation method for partial instance annotation according to claim 1, characterized in that, The method is applied to lesion segmentation in three-dimensional medical images.

8. A weakly supervised lesion segmentation device for partial instance annotation, characterized in that, This framework is used to implement a meta-learning label correction and feature contrast learning framework based on a segmentation model. In each training iteration of the learning framework, the following modules are run sequentially: The meta-label correction module is configured to perform pseudo-label correction on a subset of voxels with the highest foreground probability in the unlabeled region through a selective meta-label correction mechanism. The reliability filtering module is configured to use a prototype-guided reliability filtering mechanism to perform semantic consistency verification of the correction results using category prototypes. The joint optimization module is configured to combine segmentation loss and prototype regularization loss to optimize segmentation model parameters. The prototype regularization loss includes contrast alignment loss and diversity loss to improve class discriminativeness and feature diversity.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the weakly supervised lesion segmentation method for partial instance annotation as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the weakly supervised lesion segmentation method for partial instance annotation as described in any one of claims 1 to 7.

Images