An efficient 3D MRI segmentation method and system for knee joint based on iterative pseudo-label refinement

Abstract

This invention belongs to the field of knee joint MRI segmentation technology, specifically relating to an efficient knee joint 3D MRI segmentation method and system based on iterative pseudo-label refinement. The method includes: constructing a training set based on knee joint 3D MRI volume data; training an initial segmentation model using a pseudo-label self-refinement strategy based on the training set to obtain a refined iterative segmentation model; and implementing knee joint 3D MRI segmentation based on the refined iterative segmentation model. This invention first analyzes the relationship between different annotation ratios and segmentation performance, and based on this, proposes an iterative pseudo-label refinement strategy. By progressively optimizing the quality of pseudo-labels to reduce reliance on manually labeled data, a more efficient knee joint MRI segmentation method is explored.

Description

Technical Field

[0001] This invention belongs to the field of knee joint MRI segmentation technology, specifically relating to an efficient knee joint 3D MRI segmentation method and system based on iterative pseudo-label refinement. Background Technology

[0002] Osteoarthritis of the knee (OA) is a common degenerative joint disease characterized by progressive cartilage degeneration, osteophyte formation, and joint space narrowing. It is a leading cause of pain and functional impairment worldwide. Accurate mapping and quantitative assessment of knee joint structures are crucial for early diagnosis, progression monitoring, efficacy evaluation, and individualized treatment planning. Magnetic resonance imaging (MRI), with its excellent soft tissue contrast and ability to clearly visualize cartilage, bone, meniscus, and other intra-articular structures simultaneously, has become the primary imaging modality for assessing knee joint structures in clinical practice and research.

[0003] Despite the advantages of MRI, the analysis of knee MRI images still largely relies on manual segmentation of anatomical structures. This process is not only time-consuming but also highly subjective and dependent on the annotator's expertise. Especially when performing voxel-by-voxel annotation on 3D MRI data, each case typically requires several hours, making large-scale, high-precision annotation extremely difficult in clinical practice and large-scale studies. This high annotation cost significantly limits the application and widespread adoption of fully supervised deep learning models in knee MRI segmentation tasks.

[0004] In recent years, deep learning-based segmentation methods, especially fully convolutional networks such as U-Net and its extensions, have made significant progress in the field of medical image analysis. Numerous studies have shown that by introducing targeted network architecture design and multi-scale feature fusion strategies, excellent performance can be achieved in fully supervised knee MRI segmentation tasks. However, these methods typically rely on a large amount of high-quality voxel-level annotation, which is often difficult to obtain in 3D medical imaging tasks. Therefore, how to further improve segmentation performance under limited annotation resources has become a crucial problem that urgently needs to be solved in the field of medical image segmentation.

[0005] To reduce reliance on fully labeled data, weakly supervised and semi-supervised learning methods have received widespread attention in recent years. These methods aim to achieve segmentation performance close to fully supervised learning at a lower annotation cost by utilizing a small number of labeled samples combined with a large amount of unlabeled data. Representative strategies include pseudo-label self-training, consistency regularization, and teacher-student frameworks. Although these methods have achieved good results in natural images and some medical imaging tasks, they still face several challenges in 3D knee MRI segmentation tasks.

[0006] First, existing research largely focuses on algorithm model design, while quantitative analysis of the relationship between annotation budget and segmentation performance is relatively limited. Second, pseudo-label-based training strategies are prone to error accumulation during iteration, especially in 3D medical images of small or low-contrast anatomical structures, potentially leading to model training instability or decreased convergence performance. Third, although existing research has pointed out the prevalence of neighborhood shift in knee MRI data, systematic research on the generalization ability of weakly supervised models under different acquisition protocols and data distribution conditions remains lacking.

[0007] Voxel-level fine annotation in 3D knee MRI requires significant manual labor, which limits the application and widespread adoption of fully supervised segmentation methods on large-scale datasets. To improve annotation efficiency, this invention first analyzes the relationship between different annotation ratios and segmentation performance. Based on this analysis, an iterative pseudo-label refinement strategy is proposed. By progressively optimizing the quality of pseudo-labels, the reliance on manually annotated data is reduced, thereby exploring a more efficient knee MRI segmentation method. Summary of the Invention

[0008] To address the problems existing in the prior art, this invention provides an efficient 3D MRI segmentation method and system for the knee joint based on iterative pseudo-label refinement. It aims to improve the efficiency of annotation utilization, reduce the need for manually annotated data, and provide a practical solution for efficient MRI segmentation of the knee joint in research situations where it is difficult to obtain complete annotated data.

[0009] To achieve the above objectives, the present invention provides the following solution: An efficient 3D MRI segmentation method for the knee joint based on iterative pseudo-label refinement, the method comprising: A training set was constructed based on 3D MRI volumetric data of the knee joint; Based on the training set, a pseudo-label self-refinement strategy is used to train the initial segmentation model to obtain a refined and iterative segmentation model. A 3D MRI segmentation of the knee joint was achieved based on a refined iterative segmentation model.

[0010] Preferred methods for constructing a training set based on knee joint 3D MRI volumetric data include: The 3D MRI volumetric data of the knee joint were annotated according to different preset annotation ratios; A training set was constructed based on labeled 3D MRI volumetric data of the knee joint.

[0011] Preferably, the method for training the initial segmentation model using a pseudo-label self-refinement strategy based on the training set to obtain a refined and iterative segmentation model includes: The initial segmentation model is trained using a limited proportion of fully labeled voxel data in the training set to obtain the trained segmentation model. A pseudo-label self-refinement strategy is adopted to refine the trained segmentation model through multiple rounds of iterations, resulting in a refined segmentation model.

[0012] Preferably, the method of using a pseudo-label self-refinement strategy to perform multiple rounds of iterative refinement on the trained segmentation model to obtain a refined segmentation model includes: In the In the round of iterative refinement, the segmentation model obtained in the previous round is used for all training samples in the training set, and voxel-level pseudo-labels are generated through three-dimensional sliding window inference. Based on a screening mechanism that considers confidence and uncertainty, voxel-level pseudo-labels are screened to obtain pseudo-labels with masks. The masked pseudo-labels are merged with the real labels in the training set to form new training data, and the segmentation model is retrained until the model parameters converge.

[0013] Preferably, in the first In the iterative refinement round, the segmentation model obtained in the previous round is used for all training samples in the training set. The method for generating voxel-level pseudo-labels through three-dimensional sliding window inference includes: ; in, This is a voxel-level pseudo-label. For the first It's time for the segmentation model. This is three-dimensional MRI volumetric data.

[0014] Preferred, the first The optimization objective of the round-by-round iterative refinement is: ; in, This represents the combination of Dice loss and cross-entropy loss. For the real labels in the training set, For pseudo-tags with masks, For the first It's time for the segmentation model. and These are the weights of labeled samples and pseudo-labeled samples, respectively.

[0015] Preferably, the method further includes: A systematic generalization evaluation of the refined iterative segmentation model is performed, and a labeling budget curve is constructed to achieve a quantitative analysis of the relationship between labeling budget and segmentation performance.

[0016] The present invention also provides an efficient 3D MRI segmentation system for the knee joint based on iterative pseudo-label refinement. The system is used to implement the aforementioned method and includes: a construction module, a refinement module, and a segmentation module. The construction module is used to construct a training set based on 3D MRI volumetric data of the knee joint; The refining module is used to train the initial segmentation model based on the training set using a pseudo-label self-refining strategy to obtain a refined and iterative segmentation model. The segmentation module is used to perform 3D MRI segmentation of the knee joint based on a refined iterative segmentation model.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: We propose an efficient weakly supervised annotation framework for knee MRI segmentation tasks. By constructing an "annotation budget curve", we can characterize the quantitative relationship between annotation budget and segmentation performance.

[0018] An iterative pseudo-label self-refinement strategy was designed to improve the stability and convergence of model training under limited supervision. Experimental results show that the model performance exhibits a stable saturation trend under a 20% labeling rate, and reaches its optimal performance after two rounds of refinement.

[0019] Through multiple sets of experiments, including cross-dataset evaluation, the robustness of the proposed framework under different data distribution conditions was verified.

[0020] It provides a reproducible experimental scheme covering various annotation budgets and training configurations, which can provide a valuable experimental benchmark for future research on efficient medical image segmentation. Attached Figure Description

[0021] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of the overall structure of the efficient 3D knee joint MRI segmentation framework in an embodiment of the present invention. Figure 2 This is a schematic diagram of a three-dimensional knee joint MRI segmentation framework according to an embodiment of the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0025] Example 1 This invention provides an efficient 3D MRI segmentation method for the knee joint based on iterative pseudo-label refinement, comprising: An efficient 3D MRI segmentation method for the knee joint based on iterative pseudo-label refinement, the method comprising: A training set was constructed based on 3D MRI volumetric data of the knee joint; Based on the training set, a pseudo-label self-refinement strategy is used to train the initial segmentation model to obtain a refined and iterative segmentation model. A 3D MRI segmentation of the knee joint was achieved based on a refined iterative segmentation model.

[0026] like Figures 1-2 As shown, the specific implementation process of the present invention is as follows: This invention aims to achieve accurate 3D MRI segmentation of the knee joint under limited annotation budget conditions. To this end, this invention proposes a semi-supervised segmentation framework that organically combines limited fully supervised training, iterative pseudo-label refinement, and systematic generalization evaluation. Figure 2 An overall overview of the proposed framework is given, in which the complete training and evaluation process is organized into three sequential phases.

[0027] The first stage prepares the dataset and trains an initial seed model (i.e., the initial segmentation model) using a limited proportion of labeled voxel data (the smallest unit that constitutes a complete 3D medical image). The second stage generates full-volume pseudo-labels through sliding window inference and refines them through two rounds of iteration. During pseudo-label construction, a Gaussian weighted aggregation strategy is used to obtain voxel-level prediction results, while voxels with low confidence or high uncertainty are masked as ignored regions (confidence filtering) to avoid introducing unreliable supervision signals. Subsequently, the model undergoes progressive retraining by combining real annotations with confidence-filtered pseudo-labels. The third stage comprehensively evaluates the final model from multiple aspects, including intra-domain performance evaluation, annotation budget analysis, and cross-dataset generalization ability under zero-shot conditions.

[0028] In the first stage, a small subset of fully labeled training samples (referred to as "a small subset" here, meaning a portion of the fully labeled dataset selected from the entire dataset (e.g., approximately 20%)) are used to train the initial segmentation model, rather than using all the fully labeled data. The remaining unselected labeled data does not participate in model training at this stage. For example, approximately 20% of the available training data, i.e., 81 examples) are used to train the initial segmentation model. Training is then performed. The goal of this stage is not to pursue optimal segmentation performance, but to learn stable and anatomically consistent feature representations, enabling the model to serve as a reliable initialization for subsequent pseudo-label generation. Therefore, the seed model's ability to generate structurally coherent predictions with reasonable boundary localization is crucial to preventing error accumulation in subsequent iterations.

[0029] In the second stage, the initial segmentation model This approach is applied to unlabeled sample sets, generating voxel-level pseudo-labels through a 3D Gaussian fusion sliding window inference. Simultaneously, for samples with real labels, their labels serve as supervisory signals during model training, rather than participating in the pseudo-label generation process. Specifically, a size of [size missing] is used. Inference is performed using a sliding window with a 50% overlap rate, and a Gaussian weighted aggregation strategy is employed to mitigate potential boundary discontinuities. For each voxel, its final class label is determined by the maximum a posteriori probability.

[0030] To further improve the reliability of pseudo-supervision signals, this invention introduces a screening mechanism based on confidence and uncertainty at the voxel level. Voxels with low prediction confidence or high uncertainty are excluded from supervision by assigning an ignore label (value 255), thus preventing them from participating in loss calculations during training. This strategy allows model training to focus on high-confidence regions while avoiding reliance on complex heuristic post-processing methods such as morphological operations or conditional random fields. Subsequently, the segmentation model is retrained using the generated pseudo-labels with masks, resulting in an updated model M2. This process can be iteratively repeated, forming a progressive self-refinement mechanism: the model continuously utilizes its reliable predictions while gradually correcting potential mislabels.

[0031] In the third stage, the proposed framework is systematically evaluated in terms of both effectiveness and robustness. First, the segmentation performance is assessed on the OAI-ZIB test set, and a labeling budget curve is constructed to quantitatively analyze the relationship between the labeling ratio and model performance. Furthermore, zero-shot cross-dataset evaluation is conducted on the SKI10 dataset to examine the generalization ability of the proposed framework in the presence of domain bias.

[0032] Overall, the framework proposed in this invention provides a simple and effective solution for achieving highly efficient 3D knee MRI segmentation by jointly modeling the annotation budget, iterative pseudo-label refinement, and model generalization behavior in a unified training process.

[0033] Furthermore, in this embodiment, the method for constructing a training set based on knee joint 3D MRI volumetric data includes: The 3D MRI volumetric data of the knee joint were annotated according to different preset annotation ratios (such as 10%, 20%, 40%, and 100%). A training set was constructed based on labeled 3D MRI volumetric data of the knee joint.

[0034] Furthermore, in this embodiment, the method for training the initial segmentation model based on the training set using a pseudo-label self-refinement strategy to obtain a refined and iterative segmentation model includes: The initial segmentation model is trained using a limited proportion of fully labeled voxel data in the training set to obtain the trained segmentation model. A pseudo-label self-refinement strategy is adopted to refine the trained segmentation model through multiple rounds of iterations, resulting in a refined segmentation model. Specifically, this includes: In the In the round of iterative refinement, the segmentation model obtained in the previous round is used for all training samples in the training set, and voxel-level pseudo-labels are generated through three-dimensional sliding window inference. Based on a screening mechanism that considers confidence and uncertainty, voxel-level pseudo-labels are screened to obtain pseudo-labels with masks. The masked pseudo-labels are merged with the real labels in the training set to form new training data, and the segmentation model is retrained until the model parameters converge.

[0035] Specifically: To effectively utilize unlabeled data under limited supervision, this invention proposes an iterative pseudo-label self-refinement strategy based on self-training. For example... Figure 2 As shown, the core idea is to repeatedly use the model's predictions as supervision signals in subsequent training rounds, thereby gradually improving the quality of pseudo-labels and enhancing model performance. This process forms a lightweight self-refining loop, in which the model gradually corrects prediction errors generated in the early stages through consistency-driven learning.

[0036] In terms of format, let the training dataset be... ,in Represents three-dimensional MRI volume data (a complete three-dimensional medical image sample). Indicates the corresponding true label (if available, for unlabeled samples, y). i(No, some samples have corresponding real labels, while the rest do not contain real labels), from Different labeled subsets (10%, 20%, 40%, etc.) were used for the experimental setup. In the refinement round, the model obtained from the previous training round (denoted as...) This is applied to all training samples, and voxel-level class probability distributions are obtained through three-dimensional sliding window inference. The pseudo-label for each voxel is first determined by assigning the class with the highest posterior probability: ; in, This is a voxel-level pseudo-label. For the first It's time for the segmentation model. This is three-dimensional MRI volumetric data.

[0037] To further improve the reliability of pseudo-supervision signals, this invention introduces a screening mechanism based on confidence and uncertainty at the voxel level. Voxels with prediction confidence below a preset confidence threshold or uncertainty above a preset uncertainty threshold are excluded from supervision by being assigned an ignore label (value 255), thus preventing them from participating in loss calculation during training. The final result is a pseudo-label with a mask. This label effectively avoids interference from unreliable predictions while retaining supervision in high-confidence areas.

[0038] Unlike methods that rely on morphological operations or conditional random fields for post-processing, the proposed framework performs confidence-based masking operations directly at the voxel level. Since the seed model already provides anatomical structure predictions, the refining process is primarily driven by prediction consistency within reliable regions, rather than solely by high-confidence voxels. This design helps mitigate error accumulation in complex 3D anatomy while maintaining the simplicity of the training process.

[0039] Then, the generated pseudo-labels with masks will be... Combined with existing ground truth labels, the segmentation model is retrained to obtain an updated model. . No. The optimization objective of round refining is defined as: ; in, This represents a combination of Dice loss and cross-entropy loss, used to simultaneously optimize region overlap rate and boundary localization accuracy. For the real labels in the training set, For the first It's time for the segmentation model. and These are the weights of labeled samples (a subset of training samples with real labels) and pseudo-labeled samples (samples whose pseudo-labels are generated by model prediction from unlabeled samples).

[0040] By repeatedly generating pseudo-labels and retraining the model, the proposed self-refinement strategy can gradually suppress unstable prediction errors while enhancing the consistency of anatomical structure patterns. By combining full-volume inference with confidence-based supervised masking and avoiding additional heuristic post-processing steps, this method provides a simple and robust self-training paradigm suitable for labeling-efficient 3D knee MRI segmentation tasks.

[0041] like Figure 2 As shown, the specific iterative training process is as follows: To more clearly illustrate the overall framework of the method proposed in this invention and the dependencies between its various stages, Figure 2 A schematic diagram of a multi-round iterative training process based on a pseudo-label self-refinement strategy is presented. This process starts with an initial segmentation model and gradually optimizes the model performance through iterative processes of pseudo-label generation, confidence screening, and model retraining.

[0042] To make full use of limited labeled data, this invention employs a multi-round iterative training process based on the proposed pseudo-label self-refinement strategy. This framework starts with a seed model trained using a small amount of fully labeled data, and through repeated cycles of pseudo-label generation and model retraining, gradually corrects pseudo-label errors under weak supervision, achieving a stable improvement in model performance.

[0043] Suppose the training set contains Of the samples, only 1 The proportions are fully marked, the rest... The samples are unlabeled samples (under different labeling ratios (10%, 20%, 40%, 100%), the dataset is divided into labeled and unlabeled subsets according to the corresponding ratios. Unlabeled samples refer to the remaining samples that are not selected into the labeled subset under the current labeling ratio setting. Therefore, the unlabeled set changes dynamically with the labeling ratio. Suppose the training set contains N samples, of which a subset is randomly selected as fully labeled data according to a ratio r (e.g., 10%, 20%, 40%, etc.), and the remaining (1-r)N samples are considered unlabeled data under this experimental setting. Therefore, under different labeling ratio settings, the division between the labeled and unlabeled sets is dynamically adjusted with the change of r). First, in the labeled subset... The initial seed model is trained on a subset of samples with complete real labels selected from the training set, i.e., a fully labeled sample set used to train the initial seed model, providing reliable initialization for subsequent refinement rounds. During iterative training, both labeled and pseudo-labeled samples are optimized using the same Dice and cross-entropy combination loss. For pseudo-labeled data, voxel-level predictions are used as supervision signals in high-confidence regions, while low-confidence or high-uncertainty voxels are masked as ignored regions and excluded from the loss calculation.

[0044] In each round of refining In this process, the model obtained in the previous round is applied to all training samples, and a voxel-level class probability map is generated through 3D sliding window inference. Pseudo-labels are first determined by assigning each voxel to the class with the highest predicted probability. A Gaussian weighted window fusion strategy is employed during inference to reduce boundary artifacts and ensure spatial consistency between overlapping regions. Subsequently, voxel-level confidence and uncertainty metrics are used to mark low-confidence predictions as ignored regions before training, thereby improving the reliability of the supervision signal.

[0045] The generated pseudo-labels are then merged with the original labeled samples to form updated training data, which is used to retrain the segmentation model using a warm-start approach (initializing the training of the new model using existing model parameters or training results). To promote training stability and mitigate potential error amplification, this invention employs a coarse-to-fine training strategy, gradually reducing the learning rate and the number of training epochs in each refining epoch. This training scheme allows the model to first learn overall anatomical features, followed by gradual fine-grained adjustments in later stages.

[0046] The above process is repeated for several fixed refinement iterations until the model parameters are basically converged, at which point the performance improvement from further iterations tends to be limited. Overall, this iterative training process, through a continuous self-optimization mechanism, achieves effective utilization of unlabeled data, providing a simple and robust training framework for achieving highly efficient 3D knee MRI segmentation.

[0047] In summary, to improve annotation efficiency, this invention first analyzes the relationship between different annotation ratios and segmentation performance. Based on this, it proposes an iterative pseudo-label refinement strategy. By progressively optimizing the quality of pseudo-labels to reduce reliance on manually labeled data, a more efficient knee MRI segmentation method is explored. First, a fully supervised segmentation model is trained under different annotation ratios (10%, 20%, 40%, and 100%), and the changing trend between annotation budget and segmentation performance is analyzed. Based on this, a weakly supervised learning framework is constructed using a limited set of fully labeled voxel data (20%, 81 cases) as the initial training set. For unlabeled data, Gaussian fusion sliding window inference is used to generate pseudo-labels, and voxel-level confidence and uncertainty assessments are further performed, marking unreliable predicted regions as ignored regions. Subsequently, the model is iteratively optimized through multiple rounds of retraining, without introducing additional morphological post-processing steps throughout the process.

[0048] This invention addresses the issues of high annotation costs and unstable performance in weakly supervised knee MRI segmentation by proposing an efficient segmentation framework based on annotation budget curves and an iterative pseudo-label self-refinement strategy. Through systematic experiments and cross-dataset evaluation, the proposed method significantly reduces annotation costs while maintaining stable segmentation performance.

[0049] First, this invention constructs a systematic annotation budget evaluation protocol that covers a variety of training strategies, including fully supervised training, weakly supervised training, and iterative pseudo-label refinement, thereby quantitatively analyzing the relationship between annotation ratio and segmentation performance.

[0050] Secondly, this invention reveals the "quality versus quantity trade-off" phenomenon in weakly supervised medical image segmentation. Experimental results show that a small number of high-quality, fully labeled samples can provide a more stable and effective supervisory signal, significantly outperforming a large number of sparse or low-quality labels. This finding emphasizes the crucial role of label quality in pseudo-label learning: reliable seed labels can serve as stable supervisory anchors, effectively suppressing potential noise propagation during iterative training.

[0051] Third, the iterative refinement experiments demonstrate clear convergence characteristics. Segmentation performance continuously improves in the first two refinement rounds and stabilizes in subsequent rounds, indicating that two rounds of refinement are sufficient to achieve near-optimal performance for the OAI-ZIB knee MRI segmentation task. This result demonstrates that the proposed framework achieves a good balance between performance improvement and computational overhead without requiring complex teacher-student structures or additional supervision mechanisms.

[0052] Cross-dataset experiments on the SKI10 dataset further demonstrate that domain bias still leads to performance degradation due to differences in MRI sequences, anatomical structure distributions, and annotation protocols. However, the weakly supervised model performs essentially the same as the corresponding fully supervised model in cross-domain tests, indicating that the proposed iterative refinement strategy does not introduce additional generalization instability. Therefore, the observed performance differences mainly stem from the inherent domain differences between datasets, rather than from the weakly supervised training framework itself.

[0053] In summary, this invention re-examines the problem of weakly supervised knee MRI segmentation from the perspective of annotation efficiency and provides a basis for balancing annotation cost and segmentation performance under limited annotation resources. The results show that reliable 3D knee MRI segmentation can be achieved with a lower annotation budget by combining a small number of high-quality annotations with a stable pseudo-label refinement mechanism. Future work will further explore uncertainty-aware pseudo-label strategies and domain adaptation methods, and extend this framework to more anatomical structures and multimodal medical imaging tasks to further enhance its robustness and potential clinical application value.

[0054] Example 2 This embodiment uses specific experimental data to illustrate the specific implementation process and effects of the method described in the foregoing embodiments: 1. Dataset and Experiment Setup The experiment was conducted using the publicly available OAI-ZIB knee MRI dataset, derived from the Osteoarthritis Initiative (OAI) cohort. The dataset contains 507 three-dimensional dual-echo steady-state (DESS) MRI volumes, with pixel-level annotations provided by experienced radiologists. The segmentation task involved five anatomical structures: femur (FB), femoral cartilage (FC), tibia (TB), medial tibial cartilage (MTC), and lateral tibial cartilage (LTC).

[0055] To ensure the reproducibility and fairness of the experiments, this invention employs a fixed, patient-independent data partitioning scheme. Specifically, the dataset is divided into 404 training samples, 52 validation samples, and 51 test samples. All experiments, including ablation studies, annotation budget analysis, and iterative refinement processes, are conducted based on the same data partitioning. Different annotation budget settings are simulated by adjusting only the proportion of labeled samples in the training set (while keeping the validation and test sets unchanged).

[0056] The spatial dimensions of the original MRI volume data are approximately voxels, voxel spacing is The skeletal structure is relatively large and has a clear outline, while cartilage tissue accounts for only about 0.5%–1% of the total volume and often exhibits low contrast and blurred boundaries, which poses a significant challenge to accurate segmentation.

[0057] To further evaluate the model's cross-dataset generalization ability, this invention underwent external testing on the SKI10 challenge dataset. This dataset, acquired using different scanners and imaging protocols, provides manually labeled femoral and tibial cartilage and is commonly used to evaluate the model's robustness in the presence of domain offset. The model trained on the OAI-ZIB dataset was directly evaluated on the SKI10 dataset without any retraining or fine-tuning.

[0058] All MRI volume data were standardized and cached before training. The preprocessing workflow included: retargeting images to the RAS coordinate system; resampling images using third-order spline interpolation; and uniformly resampling labels to the RAS coordinate system. The voxel spacing was then determined; subsequently, z-score intensity normalization was performed. The knee joint region was automatically cropped and filled to a fixed size. Voxels are used to preserve as much anatomical information as possible while reducing computational costs. The training and validation phases use cached volumetric data, while inference and evaluation are performed in the original spatial domain.

[0059] The segmentation network uses 3D BasicUNet as its backbone structure, with 1 input channel and 6 output channels (including background and five types of anatomical structures). The network feature channels are set to... The model has approximately 1.2 million parameters. Training was performed on a single NVIDIA RTX 4060 GPU (8 GB VRAM). The optimizer used was AdamW, with parameters set to... , The initial learning rate is The batch size was set to 2. Both labeled and pseudo-labeled samples were optimized using a combined loss function of Dice and cross-entropy.

[0060] In the pseudo-label generation stage, a sliding window inference strategy is adopted, with a window size of [missing information]. The overlap ratio is 0.5, and boundary artifacts are reduced through Gaussian weighted aggregation. Voxel-level prediction results are obtained through the argmax operation of the softmax probability distribution. To improve the reliability of pseudo-labels, confidence and uncertainty thresholds are introduced at the voxel level. Voxels with low confidence or high uncertainty are marked as ignore labels (value 255), so they are not included in the loss calculation during training. This strategy ensures full-volume inference coverage while confining supervision to a high-reliability region.

[0061] Model performance was evaluated using the Dice similarity coefficient (DSC), the 95th percentile Hausdorff distance (HD95), and the intersection-over-union ratio (IoU). All evaluation metrics were calculated in a uniform evaluation space to ensure consistency of results. Experiments were implemented using PyTorch 1.10.0 and the MONAI framework, with a fixed random seed set in all experiments to ensure reproducibility of results.

[0062] 2. Budget Analysis To systematically analyze the relationship between annotation budget and segmentation performance, this invention conducted experiments under different supervised settings and plotted annotation budget-performance curves. This analysis aims to explore the changing trend of segmentation accuracy as the proportion of labeled data increases, and to evaluate the effectiveness of the proposed iterative pseudo-label refinement strategy in alleviating insufficient annotation.

[0063] Suppose the training set contains A sample, of which the proportion is The samples are fully labeled. Within a given labeling budget... In this case, labeled subsets and unlabeled subsets are defined as follows: First, fully supervised learning reference models are constructed, which utilize only a subset of fully labeled volume data. Training was performed without introducing any pseudo-labels. Experiments were conducted at label ratios of 10%, 20%, 40%, and 100% to characterize the upper limit trend of model performance as labeled data gradually increases.

[0064] We further construct a weakly supervised baseline method that combines sparse doodle annotations with single-round pseudo-label training. In this setting, only a subset of the annotations is provided with partial voxel-level annotations, in the form of 3D doodle annotations: approximately 20% of the voxels are marked with sparse lines or dots, while the remaining voxels are considered unlabeled (labels ignored). The remaining training samples do not contain any manual annotations and rely solely on pseudo-labels for supervision. Furthermore, the pseudo-labels are generated only once in the initial training phase and are not used in subsequent iterations. Despite utilizing additional unlabeled data, the model's performance is still significantly lower than the corresponding baseline. This phenomenon indicates that when the initial supervision signal is sparse and incomplete, the pseudo-labels generated in one round can easily introduce significant noise, thus hindering effective model convergence during training and ultimately leading to a significant performance bottleneck.

[0065] In contrast, the iterative pseudo-label refinement strategy (Round-2) proposed in this invention initializes with 20% densely labeled seed data, rather than sparse supervision. Under the same labeling budget, this strategy can significantly improve segmentation performance: when the labeling ratio is 20%, it approaches the performance level achieved by fully supervised training.

[0066] Subsequently, this invention utilizes 20% of the fully labeled data as seed supervision to evaluate the proposed iterative pseudo-label refining framework. In this setting, initial pseudo-labels are first generated for all training samples, and then iteratively refined in subsequent training rounds.

[0067] This model reflects that as the annotation budget increases, the marginal benefit of performance improvement gradually decreases. This phenomenon indicates that introducing an iterative pseudo-label refinement mechanism can significantly improve annotation utilization efficiency.

[0068] Overall, the annotation budget analysis results show that although the fully supervised model can make full use of the labeled data, the proposed iterative pseudo-label refining framework significantly narrows the performance gap with the fully supervised method under low annotation budget conditions, and can still achieve segmentation performance close to that of the fully supervised model while significantly reducing the annotation cost.

[0069] 3 Evaluation Indicators This invention uses three commonly used metrics to quantitatively evaluate segmentation performance: Dice similarity coefficient (DSC), intersection-over-union ratio (IoU), and 95th percentile Hausdorff distance (HD95).

[0070] The Dice similarity coefficient measures the degree of volume overlap between the predicted segment and the ground truth. It is defined as follows: in and These represent the predicted voxel set and the ground truth labeled voxel set, respectively. A Dice score is calculated for each anatomical structure, and the average score is obtained by averaging across all categories.

[0071] Intersection over Union (IoU) is used to evaluate the ratio of the intersection to the union between the predicted and actual results. Compared to the Dice metric, IoU provides a more stringent assessment of overlapping areas. This invention also calculates IoU for each anatomical structure separately and reports the average across all categories.

[0072] The HD95 metric assesses the geometric accuracy of the segmentation boundary by calculating the 95th percentile of the symmetrical Hausdorff distance between the predicted and actual surfaces. in Indicates from set From each point to the set The set of shortest distances on the surface. Compared to the maximum Hausdorff distance, HD95 is less sensitive to extreme outliers while still effectively reflecting geometric deviations along anatomical boundaries. All HD95 values ​​are reported in millimeters (mm).

[0073] All evaluation metrics are calculated in a unified cache space to ensure consistency of evaluation results across different experimental settings.

[0074] 4 Training Program All experiments followed a unified training scheme to ensure fair comparisons under different labeling budgets and refining strategies. For both labeled and pseudo-labeled samples, a combination of Dice loss and cross-entropy loss was used to optimize the segmentation network, with the training objective consistent with the loss function defined in Equation (1).

[0075] In the initial supervised training phase (i.e., round 1), the model is trained using 20% ​​of the fully labeled data, with a learning rate set to [value missing]. The training ran for 100 epochs. In subsequent epochs, a coarse-to-fine optimization strategy was adopted, gradually reducing the learning rate and the number of training epochs. Specifically, the second and third epochs used different learning rates. and The training rounds consisted of 50 and 30 epochs, respectively.

[0076] In each iteration, the model parameters are initialized by the checkpoint with the best performance in the previous round, thus forming a warm-start training mechanism to promote stable convergence of the training process and gradually correct potential errors in pseudo-labels.

[0077] No data augmentation was used during training. Preliminary experiments show that common spatial augmentation methods (such as random rotation and scaling) can disrupt the anatomical alignment of the cache data, leading to instability in the pseudo-label refining process. Therefore, this invention preserves the original spatial configuration to ensure the consistency of training data across different refining rounds.

[0078] 5 Discussions 5.1 Optimal annotation budget To analyze the impact of annotation scale on segmentation performance, this invention first compared fully supervised models trained with different annotation ratios (10%, 20%, 40%, and 100%), and based on this, constructed a trend showing the relationship between annotation budget and performance. Experimental results show a clear non-linear relationship between annotation ratio and segmentation accuracy. Performance improvement is more significant in the low budget range (10%–40%). In contrast, performance improvement slows down significantly when the annotation ratio increases from 40% to 100%. This phenomenon reflects the typical diminishing marginal returns law in supervised learning: once the model has obtained sufficient supervised information, the performance gain from further increasing labeled data will gradually decrease.

[0079] Based on this, by introducing an iterative pseudo-label refinement strategy, iterative pseudo-label refinement can effectively utilize unlabeled data and continuously improve labeling efficiency during training.

[0080] From a cost-benefit perspective, when annotation resources are limited, a 20% annotation ratio can be considered a practically meaningful operational point. At this budget level, segmentation performance is close to saturation, and the required manual annotation cost is only one-fifth of that of fully supervised annotation. Although the optimal annotation ratio may vary depending on the dataset size and imaging protocol, the results of this invention demonstrate that in resource-constrained medical imaging scenarios, combining a small number of high-quality annotations with an iterative pseudo-label refinement strategy can achieve a good balance between annotation cost and segmentation performance.

[0081] 5.2 Quality and Quantity Trade-offs in Weak Supervision In weakly supervised segmentation tasks, the number and quality of labeled samples are not independent; increasing the number of labeled samples does not necessarily lead to performance improvements. A typical phenomenon observed in this invention comes from the comparison between the Weak-40% and Dense-20% settings. Although Weak-40% incorporates more labeled samples and introduces pseudo-label supervision, its average Dice score is still significantly lower than that of the Dense-20% model. The latter, while using fewer labeled samples, ensures that all samples are fully labeled. This result reveals a key trade-off in weakly supervised learning: a large number of low-quality supervision signals are often less effective than a small number of high-quality labels.

[0082] This phenomenon is primarily related to the reliability of the supervisory signal during pseudo-label learning. In sparsely labeled scenarios, supervisory information only covers local areas, making it difficult to provide sufficient global structural constraints for the model in the early training stages. Therefore, the model is more prone to generating structurally inaccurate pseudo-labels, and these errors can propagate and accumulate during subsequent training. In contrast, even in limited quantities, fully labeled samples can still provide complete and consistent supervisory information, enabling the model to learn more stable feature representations and providing reliable initialization for subsequent pseudo-label refinement.

[0083] From an information representation perspective, high-density labeled samples provide more complete and structurally continuous supervisory signals, which helps the model better learn the mapping relationship between image features and anatomical structures. A similar trend has been discussed in existing weakly supervised segmentation research, where a small number of high-confidence labels can significantly improve model performance through an effective pseudo-label propagation mechanism. The experimental results of this invention are consistent with this viewpoint.

[0084] In sparse, weakly supervised settings, segmentation results often exhibit problems such as discontinuous boundaries, fragmentation of cartilage regions, and structural discontinuities. In contrast, the proposed iterative pseudo-label refining framework significantly improves the consistency and smoothness of anatomical structures, especially demonstrating greater stability in thin-layer cartilage regions. This further validates the importance of high-quality initial supervision for pseudo-label learning.

[0085] In summary, the results of this invention demonstrate that the effectiveness of pseudo-label learning depends not only on the scale of available supervised data but also closely on the quality of the initial annotations. High-quality annotations can serve as stable supervisory anchors, thereby suppressing the amplification of noise during iterative training. From a practical application perspective, when annotation resources are limited, prioritizing annotation quality is often more helpful in achieving reliable segmentation performance than simply increasing the number of annotations.

[0086] 5.3 Stability and convergence of iterative refining To evaluate the stability and convergence behavior of the proposed pseudo-label refining strategy, three rounds of continuous refining experiments were conducted with a 20% labeling budget. Experimental results show that the average Dice score improved from the first to the second round, then slightly decreased in the third round. This trend indicates that most performance improvements occur in the early stages, while the improvement in subsequent rounds gradually decreases, and the overall performance tends to stabilize.

[0087] This phenomenon of "rapid improvement in the early stages and gradual saturation in the later stages" is consistent with the typical behavior of iterative pseudo-label learning. In the initial refinement stage, the progressively improved pseudo-labels can effectively correct early prediction errors and strengthen the consistency of anatomical structures, thus bringing significant performance improvements. As the refinement process progresses, the room for improvement in pseudo-label quality gradually shrinks, while the impact of residual prediction uncertainty relatively increases. Therefore, the model performance may fluctuate slightly and no longer continue to improve.

[0088] In this invention, the model is considered to have reached the empirical convergence point when the performance change between successive refinement epochs becomes negligible. Further increasing the number of refinement epochs after this point typically yields little benefit. From an optimization perspective, the "coarse-to-fine" training strategy employed in this invention aligns with this convergence characteristic: in early epochs, a larger learning rate and more training epochs enable the model to learn global anatomical patterns from pseudo-labels; while in later epochs, a smaller learning rate and fewer training epochs allow for fine-grained adjustment of model parameters, while simultaneously reducing the risk of amplified pseudo-label noise.

[0089] Overall, these experimental results demonstrate that, under the settings of this invention, a finite number of refinement iterations is sufficient to achieve stable and near-optimal segmentation performance. Although specific convergence behavior may vary depending on dataset size and task type, for the OAI-ZIB knee MRI segmentation task, two to three rounds of refinement iterations provide a reasonable trade-off between performance improvement, training stability, and computational efficiency.

[0090] 5.4 Practical Significance and Limitations The weakly supervised knee MRI segmentation framework proposed in this invention, based on annotation budget analysis and iterative pseudo-label refinement, has significant practical implications in scenarios with limited annotation resources. In this invention, segmentation performance approaching that of a fully supervised model can be achieved using only 20% of the fully labeled data, significantly reducing the workload of voxel-level manual annotation while maintaining high accuracy.

[0091] This highly efficient annotation training strategy is particularly important for clinical research institutions and small to medium-sized medical centers, as these institutions typically possess large amounts of image data but lack sufficient experts for fine-grained annotation. By combining a small amount of high-quality annotation with automatic pseudo-label refinement, the proposed method achieves stable 3D segmentation of bone tissue and cartilage structures. This capability has potential value for applications such as early assessment of osteoarthritis, quantitative analysis of cartilage morphology, and large-scale population imaging studies.

[0092] Beyond specific segmentation tasks, the annotation budget curve proposed in this invention also provides a data-driven decision-making reference. Researchers can empirically determine an appropriate annotation ratio (e.g., 20%–40%) based on their data scale and annotation resources, and achieve near-fully supervised learning performance with a few iterative refinements. Notably, this framework does not require complex teacher-student models or multi-stage network designs, but instead employs a simple and reproducible training process, thus exhibiting good deployability in resource-constrained environments.

[0093] Overall, this invention verifies the feasibility of achieving high-precision knee joint MRI segmentation under limited annotation conditions, and shows that annotation efficiency can be significantly improved by combining a small number of high-quality annotations with a stable pseudo-label refining strategy.

[0094] Example 3 Based on the same inventive concept, the present invention also provides an efficient knee joint 3D MRI segmentation system based on iterative pseudo-label refinement, for implementing the method described in the foregoing embodiments. The system includes: a construction module, a refinement module, and a segmentation module. The construction module is used to construct a training set based on 3D MRI volumetric data of the knee joint; The refining module is used to train the initial segmentation model based on the training set using a pseudo-label self-refining strategy to obtain a refined and iterative segmentation model. The segmentation module is used to perform 3D MRI segmentation of the knee joint based on a refined iterative segmentation model.

[0095] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made by those skilled in the art to the technical solutions of the present invention without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A highly efficient 3D MRI segmentation method for the knee joint based on iterative pseudo-label refinement, characterized in that, The method includes: A training set was constructed based on 3D MRI volumetric data of the knee joint; Based on the training set, a pseudo-label self-refinement strategy is used to train the initial segmentation model to obtain a refined and iterative segmentation model. A 3D MRI segmentation of the knee joint was achieved based on a refined iterative segmentation model.

2. The method according to claim 1, characterized in that, Methods for constructing training sets based on knee joint 3D MRI volume data include: The 3D MRI volumetric data of the knee joint were annotated according to different preset annotation ratios; A training set was constructed based on labeled 3D MRI volumetric data of the knee joint.

3. The method according to claim 1, characterized in that, Based on the training set, the method for training the initial segmentation model using a pseudo-label self-refinement strategy to obtain a refined and iterative segmentation model includes: The initial segmentation model is trained using a limited proportion of fully labeled voxel data in the training set to obtain the trained segmentation model. A pseudo-label self-refinement strategy is adopted to refine the trained segmentation model through multiple rounds of iterations, resulting in a refined segmentation model.

4. The method according to claim 3, characterized in that, Methods for obtaining a refined segmentation model by employing a pseudo-label self-refinement strategy and performing multiple rounds of iterative refinement on the trained segmentation model include: In the In the round of iterative refinement, the segmentation model obtained in the previous round is used for all training samples in the training set, and voxel-level pseudo-labels are generated through three-dimensional sliding window inference. Based on a screening mechanism that considers confidence and uncertainty, voxel-level pseudo-labels are screened to obtain pseudo-labels with masks. The masked pseudo-labels are merged with the real labels in the training set to form new training data, and the segmentation model is retrained until the model parameters converge.

5. The method according to claim 4, characterized in that, In the In the iterative refinement round, the segmentation model obtained in the previous round is used for all training samples in the training set. The method for generating voxel-level pseudo-labels through three-dimensional sliding window inference includes: ； in, This is a voxel-level pseudo-label. For the first It's time for the segmentation model. This is three-dimensional MRI volumetric data.

6. The method according to claim 5, characterized in that, No. The optimization objective of the round-by-round iterative refinement is: ； in, This represents the combination of Dice loss and cross-entropy loss. For the real labels in the training set, For pseudo-tags with masks, For the first It's time for the segmentation model. and These are the weights of labeled samples and pseudo-labeled samples, respectively.

7. The method according to claim 1, characterized in that, The method further includes: A systematic generalization evaluation of the refined iterative segmentation model is performed, and a labeling budget curve is constructed to achieve a quantitative analysis of the relationship between labeling budget and segmentation performance.

8. A highly efficient 3D MRI segmentation system for the knee joint based on iterative pseudo-label refinement, the system being used to implement the method described in any one of claims 1-7, characterized in that, The system includes: a construction module, a refining module, and a segmentation module; The construction module is used to construct a training set based on 3D MRI volumetric data of the knee joint; The refining module is used to train the initial segmentation model based on the training set using a pseudo-label self-refining strategy to obtain a refined and iterative segmentation model. The segmentation module is used to perform 3D MRI segmentation of the knee joint based on a refined iterative segmentation model.

Images