Uncertainty-Guided Dynamic Pseudo-Label Enhancement for Semi-Supervised Left Atrial Segmentation

1. Introduction

Rapid advances in medical technology and enhanced quality of life have raised public health awareness. However, cardiovascular disease remains a leading health threat. The AHA 2025 Statistical Report reports a global age-adjusted mortality rate of 235.18 per 100,000 and a total prevalence of 612 million cases, underscoring its substantial burden worldwide. Atrial fibrillation (AF), a common cardiac arrhythmia originating in the atria, is a major component. Accurate segmentation of left atrial images therefore holds significant clinical and societal value, facilitating precise diagnosis, optimal treatment planning, and improved prognostic evaluation for AF patients.

With deep learning is developing, full supervised learning achieved extremely remarkable results in tasks of medical image segmentation [1,2,3]. However, such methods often rely on high-quality and large-scale labelled data, which brings two major challenges to left atrial image annotation.

First, the cost of medical image annotation is so high. Traditional manual annotation methods are extremely inefficient for annotating medical images. An experienced medical staff needs to invest lots of time in labelling. Secondly, the quality of image annotation results varies greatly. Because the collection of medical images is influenced by noise easily, and image edges are blurred, contrast is low, this will lead to significant differences in labelling results among clinical imaging physicians with different experiences [4].

To overcome annotation and generalization challenges, semi-supervised learning (SSL) has been widely adopted for medical image segmentation, combining labelled and unlabelled data to train effectively with limited annotations. The pseudo-labelling approach [5] is key, yet existing methods face three shortcomings: low-confidence predictions near ambiguous boundaries cause progressive noise accumulation; a fixed confidence threshold fails to adapt to training dynamics, potentially impairing optimization [6]; and random augmentations lack purposeful design. Uncertainty-based augmentation [7] has proven effective in guiding augmentations to mitigate these issues.

Given the above research background, we propose a semi-supervised medical image segmentation method based on uncertainty-guided dynamic pseudo-label augmentation, it’s useful for left atrial image segmentation. In this approach, the confidence threshold is dynamically optimized to better accommodate the evolving characteristics of data distribution and feature variations. An Uncertainty-Guided Region Mix (UMIX) module is further introduced to enable targeted data augmentation. Built upon the UniMatch backbone, we incorporate an additional unlabelled data stream and apply both image-level and feature-level perturbations to it, thereby strengthening the supervisory signals and improving the reliability of pseudo labels as well as the model’s generalization capability.

2. Materials and Methods

Semi-supervised learning partitions data into labelled and unlabelled subsets, using pseudo-labels from the model trained on labelled data to improve performance. Since no ground truth is available, pseudo-label quality is critical [8]. Accordingly, this paper focuses on refining the generation and filtering modules of pseudo-labels.

Current confidence thresholds are fixed or heuristic-based, causing early discarding of useful signals and later noise retention, which undermines accuracy and robustness [6]. Such global thresholds cannot adapt to dynamic data distribution changes. Su et al. proposed a novel adaptive loss function to mitigate the impact of large gradients on the model during the early training stage [9]. Medical images have ambiguous boundaries, making pseudo-labels prone to noise accumulation, so directed guidance is needed for low-confidence regions. Inspired by UniMatch, we simultaneously apply image-level and feature-level weak-strong perturbations to enhance performance [10].

The left atrium exhibits scale-dependent morphological characteristics, requiring simultaneous capture of global positional information and local boundary details. Su et al. proposed a proposal-free fully convolutional network, PF-FCN, which incorporates an innovative “box map” mechanism (comprising a bounding box size map, a correlation score map, and a classification score map) along with a channel-spatial context subnetwork. By reducing fully connected layers and filtering center points, PF-FCN significantly decreases the parameter count [11].

It should be noted that GenAI was used in this paper to assist with literature reading, grammar checking, and translation.

2.1. Overall Network Framework

The proposed method is built upon a dynamic pseudo-label augmentation strategy, with the overall architecture depicted in Figure 1. It comprises the following core modules (1) a dynamic pseudo-label refinement module, which jointly optimizes pseudo-label quality through adaptive threshold selection and UMIX-based local enhancement, (2) a dual-task multi-perturbation module, which, building upon the UniMatch model, simultaneously applies both image-level and feature-level weak and strong perturbations to one of the unlabelled data streams. The framework of the UniMatch model is illustrated in Figure 2.

2.2. Algorithmic Description of Each Module

In the dynamic pseudo-label refinement module, our method performs dynamic confidence threshold optimization, which is realized by designing a threshold update function based on exponential decay to balance the quality requirements for pseudo-labels across different training stages.

The τ(t) is a dynamic threshold that varies over time, the initial confidence threshold τ_base = 0.85. The initial confidence threshold τ_base and τ gradually increases with the number of iterations t:

τ(t)=τ_base+(1-τ_base)⋅sigmoid_rampup(t,T_max)

(1)

Here,T_max represents the maximum number of iterations, and the ramp function is defined as:

$$\mathrm{sigmoid}\text{\_}\mathrm{rampup}\left(\mathrm{t},\mathrm{T}\right)={\mathrm{e}}^{-5\text{⋅}{\left(1-\mathrm{t}/\mathrm{T}\right)}^2}$$

(2)

This function slowly raises the threshold in the early stage of training to retain more samples, and then quickly approaches 1 in the later stage to prevent noise from contaminating the correct supervision signal. Model-generated prediction probability for unlabelled data Pu and screens out high-confidence regions:

$${\stackrel{\mathrm{ˆ}}{\mathrm{y}}}_{\mathrm{u}}=\mathrm{I}\left(\max \left({\mathrm{p}}_{\mathrm{u}}\right)>\mathsf{\tau }\left(\mathrm{t}\right)\right)\cdot \arg \max \left({\mathrm{p}}_{\mathrm{u}}\right)$$

(3)

Here, I(⋅) is the indicator function, and only predictions with a confidence level higher than the threshold are retained, the arg max operator selects the class with the highest predicted probability.

2.3. UMIX Local Uncertainty Enhancement

In medical images, organ boundaries (e.g., the heart) and lesion margins are examples of regions where predictive uncertainty tends to be high. UMIX mitigates this by replacing features in such high-uncertainty regions, thereby strengthening the model’s robustness to complex boundaries. First, calculate the voxel-by-voxel entropy of the model prediction result.

$$\mathrm{H}\left({\mathrm{p}}_{\mathrm{u}}\right)=-{\displaystyle \sum _{\mathrm{c}=1}^{\mathrm{C}}}{\mathrm{p}}_{\mathrm{u},\mathrm{c}}{\log \mathrm{p}}_{\mathrm{u},\mathrm{c}}$$

(4)

Divide the feature map into local blocks of K × K × K (default is 16 × 16 × 16), and calculate the average entropy within the blocks:

$${\overline{\mathrm{H}}}_{\mathrm{k}}={\displaystyle \frac{1}{{\mathrm{K}}^3}}\sum _{\left(\mathrm{i},\mathrm{j},\mathrm{l}\right)\in {\mathrm{Block}}_{\mathrm{k}}}\mathrm{H}{\left({\mathrm{p}}_{\mathrm{u}}\right)}_{\mathrm{i},\mathrm{j},\mathrm{l}}$$

(5)

Select the k block with the highest entropy value (default k = 2) as the area to be replaced. Then extract the low-entropy block features at the corresponding positions from another unlabelled sample and perform the replacement:

$${\mathrm{x}}_{\mathrm{mix}}^{\left(\mathrm{b}\right)}\left[{\mathrm{Block}}_{\mathrm{k}}\right]={\mathrm{x}}_{\mathrm{src}}^{\left(\mathrm{b’}\right)}\left[{\mathrm{Block}}_{\mathrm{k}}\right]$$

(6)

The variable ${\mathrm{x}}_{\mathrm{mix}}^{\left(\mathrm{b}\right)}$ denotes the target sample features in the current batch after UMIX-based mixing augmentation. The notation [Block_k] serves as an index indicating the position of the k-th local block, which is determined by the high-entropy selection procedure as described in Equation 5.

The algorithm of this module is as follows:

Algorithm 1: UMIX

Input: Feature map, number of swap blocks k = 2, patch size K = 16. Output: The feature map x after exchange. 1. Begin 2. Calculate the voxel-by-voxel entropy for the model prediction result. 3. Divide the feature map into local blocks of K×K×K, where default is 16×16×16, and calculate the average entropy within the blocks. 4. Select the k block with the highest entropy value, where default k = 2, as the area to be replaced. 5. Then extract the low-entropy block features at the corresponding positions from another unlabelled sample and perform the replacement. 6. End

2.4. UMIX Local Uncertainty Enhancement

The Cross-Entropy Loss function [12] and the Dice loss function [2] are both commonly used loss functions in image segmentation tasks. The cross-entropy loss function is shown in Formula 7, and the Dice loss function is shown in Formula 8. Here, ε = 1e - 8 is the smoothing term.

CE(p,y) = −(plogy + (1 − p)log(1 − y))

(7)

$$\mathrm{Dice}\left(\mathrm{p},\mathrm{y}\right)=1-{\displaystyle \frac{2\sum \mathrm{p}\cdot \mathrm{y}+\mathsf{\epsilon }}{\sum \mathrm{p}+\sum \mathrm{y}+\mathsf{\epsilon }}}$$

(8)

Supervised Loss effectively alleviates the common class imbalance problem in medical images by jointly optimizing the classification boundaries (using CE) and the region overlap (using Dice). For the labeled data D_l = {(x_i, y_i)}, the supervised loss combines the cross-entropy loss and the Dice loss to simultaneously optimize segmentation accuracy and region overlap with the ground truth. It is shown in Formula 9.

$${\mathrm{L}}_{\mathrm{x}}={\displaystyle \frac{1}{2}}\left(\mathrm{CE}\left({\mathrm{p}}_{\mathrm{x}},{\mathrm{y}}_{\mathrm{x}}\right)+\mathrm{Dice}\left({\mathrm{p}}_{\mathrm{x}}^{\left(1\right)},{\mathrm{y}}_{\mathrm{x}}\right)\right)$$

(9)

Here, p_x = Softmax(f_θ(x)) denotes the model’s predicted probability distribution, which is typically normalized via a Softmax function, with each dimension corresponding to the probability of a respective class. y_x represents the ground-truth label corresponding to the input sample x. Moreover, ${\mathrm{p}}_{\mathrm{x}}^{\left(1\right)}$ refers to the probability map of the foreground class, i.e., the response intensity of the target region.

Unsupervised Loss enforces consistency across two augmentation branches, thereby enhancing the model’s robustness to data perturbations. For the unlabelled data D_u = {x_j}, we define the unsupervised loss based on the pseudo-labels ${\widehat{\mathrm{y}}}_{\mathrm{u}}$ generated by the model and the dynamic confidence threshold τ(t), as formulated in Equation 10.

$${\mathrm{L}}_{\mathrm{u}}={\displaystyle \sum _{\mathrm{k}=1}^2}\mathrm{Dice}\left({\mathrm{p}}_{{\mathrm{u}}_{\mathrm{k}}}^{\left(1\right)},{\stackrel{\mathrm{ˆ}}{\mathrm{y}}}_{\mathrm{u}}\cdot \mathrm{I}\left(\max \left({\mathrm{p}}_{{\mathrm{u}}_{\mathrm{k}}}\right)>\mathsf{\tau }\left(\mathrm{t}\right)\right)\right)$$

(10)

Among them, ${\mathrm{p}}_{{\mathrm{u}}_{\mathrm{k}}}^{\left(1\right)}$ is the prediction from the k-th unsupervised augmentation branch (either weak or strong), I(⋅) is the indicator function that retains only regions with confidence above the dynamic threshold τ(t).

To enhance the quality of pseudo-labels, we adopt a knowledge distillation loss based on feature perturbation, which consists of the following two components. First, the predictions obtained after image-level strong perturbation and feature-level weak-strong perturbations are both required to be consistent with the pseudo-labels derived from image-level weak perturbation, as formulated in Equation 11. Second, the predictions from feature-level weak-strong enhancements under image-level strong perturbation are enforced to align in the feature space via a KL divergence constraint [13], as shown in Equation 12.

$${\mathrm{L}}_{\mathrm{ws}}={\displaystyle \frac{1}{2}}\left(\mathrm{Dice}\left({\mathrm{p}}_{\mathrm{weak}}^{\left(1\right)},{\stackrel{\mathrm{ˆ}}{\mathrm{y}}}_{\mathrm{u}}\right)+\mathrm{Dice}\left({\mathrm{p}}_{\mathrm{strong}}^{\left(1\right)},{\stackrel{\mathrm{ˆ}}{\mathrm{y}}}_{\mathrm{u}}\right)\right)$$

(11)

Among them, ${\mathrm{p}}_{\mathrm{weak}}^{\left(1\right)}$ is the foreground probability map predicted by the weakly augmented branch for the unlabelled data (i.e., the probability distribution over the target region). ${\mathrm{p}}_{\mathrm{strong}}^{\left(1\right)}$ is the foreground probability map predicted by the strongly augmented branch for the same unlabelled data.

L_consist = KL(p_weak‖p_strong)

(12)

Among them, KL(⋅∥⋅) is The Kullback-Leibler (KL) divergence, which quantifies the dissimilarity between two probability distributions and is asymmetric in nature. p_strong is the full predictive probability distribution produced by the strongly augmented branch. p_weak The full predictive probability distribution produced by the weakly augmented branch (covering all classes).

The overall distillation loss is then defined as a weighted sum of these two terms, where η ∈ [0, 1] controls the supervision strength, as given in Equation 13.

L_KD = ηL_ws + (1 − η)L_consist

(13)

The overall objective function, which combines the supervised loss, the unsupervised loss, and the distillation loss, is presented in Equation 14.

$${\mathrm{L}}_{\mathrm{total}}={\displaystyle \frac{1}{3}}\left({\mathrm{L}}_{\mathrm{x}}+{\displaystyle \frac{1}{2}}\left({\mathrm{L}}_{{\mathrm{u}}_1}+{\mathrm{L}}_{{\mathrm{u}}_2}\right)+{\mathrm{L}}_{\mathrm{KD}}\right)$$

(14)

The design of this total loss function achieves the following goals by balancing the contributions of each loss term. First, labelled data dominates the optimization direction. L_x provides reliable gradients to avoid model shifts caused by pseudo-annotation noise. Second, unlabelled data enhances generalization ability. Dual branch ${\mathrm{L}}_{{\mathrm{u}}_1}$and ${\mathrm{L}}_{{\mathrm{u}}_2}$ improve the model’s adaptability to the variability of medical image data. Third, the knowledge distillation term prevents the model from over-learning erroneous information. Specifically, L_KD enforces prediction consistency under feature-space perturbations, which mitigates the adverse effects of pseudo-label noise and thereby strengthens model robustness.

3. Experimental Design and Result Analysis

3.1. Dataset

According to the experimental requirements, in this work, the training sets of the LA dataset and the ACDC dataset were randomly divided into subsets with a proportion of labelled data of 5%, 10%, and 20%, and subsets with a proportion of labelled data of 5% and 10% respectively, for model training. All the datasets used are public datasets.

The LA dataset was used in the 2018 Left Atrial Segmentation Challenge of the Heart. This dataset contains 100 cases of 3D cardiac MRI data, among which 80 cases are used as the training set and the remaining 20 cases are used as the validation set, with the labelled target being the left atrium.

The ACDC dataset is a public dataset used by the Automated Cardiac Diagnosis Challenge. This dataset covers 100 cases. Among them, 70 cases were used as the training set and 30 cases as the validation set.

3.2. Evaluation Metrics

All the datasets in this paper are evaluated for the performance of the proposed model using the same four evaluation indicators.

Dice score is a statistical indicator used to evaluate the degree of overlap between segmentation results and real labelling. Its value ranges from 0 to 1, with 1 indicating complete overlap and 0 indicating no overlap at all. We use it to evaluate the accuracy of medical image segmentation, such as organ or tumour segmentation [2], as given in Equation 15.

$$\mathrm{Dice}={\displaystyle \frac{2\left|{\mathrm{Y}}_{\mathrm{pred}}\cap {\mathrm{Y}}_{\mathrm{gt}}\right|}{\left|{\mathrm{Y}}_{\mathrm{pred}}\right|+\left|{\mathrm{Y}}_{\mathrm{gt}}\right|}}$$

(15)

where Y_pred is the predicted partition region,Y_gt is the true partition region, and |Y_pred ∩ Y_gt| is the intersection area of the two regions.

The IoU (Intersection over Union) or Jaccard index is used to measure the overlap of two sets and also has a value between 0 and 1. Similar to Dice score, it is used to evaluate the accuracy of segmentation, but IoU is more sensitive to the case of small overlapping areas [14], as given in Equation 16, where |Y_pred ∪ Y_gt| denote the area of the union of the two regions.

$$\mathrm{IoU}={\displaystyle \frac{\left|Y_{\mathrm{pred}}\cap Y_{\mathrm{gt}}\right|}{\left|Y_{\mathrm{pred}}\cup Y_{\mathrm{gt}}\right|}}$$

(16)

Hausdorff Distance, which measures the maximum distance between two-point sets, that is, the maximum minimum distance of all points from one-point set to another. Evaluate the maximum deviation between the segmentation boundary and the true boundary [15], as given in Equation 17.

$$\mathrm{HD}\left({\mathrm{Y}}_{\mathrm{pred}},{\mathrm{Y}}_{\mathrm{gt}}\right)=\max \left(\underset{\mathrm{p}\in {\mathrm{Y}}_{\mathrm{pred}}}{\max }\underset{\mathrm{g}\in {\mathrm{Y}}_{\mathrm{gt}}}{\min }‖\mathrm{p}-\mathrm{g}‖,\underset{\mathrm{g}\in {\mathrm{Y}}_{\mathrm{gt}}}{\max }\underset{\mathrm{p}\in {\mathrm{Y}}_{\mathrm{pred}}}{\min }‖\mathrm{p}-\mathrm{g}‖\right)$$

(17)

where ‖p − g‖ is the Euclidean distance between points p and g, typically defined as the straight-line distance in two-dimensional or three-dimensional space. To mitigate the influence of outliers, the 95% Hausdorff Distance is commonly employed, which is defined as the 95th percentile of the Hausdorff Distance. A smaller value of this metric indicates that the segmentation boundary is closer to that of the ground truth.

The Average Surface Distance (ASD) serves as a metric for quantifying the mean distance between the model’s segmentation output and the corresponding ground-truth labels and is computed as Equation 18.

$$\mathrm{ASD}={\displaystyle \frac{1}{\left|{\mathrm{Y}}_{\mathrm{pred}}\right|}}\sum _{\mathrm{p}\in {\mathrm{Y}}_{\mathrm{pred}}}\underset{\mathrm{g}\in {\mathrm{Y}}_{\mathrm{gt}}}{\min }‖\mathrm{p}-\mathrm{g}‖+{\displaystyle \frac{1}{\left|{\mathrm{Y}}_{\mathrm{gt}}\right|}}\sum _{\mathrm{g}\in {\mathrm{Y}}_{\mathrm{gt}}}\underset{\mathrm{p}\in {\mathrm{Y}}_{\mathrm{pred}}}{\min }‖\mathrm{p}-\mathrm{g}‖$$

(18)

3.3. Experimental Evaluation and Comparative Analysis

Table 1 summarizes the quantitative experimental results on the LA dataset. As can be observed from the table, our proposed method achieves competitive performance across all four-evaluation metrics. When using only 5% labelled data, our method attains a Dice score of 90.35%, which outperforms the BCP method by 2.83 percentage points. Furthermore, with 8% labelled data, our method surpasses all competing approaches across all four metrics. Table 2 presents the quantitative results on the ACDC dataset. It is evident that our method also yields strong performance on this dataset against the same four evaluation criteria. With 5% labelled data, our method achieves superior scores on all metrics compared with the baselines, with a Dice score of 87.83%, exceeding that of the URCA method by 4.52 percentage points.

Table 3 shows the experimental results of ablating different modular components on the LA dataset, where A denotes the baseline model without any of the proposed modules, B represents the UMIX uncertainty-guided module, C stands for the dynamic confidence threshold module, and D refers to the dual-task multi-perturbation module. For a fair comparison, all ablation studies were performed under identical parameter settings.

From the experimental results, we can make the following observations:

• ▪ UMIX uncertainty-guided module: Removing this module leads to a decrease in Jaccard index scores by 0.88%, 2.51%, and 0.67% respectively, indicating that uncertainty-guided local feature replacement effectively suppresses edge noise.
• ▪ Dynamic confidence threshold module: Its removal consistently lowers the Dice scores across all settings, which verifies the validity of the progressive adjustment strategy.
• ▪ Dual-task multi-perturbation module: A comparison of Dice scores before and after removing this module reveals that the performance gap enlarges as the proportion of unlabelled data decreases.

Figure 3 visually presents a comparative display of segmentation results produced by various methods on the LA dataset, where the GT (ground truth) images denote the manually annotated labels. As can be observed, our proposed method yields superior segmentation outcomes relative to the other comparative approaches, which is consistent with the qualitative findings discussed above. Notably, the segmentation results achieved by our method exhibit better boundary continuity and preserve more detailed structures, aligning more closely with the GT images.

The UMIX uncertainty-guided module proposed in this paper effectively reduces the uncertainty associated with pseudo-labels and enhances segmentation performance along the boundaries of medical structures such as the left atrium. Figure 4 provides a visual comparison of uncertainty heatmaps with and without the UMIX module. The lack of interpretability can severely undermine clinical confidence [32], while the UMIX module provides a certain degree of interpretability through entropy heatmaps. The heatmap colour mapping adopts the Plasma colormap from the Matplotlib library, where colours range from deep purple to bright yellow, representing low to high uncertainty. A representative sample from the LA dataset is shown, with the central slice selected for display and visualized across three different views. The results demonstrate that the heatmap of the model optimized with the UMIX module (on the right) exhibits a more stable and consistently deeper purple tone overall compared to the unoptimized version (on the left), with finer and more detailed processing along the edges. These observations substantiate the effectiveness of the proposed module.

4. Discussion

Table 1, Table 2 and Table 3 quantitatively validate the effectiveness of our proposed semi-supervised medical image segmentation method based on dynamic pseudo-label augmentation. By comparing our method with state-of-the-art approaches reported in recent literature, we observe that our method consistently achieves competitive performance across various proportions of labeled data, as measured by four evaluation metrics, namely the Dice coefficient, Intersection over Union (IoU), Hausdorff distance, and Average Surface Distance (ASD). Furthermore, the ablation study results reveal that the inclusion or removal of individual modules directly affects the Dice coefficient and other metrics, confirming that the UMIX uncertainty-guided module effectively suppresses edge noise through uncertainty-guided local feature replacement, while the dual-task multi-perturbation strategy enhances supervision both among unlabeled data and from labeled to unlabeled data, thereby effectively reducing the generation of erroneous pseudo-labels. Furthermore, Figure 3 and Figure 4 qualitatively illustrate that our method yields more stable segmentation results with more refined handling of boundary regions, further substantiating the effectiveness of the proposed approach.

5. Conclusions

Targeting the task of left atrial image segmentation, this paper proposes a semi-supervised medical image segmentation method based on uncertainty-guided dynamic pseudo-label augmentation, with the primary aim of addressing the issues of rigidly fixed confidence thresholds for pseudo-labels and the lack of purposeful guidance in data augmentation. The main contributions are summarized as follows:

To overcome the limitation that a fixed confidence threshold fails to adapt to the dynamic training process, leading to suboptimal utilization of pseudo-labels, we design a threshold update function inspired by exponential decay, which meets the varying quality demands for pseudo-labels at different training stages. In response to the fact that conventional data augmentation methods are often applied indiscriminately and may disrupt the local structures of medical images, we introduce uncertainty-guided image mixing to purposefully expand data diversity and enhance the model’s generalization capability.

Motivated by the ablation study on weak-strong perturbations in the UniMatch model, we incorporate an additional unlabelled data stream and apply both image-level and feature-level weak and strong perturbations, thereby obtaining stronger supervision signals and improving segmentation performance.

Extensive experiments are conducted on two public medical image datasets, including comparisons with state-of-the-art methods, ablation studies on each innovative module, and parameter selection analyses. The experimental results demonstrate that our proposed method effectively improves the utilization of pseudo-labels, reduces predictive uncertainty, and enhances overall segmentation performance.

Nevertheless, the proposed method has certain limitations. First, it entails considerable computational resource demands, the feature replacement operation results in a training memory footprint of 23.6 GB, which constrains its deployment on low-end devices. Second, the method is sensitive to hyperparameter selection where the initial threshold τ_base = 0.85 may require further tuning in scenarios with extreme class imbalance.

To address these limitations, future work can be directed toward the following aspects. (1) Lightweight model design, employing techniques such as channel pruning and dynamic adjustment of network depth to reduce computational requirements. Peng, X. et al. proposed a Fourier-spatial synergistic learning framework that operates under a frame-by-frame real-time inference setting, achieving “clinical-grade” segmentation [33]. (2) Multimodal fusion, exploring joint training on paired CT-MRI data and aligning feature spaces via contrastive learning. The future progress will depend on multi-institutional data curation, self-supervised and foundation model pre-training, privacy-preserving collaboration, and multimodal designs that remain reliable under missing modalities and real-world distribution shifts [34]. (3) Interactive segmentation, enabling clinicians to correct a small number of erroneous predictions and updating the model in real time through online learning, thereby establishing a human-in-the-loop system.