2.1. Early Development of TMJ MRI Segmentation

Medical image segmentation has evolved from traditional manual or semiautomated methods to deep learning-based approaches. Early studies of the TMJ primarily relied on manual delineation or simple thresholding and edge detection techniques that were time-consuming and suffered from high inter-observer variability. With the advent of CNNs, fully convolutional networks and U-Net ⁶ became the foundation for biomedical segmentation. The symmetric encoder–decoder architecture with skip connections in U-Net enabled precise delineation of small anatomical structures. Subsequent variants, including 3D U-Net ⁷ and Attention U-Net ²⁰, further improved volumetric segmentation performance.

Specific to TMJ MRI, several deep learning studies have targeted articular disc and condyle segmentation ²¹. Ito et al. ⁹ proposed an automated segmentation method for articular discs using deep learning on magnetic resonance images. Li et al. ¹⁰ and Vinayahalingam et al. ¹² developed CNN-based frameworks for TMJ structure delineation. These studies demonstrated the feasibility of deep learning for TMJ analysis but revealed persistent limitations when dealing with ultra-thin (1–2 mm) articular disc.

2.2. Limitations of Standard CNN-Based Frameworks

Although U-Net and its variants have achieved promising results in medical segmentation, they exhibit inherent limitations for small and low-contrast targets such as TMJ articular discs. Standard downsampling operations (typically in four stages) cause an irreversible loss of fine spatial details, whereas severe class imbalance biases predictions toward the dominant background and condylar regions. The self-configuring nnU-Net framework ⁸ has attained state-of-the-art performance across numerous biomedical tasks by automatically adapting preprocessing and network topology. However, direct application of the default nnU-Net configuration to TMJ MRI still yields suboptimal disc segmentation accuracy owing to the insufficient preservation of thin-structure features and inadequate boundary refinement.

Recent publications in Journal of Advanced Computational Intelligence and Intelligent Informatics have addressed similar challenges in lightweight and efficient medical segmentation. For instance, He et al. ¹³ proposed a linear transformer-based U-shaped lightweight network that combines CNNs and transformers to achieve a balance between accuracy and computational efficiency, highlighting the importance of task-specific architectural optimizations.

2.3. Advances in Attention Mechanisms, Loss Functions, and Task-Oriented Improvements

Researchers have integrated attention modules and hybrid loss functions to overcome boundary ambiguity and feature discrimination issues. Schlemper et al. ²⁰ introduced attention-gated networks, whereas the convolutional block attention module (CBAM) has been widely adopted to enhance both channel-wise and spatial feature representations. In parallel, joint loss strategies combining Dice loss with focal loss or boundary-aware terms have proven effective for small-object and imbalanced segmentation tasks ²²,¹¹.

Transformer-based approaches such as TransUNet ²³, Swin-UNet ²⁴, and foundational vision transformer models ²⁵ have recently gained popularity for capturing long-range dependencies; however, they typically require large datasets and incur high computational costs, limiting their applicability to clinical TMJ MRI with limited annotated samples. By contrast, targeted modifications to the nnU-Net framework, such as reduced downsampling, bottleneck attention mechanisms, and customized joint loss functions, provide a more efficient and clinically practical solution for thin-structure segmentation.

Despite these advances, few studies have simultaneously addressed the dual-scale challenge of TMJ (large condyle versus ultra-thin disc) within a single nnU-Net-based pipeline. This study bridges this gap by introducing three synergistic modifications to the nnU-Net framework to achieve superior articular disc segmentation without compromising condylar accuracy.

3.1. Overall Framework

The proposed method is built upon the nnU-Net framework and is specifically designed for simultaneous segmentation of the mandibular condyle and articular disc from TMJ MRI ⁸. Unlike conventional segmentation tasks that focus on relatively large and homogeneous anatomical structures, TMJ segmentation presents a dual-scale challenge: the condyle occupies a comparatively large region with relatively clear boundaries, whereas the articular disc is a thin, small, and low-contrast structure embedded in complex surrounding tissues.

To address this imbalance, we introduce a task-oriented modification of the nnU-Net framework that preserves the fine spatial details critical for disc segmentation while maintaining sufficient contextual understanding for condylar structures. The overall framework follows an encoder–decoder paradigm with skip connections but incorporates targeted architectural and optimization-level adaptations to mitigate information loss, enhance feature discrimination, and improve boundary delineation.

Figure 1 illustrates the proposed improved nnU-Net framework for simultaneous segmentation of the mandibular condyle and articular disc from TMJ MRI. The architecture reduces the number of downsampling stages from four to three to preserve fine spatial details of the thin articular disc, integrates a CBAM at the bottleneck, and is trained with a joint loss function.

3.2. Customized nnU-Net Architecture

3.3. Joint Loss Function Design

Severe class imbalances and boundary ambiguities present significant challenges in TMJ articular disc segmentation. The disc occupies only a small fraction of the image volume, making standard loss functions prone to bias toward the dominant background and condylar regions. Moreover, partial volume effects and blurred boundaries further complicate accurate delineation.

To address these issues, we design a joint loss function that combines Dice, focal, and boundary losses, each targeting a specific limitation of conventional optimization strategies. The Dice loss directly optimizes region overlap and is effective in mitigating class imbalances at the global level. The focal loss dynamically downweights easily classified samples, encouraging the network to focus on hard-to-segment disc regions. The boundary loss explicitly penalizes discrepancies along object contours, improving boundary sharpness and reducing edge blurring.

The final loss function is defined as a weighted sum of these three components, enabling simultaneous optimization of region accuracy, hard-sample sensitivity, and boundary precision. This joint formulation is particularly well-suited for thin anatomical structures such as TMJ articular discs, in which accurate boundary localization is critical for clinical interpretation.

Figure 2 show the detailed view of the three proposed modifications to the nnU-Net framework.

4.1. Dataset and Preprocessing

This study was conducted using a retrospective TMJ MRI dataset collected from patients diagnosed with temporomandibular disorders. The dataset consisted of MRI scans of 50 patients acquired using a 1.5T/3.0T MRI scanner with a dedicated TMJ imaging protocol. All images were obtained in the closed-mouth position that is commonly used for clinical evaluation of TMJ disc morphology.

Manual annotations of the mandibular condyle and articular disc were performed slice-by-slice by experienced oral and maxillofacial radiologists and subsequently reviewed by a senior expert to ensure annotation consistency. These annotations served as the reference standard for supervised learning.

Prior to training, all MRI volumes were resampled to a uniform voxel spacing and intensity-normalized following the default preprocessing pipeline of nnU-Net. Data augmentation strategies, including random rotation, scaling, and elastic deformation, were employed to enhance the model robustness and mitigate overfitting.

4.2. Computational Environment

All experiments were implemented using the PyTorch framework and conducted on a workstation equipped with an NVIDIA GPU. The proposed method was trained following the nnU-Net training strategy, with modifications strictly limited to the architectural components described in Section 3 to ensure a fair comparison.

The dataset was divided into training, validation, and testing sets at the patient level to prevent data leakage. Model training was performed using a fixed number of epochs, and the best-performing model on the validation set was selected for final evaluation. The Adam optimizer was employed with an initial learning rate consistent with the nnU-Net defaults, and learning rate decay was applied during training.

For a baseline comparison, the original nnU-Net framework with the standard configuration was trained under identical experimental conditions, including data preprocessing, augmentation strategies, and training schedules. This setup ensured that the observed performance differences could be attributed to the proposed modifications rather than to implementation or training discrepancies.

The segmentation performance was quantitatively evaluated using the Dice similarity coefficient (Dice) and Hausdorff distance (HD). The Dice measured the volumetric overlap between the predicted segmentation and the ground truth, providing an overall assessment of the region-level accuracy. The HD evaluated the maximum surface distance between two segmentations and was particularly sensitive to boundary errors.

Given the thin morphology of the TMJ articular disc, boundary accuracy is critical for clinical interpretation. Therefore, the HD was included as a metric complementary to the Dice to better capture the segmentation discrepancies along the object contours. All metrics were computed separately for the mandibular condyle and articular disc and averaged to provide an overall performance assessment.

4.3. Comparison with Baseline Methods

The proposed method was compared with the standard nnU-Net framework to evaluate the effectiveness of the introduced modifications. The quantitative results are summarized in Table 1.

Both methods achieved high Dice scores for mandibular condyle segmentation, reflecting the relatively large size and clear boundaries of the structure. The proposed method demonstrated a comparable or slightly improved performance, indicating that the architectural modifications did not compromise the segmentation of larger anatomical regions.

In contrast, substantial improvements were observed in articular disc segmentation. The proposed method achieved a higher Dice score and noticeably reduced HD compared with the baseline nnU-Net. These results suggested that the reduced downsampling strategy and attention-enhanced bottleneck effectively preserved the fine structural details and improved boundary delineation for the thin-disc structure.

To further illustrate these improvements, representative segmentation results from the four subjects are shown in Fig. 3. The proposed method demonstrated significantly sharper boundaries and fewer missed thin-disc regions than the baseline nnU-Net and other state-of-the-art models.

4.4. Ablation Study

An ablation study was conducted by incrementally introducing the proposed modifications to assess the contribution of each component in the proposed framework ¹³,¹⁴. Specifically, we evaluated (1) the baseline nnU-Net, (2) nnU-Net with reduced downsampling, (3) nnU-Net + DynFocus, and (4) the full model with the joint loss function.

As presented in Table 2, reducing the number of downsampling stages led to noticeable improvements in articular disc Dice and HD, confirming the importance of preserving spatial resolution for thin structures. The ablation results confirmed that the DynFocus framework served as a pivotal component for boundary refinement. By implementing DynFocus at the bottleneck, the model achieved a significantly lower HD (1.45 vs. 1.55), proving its superior capability in delineating the precise contours of the articular disc.

In this study, the Dice coefficient was selected as the evaluation metric. The results obtained from the dataset used in this study are summarized in Table 1. Our model achieved an average Dice coefficient of 0.80, with 0.89 for the condylar region and 0.71 for the articular disc. In comparison, U-Net yielded a lower average Dice coefficient of 0.63, with 0.81 for the condylar region and only 0.45 for the articular disc. These results were significant because they demonstrated that the proposed model detected the presence of the articular disc and successfully located it. The proposed model consistently outperformed U-Net, U-Net++, and SWin-UNet in terms of segmentation accuracy. For both the condyle and articular disc, the Dice coefficient of the proposed model exceeded those of the other models. Although the Dice coefficient of the proposed model for the condylar region was slightly lower than that of nnU-Net, it outperformed for the articular disc, the target area of this study.