Semi-supervised 3D diabetic macular edema segmentation in OCT volumes via independent dual-branch consistency learning

Abstract

Accurate 3D lesion segmentation in optical coherence tomography (OCT) images is crucial for the detection and treatment evaluation of diabetic macular edema (DME), but is hindered by scarce labeled data and complex boundaries. To address these challenges, we propose a semi-supervised 3D DME segmentation method that leverages a dual independent-branch co-training framework with mask perturbation consistency. The dual branches jointly exploit complementary multi-view features, while an independence-measuring strategy encourages parameter divergence to break performance bottlenecks. To improve robustness against diverse lesion morphologies, we introduce dynamic region-mask augmentation and a bidirectional confidence filtering mechanism, enforcing strong consistency constraints between the two branches. Additionally, a lightweight channel-attention module integrates multi-scale context. Extensive experiments demonstrate that our method surpasses state-of-the-art approaches, validating its effectiveness and superiority.

1. Introduction

Diabetic macular edema (DME) is associated with increased retinal vascular permeability, where intravascular fluid leaks into the macular region, leading to macular swelling. In the case of DME, fluid accumulation in the macular region directly affects central vision, causing blurring and distortion of central vision [1–3]. It is one of the major causes of vision loss in the working-age population worldwide [4]. In light of this, the screening and monitoring of DME are crucial for disease control, with macular volume being a core indicator for the diagnosis and therapeutic effect evaluation of DME [5]. While non-stereoscopic fundus photography was used at the early stage, which resulted in inaccurate assessments, the emergence and related advancements of optical coherence tomography (OCT) have enabled the accurate segmentation of DME. OCT has demonstrated significant utility in various sectors, including non-destructive testing for material structure inspection, industrial metrology, and particularly for biomedical applications, providing crucial support for the diagnosis, therapeutic efficacy, and prognosis of DME [6–9]. Specifically, High-Resolution OCT (HR-OCT) has revolutionized medical diagnostics by enabling the in situ visualization of cellular structures to uncover disease mechanisms [10]. Simultaneously, Multi-Contrast Jones-Matrix OCT (JM-OCT) enhances tissue characterization by extracting diverse optical contrasts like birefringence [11]. Furthermore, recent algorithmic and functional innovations have significantly enhanced imaging precision through artifact-free deblurring and wide-field, high-dynamic-range blood flow quantification [12,13].

Deep learning has achieved remarkable success in OCT image segmentation tasks, ranging from retinal blood vessels [14] to choroidal layers [15,16] and RNFL [17]. Though these methods have shown great potential in image segmentation, the training of the frameworks is still limited by labeled data. Due to the scarcity of labeled data, it has been a challenge to train deep neural networks with large-scale unlabeled data. The issue prompts related research based on semi-supervised learning.

To address challenges like limited labeled data and ambiguous boundaries, various innovative semi-supervised frameworks have been proposed. Among semi-supervised learning approaches, consistency regularization has emerged as a research hotspot [18–22]. Several methods leverage the Mean Teacher (MT) framework for their implementation [23]. Representative works such as Uncertainty Aware Mean Teacher (UAMT) [24] improved the MT architecture by incorporating an uncertainty-aware mechanism to filter out unreliable pseudo-labels. Similarly, Uncertainty Rectified Pyramid Consistency (URPC) [25] leveraged multi-scale pyramid consistency to enhance sensitivity to targets of varying sizes. Besides, H. Zhang et al. [26] developed a novel semi-supervised segmentation approach, prototype-augmented mean teacher (PAMT), which incorporates a non-learnable prototype to enhance feature space supervision. However, Mean Teacher-based approaches rely heavily on the Exponential Moving Average (EMA) of the student model to update the teacher model, lacking source supervision. This tight parameter coupling makes the teacher model vulnerable to improper updates of the student model, where early errors from the student are inherited and reinforced by the teacher—a critical issue when delineating the ambiguous and complex 3D boundaries of DME lesions. Meanwhile, the student model tends to replicate the erroneous pseudo-labels generated by the teacher model, falling into local optima [27].

To mitigate such coupling, Cross Pseudo Supervision (CPS) [28] introduces a dual-network strategy where two parallel networks with different initializations supervise each other via pseudo-labels. While this design avoids the lag of EMA, it lacks explicit constraints on the divergence between branches. Consequently, the two networks tend to converge towards similar feature representations during training, diminishing the view complementarity essential for effective co-training. Subsequent frameworks like UniMatch [29] employ a dual-branch perturbation strategy, performing cross-pseudo supervision between views with different augmentation intensities. Although UniMatch has achieved state-of-the-art performance, it still suffers from implicit parameter homogenization. Furthermore, existing perturbation strategies often fail to generate sufficiently diverse views to capture both the local texture and global spatial context required for accurate DME segmentation. As a result, existing methods struggle to break the performance bottleneck when segmenting the ambiguous boundaries of 3D DME lesions.

Considering these limitations, we propose a semi-supervised framework based on independent dual-branch co-training. Distinct from CPS and UniMatch, we introduce a convolutional-layer independence regularization that minimizes the linear correlation between the convolutional kernels of the two decoders. This explicitly enforces the learning of diverse feature representations, preventing model collapse and error accumulation. The complementary divergence of the dual branches adheres to the multi-view learning principle of co-training, enabling the model to capture richer and more discriminative features [30,31]. Additionally, we replace standard strong augmentations with a dynamic mask perturbation strategy, leveraging the spatial redundancy of 3D volumes to construct complementary views. This forces the model to recover global lesion context from corrupted volumetric data, significantly enhancing robustness. To suppress pseudo-label noise, we further design a bidirectional confidence filtering strategy that selects high-confidence predictions from each branch to supervise the other, enforcing consistent learning on unlabeled data. A channel-attention module is incorporated to fuse multi-scale representations at the decoding output. The shared-encoder and independent-decoder architecture maximizes convolutional feature learning capacity while improving robustness and generalization. It is worth noting that this study is a substantial extension of our preliminary work [32]. While [32] explored basic dual-branch consistent learning, it did not address the critical issue of feature homogenization. The proposed independence constraint and mask perturbation in this work specifically fill this gap. Finally, our contributions are summarized as follows:

• (1)Independent dual-branch co-training framework

  We develop a novel dual-branch co-training learning strategy that effectively leverages complementary structural information for 3D retinal fluid segmentation. Extensive results on a customized DME dataset demonstrate clear performance improvement over existing methods.

• (2)Dynamic mask perturbation with bidirectional confidence consistency

  We introduce regional mask perturbation with reciprocal strengths across the two branches, together with a confidence-aware mutual supervision mechanism operating on shared unmasked regions, significantly enhancing robustness against pseudo-label noise.

• (3)Branch-level parameter independence enhancement

  We propose a convolutional-layer independence regularization method based on linear correlation minimization, preventing parameter collapse between branches and breaking performance bottlenecks.

2. Method

In this section, we will introduce the framework of the proposed model, as illustrated in Fig. 1. This framework comprises a shared encoder and two decoders with independent parameters in their convolutional layers, employing operations such as 3D convolution, normalization, down-sampling, and up-sampling. Suppose that there are a set of labeled data $D_l=\{x_l^i,y_l^i{\}}_{i=1}^{N_l}$ and a set of unlabeled data $D_u=\{x_u^i{\}}_{i=1}^{N_u}$ . Noting that the quantity of unlabeled is far greater than that of labeled data, i.e., $N_l\ll N_u$ . Before inputting labeled and unlabeled data into the network, we perform strong augmentation on the data, specifically by applying masks with different quantities and positions, setting the pixel values of some regions to zero. The specific structure of the network consists of convolutional layers composed of 3D convolution, batch normalization, and the ReLU activation function. In the encoding process, $2\times 2\times 2$ down-sampling is utilized to extract features layer by layer, which doubles the number of feature map channels and reduces the spatial dimensions to half their original size. During decoding, the two decoders adopt $2\times 2\times 2$ trilinear up-sampling and nearest-neighbor up-sampling, respectively, to restore the feature map dimensions. Additionally, after the final up-sampling layer of the decoder and before the final prediction head, we utilize the channel attention module to integrate information from feature maps of different resolutions. Furthermore, threshold-based bidirectional confidence pseudo-supervision is also used to filter out poor-quality noisy pseudo-labels that may be generated during training.

Fig. 1.

Overview of proposed semi-supervised framework. $\Psi (x)$ denotes the operations of random cropping, scaling, flipping, and rotation applied to the original image. Additionally, $H,W,D,$ and C represent the height, width, depth, and number of channels of the image, respectively. $p_u^{(1)}$ , $p_l^{(1)}$ are the predictions of the branch for unlabeled and labeled data, respectively, $p_u^{(2)}$ and $p_l^{(2)}$ are similar. $Enc.$ denotes the encoder, and $Dec.$ denotes the decoder.

2.1. Mask perturbation-based bidirectional pseudo-supervision

For the two decoder branches, we have designed a novel bidirectional pseudo-supervision method, as shown in Fig. 2(a). For the same data input to the network, it is subjected to similar perturbations before being fed into different decoder branches, and the final outputs of the model should maintain consistency. In our design, the number of masks represents the intensity of data perturbation. We set the size of each small mask to $16\times 16\times 16$ , and the number of masks in one branch increases with training iterations, indicating that the perturbation becomes progressively stronger. At the same time, the number of masks in the other branch decreases with the number of iterations. These two opposite perturbation strategies allow the model to use the weak-perturbation branch to correct the strong-perturbation branch throughout the training process. Meanwhile, the strong-perturbation branch can guide the weak-perturbation branch to avoid overfitting, thereby enhancing the model's robustness. The number of masks in the two branches is defined as Eq. (1):

$$\begin{array}{rl}& nu{m}_{Mask1}=max\left\{\left[{\omega }_1\cdot (\frac{H\times W\times D}{siz{e}_{mask}})-{\omega }_1\cdot \frac{iteration}{step}\right],0\right\}\\ & nu{m}_{Mask2}+nu{m}_{Mask1}={\omega }_1\cdot (\frac{H\times W\times D}{siz{e}_{mask}})\end{array}$$ (1)

where $H,W,D$ denote the height, width, and depth of the input image, respectively. $siz{e}_{mask}$ represents the size of each mask. Here we set it to $16\times 16\times 16$ , and $step$ is set to 15, determining the rate of quantity decrease. $iteration$ is the current iteration round in model training, and ${\omega }_1$ determines the maximum number of masks, which is set to 0.5.

Fig. 2.

Overview of our methods. Sub-figure (a) illustrates the bidirectional confidence filtering process based on shared regions. Sub-figure (b) demonstrates the principle underlying the computation of parameter independence across convolutional layers. Sub-figure (c) depicts the specific implementation steps of the CAM designed to recalibrate the feature weights across different channels.

To ensure that this consistency preserves valid information while reducing noise interference, we take the shared regions of the input data that are not zeroed out by masks as potentially usable pseudo-labels. Confidence is then calculated for these regions, and only the one with high confidence is retained as a pseudo-label to guide the parameter update of the other branch. Entropy is utilized to quantify the model confidence, and can be formulated as follows:

$$E^{(i)}=-[p^{(i)}\cdot \log (p^{(i)})+(1-p^{(i)})\cdot \log (1-p^{(i)})]$$ (2)

where i denotes the $i$ -th branch, $p^{(i)}$ represents the prediction output by the branch, $i=1,2$ .

The threshold we set is a non-linear value that dynamically changes with training iterations, and its definition is as follows:

$$T=1-{\mu }_1\cdot {\lambda }_1$$ (3)

where ${\lambda }_1$ is tuned via a time-dependent Gaussian warming-up function [33], and ${\mu }_1$ is a to-be-determined parameter, which is set to 0.6 in our experiments. Through such a setup, the threshold can be continuously adjusted during the training process, thereby enabling different branches to perform consistent learning with increasingly high confidence levels.

For any pixel, the greater the uncertainty in the model's prediction, the larger the corresponding entropy value is. Thus, we consider labels with low entropy values to contain more reliable learnable information. Then, for the two branches, using a threshold to filter pseudo-labels will result in three meaningful scenarios: the output of $Branch1$ has high confidence while the output of $Branch2$ has high uncertainty, thus only the gradients of $Branch2$ participate in the update; only the output of $Branch2$ has high confidence and guides the training of $Branch1$ ; the outputs of both have low uncertainty and supervise each other. Gradients are separated to create pseudo-labels, thereby maintaining the stability of training.

2.2. Independence of convolutional layers in branches

Co-training is often applied to semi-supervised learning tasks. Its basic principle is based on the multi-view nature of data, which can be described from multiple perspectives. By training multiple classifiers simultaneously that each can capture a different subset of features, the complementary information among these classifiers can be utilized to improve classification performance [34]. To address the issue that the weights of each convolutional layer in the branches gradually become similar during training, we introduce the network independence calculation method from CauSSL [35]. The core idea of the independence loss is to enforce structural divergence between the two decoders. By minimizing the linear correlation between the convolutional kernels of the two branches, we ensure they extract diverse features. The min-max optimization process works by first finding the maximum correlation (fitting the dependency) and then updating network weights to minimize this correlation, effectively pushing the branches apart in the parameter space. The specific process is shown in Fig. 2(b). We can use A and B to respectively represent the matrix version of kernel weights in convolutional layers, with a size of $D_{out}\times d$ . $D_{out}$ is the number of output channels. We suppose that the number of input feature channels is $D_{in}$ and the kernel size is $k\times k\times k$ , then $d=D_{in}\times k\times k\times k$ . Specifically, the dependency between the same convolutional layers in the network is defined as follows:

$$\begin{array}{rl}{L}_{in}(A,B;M_B)& =\frac{1}{D_{out}}\sum _{i=1}^{D_{out}}{(\frac{v_{A,i}\cdot q_{B,i}}{|v_{A,i}|\times |q_{B,i}|})}^2\\ q_{B,i}& ={(M_B\times B)}_i\end{array}$$ (4)

where $v_{A,i}$ is the $i$ -th row vector in matrix A, and $q_{B,i}$ denotes the optimal linear combination vector formed by the vector group of B, which approximates $v_{A,i}$ as closely as possible. ${\mathrm{M}}_{\mathrm{B}}$ is the optimal coefficient matrix, whose elements are the optimal linear combination coefficients and which has a size of $D_{out}\times D_{out}$ . For network branches containing multiple convolutional layers, the dependency between branches is defined by the average over all convolutional layers:

$$L_{in}({\theta }_1,{\theta }_2;M_2)=\frac{1}{N}\sum _{i=1}^N{L}_{in}({\theta }_{1,i},{\theta }_{2,i};M_{2,i})$$ (5)

where ${\theta }_{1,i}$ and ${\theta }_{2,i}$ are the weight parameters in the format of matrices. $M_{2,i}$ is the optimal coefficient matrix of the $\mathrm{i}$ -th convolutional layer, and $\mathrm{N}$ represents the number of convolutional layers.

In order to enhance the algorithmic independence, a min-max optimization framework is employed during the training stage. To be specific, at the start of iteration, we first fix the weights of the convolutional layers and update the coefficient matrix to maximize $L_{in}$ , thereby establishing accurate network dependencies. Then, the linear matrix is fixed, and the network parameters are updated to minimize $L_{in}$ . The optimization objective can be expressed as follows:

$$\underset{{\theta }_1,{\theta }_2}{min}\underset{M_1,M_2}{max}[L_{in}({\theta }_1,\overline{{\theta }_2};M_2)+L_{in}({\theta }_2,\overline{{\theta }_1};M_1)]$$ (6)

where ${\mathrm{M}}_1$ and ${\mathrm{M}}_2$ are the coefficient matrix, ${\mathrm{\theta }}_1$ and ${\mathrm{\theta }}_2$ denote the weights of the convolutional layers after gradient separation.

2.3. Channel attention module

In neural networks, shallow high-resolution feature maps possess strong geometric detail representation capabilities, while deep low-resolution feature maps often contain richer semantic information in the feature space. However, such valuable information may be lost during the progressive up-sampling process. Therefore, in the proposed framework, we introduce the channel attention module [36] to integrate feature information across different resolution levels, enabling more accurate pixel-level classification. As shown in Fig. 2(c), this module primarily operates by compressing the spatial dimensions of the input features while preserving the channel dimension. A multi-layer perceptron (MLP) is then employed to recalibrate the channel-wise weights, producing the channel attention map. By performing element-wise multiplication between the attention weights and the original input features, the module effectively integrates multi-scale feature representations, thereby preserving both semantic information and geometric details to the greatest extent as much as possible.

2.4. Loss function

Considering that the model incorporates supervision, pseudo-supervision, and branch independence, we use a joint loss function as the overall optimization objective. In DME lesion segmentation, the volume of the lesion is typically much smaller than the background retinal tissue, leading to a severe class imbalance problem. Standard Cross-Entropy loss often tends to bias towards the background. Therefore, we adopt a combination of Weighted Cross Entropy (WCE) and Tversky Loss. WCE assigns higher weights to lesion voxels to penalize false negatives, while Tversky Loss allows for a more flexible balance between precision and recall, effectively improving segmentation performance for small and irregular lesion regions. They are defined as follows:

$$L\text{os}{{\mathrm{s}}^{(n)}}_{WCE}=-\frac{1}{N}\sum _{i=1}^N{w}_1\cdot y_l^i\cdot \log (p_l^i)+w_2\cdot (1-y_l^i)\cdot \log ((1-p_l^i)),n=1,2$$ (7)

$$L\text{os}{{\mathrm{s}}^{(n)}}_{tversky}=1-\frac{\sum _{\mathrm{i}}{y}_l^i{p}_l^i+\epsilon }{\sum _{\mathrm{i}}{y}_l^i{p}_l^i+\alpha \cdot \sum _{\mathrm{i}}(1-y_l^i)p_l^i+\beta \cdot \sum _{\mathrm{i}}{y}_l^i(1-p_l^i)+\epsilon },n=1,2$$ (8)

$$L\text{os}{\mathrm{s}}_{sup}=\eta \cdot \sum _{n=1}^2(L\text{os}{{\mathrm{s}}^{(n)}}_{WCE}+L\text{os}{{\mathrm{s}}^{(n)}}_{tversky})$$ (9)

where N denotes the number of image voxels, and n represents the different branches. ${\mathrm{w}}_1$ and ${\mathrm{w}}_2$ are coefficients that adjust the weights of foreground and background points, with values inversely proportional to their respective voxel counts. $p_l^i$ and $y_l^i$ respectively represent the predicted pixels of the data and the corresponding true pixels, and $\mathrm{\epsilon }$ is a smoothing term. $\mathrm{\alpha }$ , $\mathrm{\beta }$ , and $\eta$ are coefficients that can be customized according to specific tasks, and are set to 0.7, 0.3, and 0.5 based on preliminary experiments. The total supervised loss is calculated as the sum of the supervised losses from both branches.

For unlabeled data, we constrain the training process by designing a joint consistency loss based on non-masked regions. The computation proceeds in three steps: First, we identify the non-masked spatial regions shared by both branches. Second, we compute the pixel-wise uncertainty using entropy for predictions in these regions. Third, we apply a dynamic threshold T; pixels with uncertainty lower than T are considered “high-confidence”. Finally, these high-confidence predictions from one branch are used as pseudo-labels to supervise the other branch, as formulated in Eq. (11). This loss function is mainly defined by the KL divergence and entropy:

$$D_{KL}^{(i)}=\sum _{i,j=1}^C{p}_u^{(i)}\log (\frac{p_u^{(i)}}{p_u^{(j)}}),i\ne j$$ (10)

$$Los{s}_{con}=\frac{\sum _{i=1}^2[D_{KL}^{(i)}+{\lambda }_2\cdot E(p_u^{(i)})]\cdot (Mask1\cap Mask2)\cdot \delta (E^{(i)}<T)}{\sum _{i=1}^2\delta (E^{(i)}<T)}$$ (11)

where $C$ is the number of classes, and $p_u^{(i)}$ is the predicted image of the $i$ -th branch. T denotes the confidence threshold, used to filter out noisy labels generated by different branches. ${\mathrm{\lambda }}_2$ is a coefficient that balances KL divergence and entropy, usually set to 0.1 based on empirical values. $\delta$ is the indicator function where $\delta (condition)=1$ if the condition is met, else 0.

For the entire model, the final total loss function is defined as follows:

$$Los{s}_{total}={\alpha }_1\cdot \{Los{s}_{sup}+{\beta }_1\cdot [L_{in}({\theta }_1,\overline{{\theta }_2};M_2)+L_{in}({\theta }_2,\overline{{\theta }_1};M_1)]\}+{\gamma }_1\cdot Los{s}_{con}$$ (12)

where ${\mathrm{\alpha }}_1$ and ${\mathrm{\beta }}_1$ are set to 0.5 and 3.0 based on preliminary experiments. And the coefficient ${\mathrm{\gamma }}_1$ is tuned via a time-dependent Gaussian warming-up function [33].

2.5. Overall training procedure

To clarify the logical flow of our proposed framework, specifically the interaction between the mask perturbation, bidirectional confidence filtering, and the independence constraint, the complete training procedure is summarized in Algorithm 1, shown in Fig. 3.

Fig. 3.

Pseudocode for the framework algorithm.

3. Experiment and results

3.1. Experimental setup

3.1.1. Datasets

In this paper, our proposed model is evaluated on a customized volumetric DME dataset for both training and testing. The dataset contains 40 3D volume images (512 × 1024 × 512) from patients diagnosed with DME at Sichuan Provincial People’s Hospital from February 2022 to September 2023. All subjects underwent comprehensive ophthalmic examinations, including non-contact intraocular pressure, slit lamp, wide-angle fundus imaging, ocular axis, and optical coherence tomography. All participants signed written informed consent forms after being informed of the nature of the study and its potential consequences.

The dataset was collected by using a swept-source optical coherence tomography (SS-OCT) device for fundus imaging. This device utilizes a scanning vertical-cavity surface-emitting laser with a wavelength of 1060 $nm$ , featuring an A-scan rate of 400,000 scans per second. It has a scanning range of 6 × 6 $mm$ , providing a lateral resolution of 10 $\mu m$ and an axial resolution of 3.8 $\mu m$ . Within the scanning range, there are 512 B-scan images, each composed of 512 × 1044 pixels. The original images contain a large amount of unnecessary black backgrounds and redundant tissues, which increases the computational load. Therefore, following the approach in [32], we performed appropriate cropping to retain the main lesion areas, resized the images to 256 × 256 pixels, and applied standard normalization for model training.

3.1.2. Implementation details

We used the 3D VNet [37] as the backbone, and the structure of the encoder and decoder in our model is basically consistent with it. The proposed framework was trained on a computer equipped with NVIDIA RTX A6000 (48GB) with PyTorch v1.9.1, CUDA 11.2, and Python 3.8.20. We employed the SGD optimizer with an initial learning rate of 5e-2. During training, the batch size was set to 4, containing an equal number of labeled and unlabeled data, and the number of iterations was set to 9000. In this experiment, a total of 35 3D samples were used for training, which contained 5879 slices, and the remaining 5 samples containing 878 slices were used for testing.

3.1.3. Evaluation metrics

For a comprehensive evaluation of the segmentation results, dice similarity coefficient (DSC), jaccard, 95% hausdorff distance (95HD), and relative volume error (RVE) are employed to assess the segmentation performance. These metrics are shown in the following equations:

$$DSC=\frac{2TP}{2TP+FP+FN}$$ (13)

$$Jaccard=\frac{TP}{TP+FP+FN}$$ (14)

$$\begin{array}{rl}95HD(X,Y)& =95\mathrm{\%}max(hd(X,Y),hd(Y,X))\\ hd(X,Y)& =max\{x\in X\}[d(x,Y)]\end{array}$$ (15)

$$RVE=\frac{|V_{pre}-V_{gt}|}{V_{gt}}$$ (16)

where $\text{TP}$ , $\text{FP},$ and $\text{FN}$ are true positive, false positive, and false negative for binary classification, respectively. X and Y denote the volumetric data composed of all voxel points, and $|X|$ , $|Y|$ are respectively the total number of their corresponding voxels. $d(x,Y)$ represents the minimum Euclidean distance from point x to all points on the surface of Y. $V_{pre}$ is the total volume of model-predicted segmented regions, while $V_{gt}$ is the total volume of segmented regions in the ground truth.

3.2. Performance comparison and analysis

3.2.1. Comparison with other segmentation methods

To validate the effectiveness of our proposed framework, we compare our framework with some classical supervised and semi-supervised segmentation methods on the customized DME dataset, where rigorous hyperparameter tuning and 5-fold cross-validation were implemented to ensure optimal baseline performance, including VNet [37], nnUNet [38], UNETR [39], URPC, UAMT, UniMatch, and TraCoCo [40]. We also include a direct comparison with our previous work, Cao et al. [32], to further demonstrate the advancements of this work. To ensure the fairness and credibility of the result comparison, all networks were trained using the same amount of data. Furthermore, all semi-supervised networks employed VNet as the backbone architecture and utilized the same training configurations. Additionally, 20% and 9% labeled data were respectively used for all semi-supervised methods, while full supervised networks were trained using the complete set of labeled data.

The results of our quantitative comparison with semi-supervised methods are shown in Table 1. Among the existing techniques with 20% labeled data used, it is evident that ours demonstrates the best segmentation performance, which achieved an improvement of 0.0107 in DSC compared with the previous work and also showed an increase of 0.0122 in Jaccard, effectively highlighting the ability to capture multi-angle information from unlabeled data. As a substantial extension of [32], this comparison allows for a clear quantification of the performance gains achieved by addressing the feature homogenization issue. In addition, RVE drops to 0.1446, exhibiting a lower false segmentation rate compared with other networks. Overall, it demonstrates better adaptability to the 3D features of volumetric data. A similar situation also occurs in the experiments with 9% labeled data. Our method outperforms other methods in overall performance, with DSC, Jaccard, 95HD, and RVE values of 0.8067, 0.6779, 8.4548, and 0.2082, respectively.

Table 1. Performance comparison with other segmentation methods using 5-fold cross-validation on the customized DME dataset.

Method	Training		Metrics				Parameters
	Labeled	Unlabeled	DSC $\uparrow$	Jaccard $\uparrow$	95HD $\downarrow$	RVE $\downarrow$
UAMT	7	28	0.7659 ± 0.0131	0.6275 ± 0.0150	20.3434 ± 12.7557	0.2551 ± 0.0548	29.2701 M
URPC			0.7951 ± 0.0290	0.6710 ± 0.0373	24.0824 ± 14.5214	0.2185 ± 0.1252	5.8845 M
UniMatch			0.7886 ± 0.0405	0.6576 ± 0.0523	4.9443 ± 1.7727	0.1558 ± 0.0567	59.4632 M
TraCoCo			0.6785 ± 0.0411	0.5237 ± 0.0419	28.3166 ± 4.8967	0.3039 ± 0.1274	9.4489 M
Cao et al. [32]			0.8030 ± 0.0238	0.6763 ± 0.0311	8.2369 ± 6.2061	0.2000 ± 0.0341	15.6523 M
Ours			0.8137 ± 0.0209	0.6885 ± 0.0287	8.2452 ± 6.1020	0.1446 ± 0.0741	13.1761 M
UAMT	3	32	0.7427 ± 0.0429	0.6059 ± 0.0355	31.7701 ± 20.7417	0.3787 ± 0.5167	29.2701 M
URPC			0.6742 ± 0.0464	0.5321 ± 0.0552	52.7517 ± 21.7984	0.6304 ± 0.1101	5.8845 M
UniMatch			0.7735 ± 0.0334	0.6379 ± 0.0420	9.4628 ± 10.3632	0.1633 ± 0.0299	59.4632 M
TraCoCo			0.6344 ± 0.0842	0.4509 ± 0.0518	25.1493 ± 14.5858	0.4158 ± 0.1145	9.4489 M
Cao et al. [32]			0.7617 ± 0.0493	0.6234 ± 0.0619	22.4784 ± 23.9073	0.1968 ± 0.0268	15.6523 M
Ours			0.8067 ± 0.0125	0.6779 ± 0.0173	8.4548 ± 8.9046	0.2082 ± 0.0583	13.1761 M
VNet	35	0	0.8143 ± 0.0236	0.6903 ± 0.0319	6.7931 ± 3.8107	0.1872 ± 0.0641	9.4553 M
nnUNet			0.8076 ± 0.0285	0.6764 ± 0.0400	13.4792 ± 17.9676	0.2034 ± 0.0483	30.7854 M
UNETR			0.8072 ± 0.0226	0.6837 ± 0.0275	7.2055 ± 5.0185	0.1648 ± 0.0488	74.1245 M

We further perform quantitative experiments involving various supervised approaches, as shown in Table 1. It can be observed that after parameter fine-tuning, our method achieves performance comparable to that of fully supervised methods. While nnUNet and UNETR utilize advanced architectures and different backbones compared to our VNet-based framework, they were included in Table 1 primarily to establish a high-level performance upper bound using the full set of 35 labeled volumes. Specifically, compared with UNETR, the DSC increases from 0.8072 to 0.8137, and compared with nnUNet, the 95HD decreases from 13.4792 to 8.2452, underscoring the efficiency of the proposed dual-branch consistency learning framework in bridging the gap between semi-supervised and fully supervised learning. While the aggregate 5-fold cross-validation performance is summarized in Table 1, we specifically present the detailed results of a representative fold in Table 2, Table 3, and Fig. 4. The inclusion of these tables is motivated by the need for a granular performance analysis beyond simple statistical averaging. This fold encompasses a more diverse range of clinically challenging cases, such as volumes with blurred lesion boundaries and multiple small-scale fluid components. These cases provide a more rigorous testbed for evaluating the efficacy of the proposed independence constraint and mask perturbation mechanisms. Furthermore, maintaining this representative fold for ablation studies and detailed comparisons allows for a more granular inspection of model behavior under complex pathological conditions without loss of generality in performance trends. In Fig. 4 and Table 3, we include ConDSeg [41] (based on single-fold results) as a reference method to further evaluate segmentation capabilities in specific challenging cases. Note that in this single-fold experiment, our method outperforms other supervised and semi-supervised methods in segmenting lesions, especially when the lesion region is blurred and contains multiple small components (yellow box in Fig. 4). Furthermore, in the 3D views of segmentation results, other methods tend to exhibit more missing segmentations or false segmentations, which further leads to larger overall volume errors.

Table 2. Performance comparison with semi-supervised methods on the customized DME dataset (results from a single fold).

Method	Training		Metrics
	Labeled	Unlabeled	DSC $\uparrow$	Jaccard $\uparrow$	95HD $\downarrow$	RVE $\downarrow$
UAMT	7	28	0.7879 ± 0.0570	0.6530 ± 0.0779	3.2782 ± 1.1497	0.3210 ± 0.2314
URPC			0.8254 ± 0.0324	0.7039 ± 0.0458	23.8441 ± 38.1743	0.1232 ± 0.0760
UniMatch			0.7864 ± 0.0600	0.6517 ± 0.0789	5.6330 ± 5.9777	0.1770 ± 0.1052
TraCoCo			0.7377 ± 0.0560	0.5867 ± 0.0687	26.1994 ± 30.3101	0.2430 ± 0.1042
Ours			0.8455 ± 0.0120	0.7325 ± 0.0180	2.9215 ± 1.9576	0.0782 ± 0.0823
UAMT	3	32	0.7658 ± 0.0625	0.6238 ± 0.0829	5.7914 ± 2.3076	0.1452 ± 0.0728
URPC			0.5916 ± 0.1673	0.4378 ± 0.1811	84.9154 ± 43.0940	0.4686 ± 0.2021
UniMatch			0.7587 ± 0.0662	0.6152 ± 0.0860	5.4829 ± 5.4094	0.1843 ± 0.1151
TraCoCo			0.6517 ± 0.0913	0.4891 ± 0.1064	16.7867 ± 22.3883	0.2601 ± 0.1703
Ours			0.8236 ± 0.0148	0.7003 ± 0.0218	3.6543 ± 2.7491	0.1234 ± 0.0717

Table 3. Performance comparison of supervised methods on the DME dataset (results from a single fold).

Method	Training		Metrics
	Labeled	Unlabeled	DSC	Jaccard	95HD	RVE
VNet	35	0	0.8031 ± 0.0495	0.6732 ± 0.0678	3.7001 ± 2.0948	0.2676 ± 0.1839
nnUNet			0.8083 ± 0.0517	0.6525 ± 0.6310	44.7282 ± 31.2225	0.1511 ± 0.1250
UNETR			0.7974 ± 0.1058	0.6733 ± 0.1461	4.2887 ± 3.1636	0.2284 ± 0.2090
ConDSeg			0.7911 ± 0.0444	0.6562 ± 0.0610	3.6068 ± 2.2335	0.1658 ± 0.0999
Ours	7	28	0.8455 ± 0.0120	0.7325 ± 0.0180	2.9215 ± 1.9576	0.0782 ± 0.0823

Fig. 4.

Visualization of the segmentation results from different methods. a(1)-a(3) demonstrate the raw 3D OCT data, a cross-section of it, and the corresponding 2D image of the cross-section, respectively. b(1)-b(3) show the labeled lesion, a cross-section of it, and the corresponding 2D image of the cross-section, respectively. The first and second lines are overlays of the 2D slices of the original images and the cross-sections of the predictions, and the third and fourth lines are 3D views of the segmentation results.

3.2.2. Ablation study

We conduct ablation experiments on the 3D DME dataset to validate the effectiveness of each component in our framework, with 20% labeled data used for training. The VNet served as the baseline model after appropriate adjustments (with two identical decoders), and other components were added incrementally to it. Table 4 shows the results, from which we can observe that with three designs of PI, RMBC, and CAM, the Jaccard achieved an improvement of 5.37% compared to the baseline, while the 95HD notably decreases by 6.311. Furthermore, the table also indicates that when each component acts individually, CAM achieves the most significant improvement in the model's segmentation similarity, while PI enhances the model's sensitivity to boundaries. When the components are combined in pairs, both PI and RMBC require the synergistic effect of CAM to enhance the model's performance. Overall, the combined use of PI, CAM, and RMBC not only absolutely improves the similarity between the segmentation and the ground truth, but also makes more accurate and reliable predictions on the surface boundaries of lesions. These results demonstrate the effectiveness of our design in capturing multi-dimensional information of the data.

Table 4. Ablation study on the DME dataset with 20% labeled data. “PI” denotes parameter independence, “RMBC” denotes regional mask-based bidirectional confidence, and “CAM” denotes channel attention module, respectively.

Setting			Metrics
PI	RMBC	CAM	DSC	Jaccard	95HD	RVE
√	√	√	0.8455 ± 0.0120	0.7325 ± 0.0180	2.9215 ± 1.9576	0.0782 ± 0.0823
√	√		0.8054 ± 0.0361	0.6756 ± 0.0549	3.4895 ± 1.0272	0.1292 ± 0.1194
	√	√	0.8401 ± 0.0128	0.7244 ± 0.0175	3.3069 ± 2.6428	0.0958 ± 0.0484
√		√	0.8433 ± 0.0112	0.7295 ± 0.0143	3.2826 ± 2.2116	0.0948 ± 0.0478
√			0.8206 ± 0.0218	0.6963 ± 0.0318	3.2930 ± 1.6989	0.1342 ± 0.0717
	√		0.7960 ± 0.0401	0.6625 ± 0.0579	5.8575 ± 3.5086	0.1271 ± 0.1209
		√	0.8232 ± 0.0169	0.6998 ± 0.0247	3.5725 ± 2.0450	0.1204 ± 0.0694
			0.8088 ± 0.0112	0.6788 ± 0.0156	9.2525 ± 10.8332	0.1357 ± 0.0341

To further illustrate the role of each component in the proposed framework, we also visualize the segmentation results. Figure 5 shows the visualization results, and the degree of overlap between the yellow and red regions reflects the predictive accuracy. Without any components, the regions of false segmentation increase significantly (green rectangular box in Fig. 5). Additionally, when some of the components are added to the network, it still performs poorly in tiny lesion regions (blue arrow in Fig. 5). The joint use of all components enables accurate segmentation of complex lesion regions. Figure 6 also intuitively demonstrates the effectiveness and stability of our method by plotting graphs.

Fig. 5.

Visualization results of ablation studies on the DME dataset with 20% labeled data. All images are obtained by superimposing original images, lesions, and segmentation results. Yellow regions represent lesions segmented correctly, red regions represent lesions predicted by networks, and green regions indicate the under-segmented lesions.

Fig. 6.

Mean-standard deviation charts of the results from different methods in DSC and Jaccard. The figure at the top-left corner displays the results of all methods, where all semi-supervised methods utilized 20% labeled data. The figure at the top-right corner shows the results of the ablation experiments. The second row presents the models’ results under 20% and 10% labeled data, respectively.

4. Discussion

4.1. Parameter analysis

We further conduct a discussion on two parameters and a training setting to investigate the impact on the model’s performance.

Table 5 and Fig. 7(a) show the effects of using different values of ${\mu }_1$ mentioned in Eq. (3) on the model’s final performance, and the experiments were performed on the customized dataset with 20% labeled data. The table shows that regardless of the value of ${\mu }_1$ , our method outperforms the others in terms of the segmentation performance. In addition, with different ${\mu }_1$ values, it can be observed that our model demonstrates relatively stable performance without significant data fluctuations. However, it is worth noting that when the ${\mu }_1$ value is set too large, the metrics will decrease; when the ${\mu }_1$ value is set increasingly small, the performance also shows slight fluctuations and degradation. Theoretically speaking, an excessively large ${\mu }_1$ value will cause the model to filter out too much valid information in the later stage, thereby leading the two branches to be trained under unsupervised conditions. Conversely, an overly small ${\mu }_1$ value may cause the model to learn some noise, resulting in model degradation. Therefore, based on our experimental results, we set the ${\mu }_1$ value to 0.6 to ensure that while filtering out noise, the quality of pseudo-label supervision is improved.

Table 5. The average DSC under different values of ${\mathit{\mu }}_1$ .

${\mu }_1$	1.0	0.9	0.8	0.7	0.6	0.5	0.4	0.3	0.2	0.1	0.0
20% labeled	0.8270	0.8396	0.8388	0.8421	0.8455	0.8413	0.8412	0.8437	0.8403	0.8406	0.8332
9% labeled	0.8112	0.8165	0.8205	0.8201	0.8236	0.8221	0.8194	0.8157	0.8166	0.8146	0.8063

Fig. 7.

Result of different parameters on the DME dataset. (a) shows the mean DSC performance under different ${\mu }_1$ , and (b) shows the mean DSC performance under different ${\omega }_1$ .

Table 6 and Fig. 7(b) illustrate the influence of the number of masks on segmentation performance under different values of ${\omega }_1$ (Eq. (1)). The data indicates that when the number of masks is set to 0.5 of the maximum number, the model achieves the optimal performance. When there are an excessive number of masks during training (i.e., overly strong perturbations), the useful knowledge that can be learned in the co-training process decreases, leading to suboptimal results; if the number of masks is reduced from the optimal value, the model's generalization ability will be limited, and its adaptability to noise in image data will decline. Therefore, in this experiment, we set parameter ${\omega }_1$ to 0.5 to optimize the model performance.

Table 6. The average DSC under different values of ${\mathit{\omega }}_1$ .

${\omega }_1$	1.0	0.9	0.8	0.7	0.6	0.5	0.4	0.3	0.2	0.1	0.0
20% labeled	0.7748	0.8261	0.8281	0.8381	0.8368	0.8455	0.8427	0.8352	0.8377	0.8366	0.8276
9% labeled	0.7458	0.8105	0.7772	0.8120	0.8019	0.8236	0.8217	0.8208	0.8210	0.8134	0.8174

Table 7 shows the impact of using labeled data with different proportions on the model’s performance in each batch. Results indicate that when there is only one labeled sample in each batch, the model performs the worst. When the number of labeled and unlabeled data samples in each batch is balanced, the model achieves the best performance; however, if the proportion of labeled data is further increased, it leads to a slight decrease in segmentation accuracy. This may be because when there is only one labeled sample, the supervisory effect of model training is insufficient, while an excessive number of labeled samples may lead to overfitting of the samples, resulting in a slight negative impact.

Table 7. The average DSC under different labeling proportions in each batch.

labeled: unlabeled	1:3	2:2	3:1
DSC	0.8300	0.8455	0.8443

4.2. Clinical analysis

To clarify the critical significance of diabetic macular edema lesion segmentation results for clinical treatment staging assessment, we focused on the correlation between DME lesions and retinal thickness to conduct a qualitative analysis. Given that DME edematous regions are highly concentrated between the Outer Plexiform Layer (OPL) and the Inner Nuclear Layer (INL), and that thickness changes in this region are a direct indicator of lesion severity. We first completed the accurate annotation of OPL and the inner segment (IS) in OCT images using LabelMe software. This annotation process essentially constitutes an indirect segmentation of the involved range of DME lesions, which locks in the core region for subsequent thickness assessment. On this basis, we specifically visualized the interlayer thickness between OPL and IS, and combined it with the top view of predicted DME lesion segmentation results to establish the corresponding correlation between thickness heatmaps and lesion distribution.

As shown in Fig. 8, the central macular thickness heatmap of a normal fundus (a1) exhibits a typical characteristic of “central depression with gradual peripheral transition”. In contrast, the central macular region of DME patients shows significantly greater thickness than the peripheral area, forming a distinctive “central protrusion” manifestation. Furthermore, there is a high spatial correlation between the DME lesion segmentation results and the regions with abnormal OPL-IS interlayer thickness, such that the edematous range defined by lesion segmentation exactly corresponds to the regions with abnormally increased values in the thickness heatmap, as indicated by the white dashed circles. This finding directly confirms that DME lesion segmentation results can accurately locate the involved range of edema, and the quantitative assessment of thickness in this range (especially the OPL-INL interlayer thickness) serves as a core indicator reflecting the degree of macular edema, ultimately leading to abnormal thickness of the entire central macula.

Fig. 8.

Qualitative comparison of retinal thickness. a(1)-a(3) show the retinal thickness heatmap, OPL-IS thickness heatmap, and top view of the predicted lesion of a normal fundus, respectively. b(1)–b(3) and c(1)–c(3), respectively, display the retina thickness and OPL-IS thickness heatmaps of fundi from two DME patients and the corresponding top views of predicted lesions.

Overall, our proposed method exhibits significant advantages in DME lesion segmentation. With limited labeled data, it leverages unlabeled data for consistency learning and explores the independence of network parameters. It utilizes data augmentation and the confidence filtering method based on multiple random masks, without adding extra redundant components, reducing computational overhead while significantly improving the model's segmentation performance. Meanwhile, it further clarifies the clinical value of DME lesion segmentation—not only is it a prerequisite for the accurate assessment of retinal thickness, but it also provides clinicians with a visualizable and quantifiable lesion assessment tool by establishing a quantitative correlation between lesion scope and thickness changes. This directly supports the accurate determination of DME treatment staging and significantly enhances the objectivity and scientificity of clinical diagnosis.

5. Conclusion

In this paper, we propose a novel semi-supervised framework for 3D retinal lesion segmentation in OCT images, which effectively exploits limited labeled data and fully leverages feature information from unlabeled data. The framework adopts a shared encoder with two parameter-independent decoders. A dynamic mask perturbation strategy and a bidirectional confidence filtering mechanism are designed to enforce consistency between the two branches, suppressing noise propagation in pseudo-labels and enhancing mutual learning. In addition, a channel-attention module is employed to integrate multi-scale contextual information and improve boundary localization accuracy.

Extensive experiments on a customized DME dataset demonstrate that the proposed method consistently outperforms state-of-the-art semi-supervised and fully supervised segmentation approaches. Moreover, the strong spatial correspondence observed between segmented lesion regions and areas of abnormal retinal thickness highlights the clinical value of accurate 3D lesion quantification for DME assessment, supporting objective diagnosis and treatment planning.

Despite these advantages, our method still faces certain limitations. The performance may be affected by variations in data distribution introduced by different imaging devices. In future work, we will further enhance the model’s generalizability across heterogeneous datasets and improve robustness against domain shifts in OCT acquisition.

Funding

National Natural Science Foundation of China https://ror.org/01h0zpd94 ( 61905036); MyBright Medical Science Funding ( 2025-5100-07-000317); China Postdoctoral Science Foundation https://ror.org/0426zh255 ( 2021T140090, 2019M663465); Fundamental Research Funds for the Central Universities https://ror.org/02jnxpg76 (University of Electronic Science and Technology of China) ( ZYGX2021J012); Medico-Engineering Cooperation Funds from University of Electronic Science and Technology of China ( ZYGX2021YGCX019).

Disclosures

The authors declare no conflicts of interest.

Data availability

The data underlying the results presented in this paper are not publicly available at this time but may be obtained from the authors upon reasonable request.