Outlier-robust self-training for semi-supervised clavicle semantic segmentation in computed tomography images

Introduction

The field of image segmentation continues to evolve, with deep learning emerging as a pivotal technology in medical image analysis. Deep learning techniques have demonstrated superior performance regardless of texture or object morphology. Notably, in the field of biomedical science, there are state-of-the-art models such as U-Net (1) and 3-dimensional (3D) U-Net (2), which utilize 3D datasets. These models typically consist of contracting paths for feature extraction and expanding paths for full-resolution reconstruction, resulting in precise segmentation of regions of interest (ROI). U²-Net (3) enhances feature extraction by integrating sequential contraction and expansion operations at each layer. Various applications leverage these structured networks, with U-Net++ (4) also demonstrating remarkable performance in segmentation tasks. Additionally, several networks employing compression-expansion architectures have been proposed (5-7). Furthermore, research on bone segmentation using computed tomography (CT) images is prevalent (8-10). Such studies facilitate the rapid acquisition of segmented images through auto-mated segmentation processes.

Segmenting bones in CT images is crucial for studying anatomical structures. In (11), it was reported that clavicle fractures occur periodically, and in severe cases, treatment is performed using titanium plates. Additionally, as reported in (12), clavicle fractures can be caused by sports activities, road traffic accidents (RTAs), and even simple falls. Accordingly, the acquisition of diverse clavicle morphologies and extensive three-dimensional clavicle datasets is required for the proper design of such plates. Accurate bone segmentation aids in efficient fracture treatments. However, due to variations in bone structures among individuals, conventional computer vision techniques may struggle with segmentation tasks. Although segmentation via thresholding is feasible (13), additional efforts are required for outlier removal. Femur segmentation, for instance, involves reducing the ROI at both ends, characterized by elongated shapes (14). Similarly, clavicular shapes exhibit elongation, with reduced ROIs at the sternal and acromial ends. Moreover, midshaft regions demonstrate bending and twisting characteristics. The characteristics of bending exhibit similar patterns in structures other than bones, such as blood vessels and nerves (15). In Hogeweg et al.'s study (16), the authors segmented the clavicle, but analyzing chest radiograph data is limited. These complexities necessitate various approaches for enhancing segmentation accuracy.

In order to utilize deep learning techniques, annotated data acquisition, also known as annotation, is required. Hospitals and medical institutions have accumulated medical imaging data such as CT scans, magnetic resonance imaging (MRI), and X-rays over several years. The annotation process of medical image data is tedious and repetitive, incurring significant time and cost. The quantity of labeled data created is often insufficient compared to the increasing amount of data. Therefore, attempts have been made to utilize unlabeled data for training. Active learning is one such approach where the model efficiently selects which data to use for training from the labeled and unlabeled dataset (17), or creates uncertainty queries from the unlabeled dataset for submission to an oracle (e.g., human annotator) for labeling (18). Similarly, a semi-supervised learning approach known as self-training (19) utilizes unlabeled datasets. This involves the model inferring labeled data from the unlabeled dataset and automatically adding it to the training dataset without external intervention. LightSegMamba (20) is a lightweight Mamba-based 3D segmentation model developed for accurate tumor detection in ABUS imaging. By leveraging a state-space Mamba backbone, it reduces computational cost while effectively modeling spatial and volumetric information. Furthermore, its Dense Atrous Spatial Pyramid Pooling (DASPP) module extends Atrous Spatial Pyramid Pooling (ASPP) into 3D to capture multiscale features, enabling robust segmentation of tumors with varying sizes and shapes. Efforts are made to effectively utilize unlabeled data as an alternative to enhance model performance for such problems.

Labeled data obtained through the self-training approach may contain unexpected outliers. Datasets with outliers removed are effective in improving model performance (21). Therefore, training strategies should be developed to detect and remove outlier data beforehand. Furthermore, when considering the application of the developed model in real-world industrial settings, observing sensitivity to Outlier-robust Self-training Segmentation in Computed Tomography image such outlier data is essential. Prior studies have shown that outliers can severely affect model reliability and robustness, prompting extensive research on outlier detection methods and their limitations (22). Such sensitivity becomes especially critical in safe-ty-critical applications, where unexpected model behavior may lead to severe consequences. For instance, even if the overall model performance improves, the occurrence of outlier data, however minor, may necessitate additional post-processing steps for outlier removal despite successful segmentation of objects with high performance. As advancements in deep learning technology continue to improve segmentation accuracy in image segmentation tasks, there is still a demand for performance improvement, particularly in small ROI. Moreover, to minimize the occurrence of outliers, models are often trained with a bias for recall scores to be lower than precision scores. To address these issues, we propose a self-training uncertainty query strategy to suppress the occurrence of outlier data. The objective of this study is to develop an outlier-robust self-training strategy for CT image segmentation by suppressing unreliable pseudo-labels generated during the self-training process. Through the proposed uncertainty-based query strategy, model sensitivity to outlier data is reduced while maintaining segmentation performance, particularly in scenarios involving small regions of interest.

Methods

Figure 1 illustrates our proposed 3D OrSeg framework. This framework adopts an approach of generating labels from unlabeled data through outlier uncertainty and subsequently leveraging the generated data for automatic learning. A 3D OrSeg network based on 3D U-Net was proposed for learning. The network consists of an encoder, two latents, and a decoder layer, and has four output layers obtained from the decoder layer. To accelerate segmentation performance, the RASPP module, which combines Residual with the ASPP module (23), is added. Additionally, by replacing skip-connection, the attention gate shares the features of the encoder layer with the decoder layer.

Figure 1 Proposed self-training framework is introduced, specifically designed to prioritize the avoidance of outliers in the segmentation process, iterating until M≤τ. 3D OrSeg, 3-Dimensional Outlier-robust Segmentation; BFS, breadth first search; Morp, morphology.

Network architecture

Our proposed segmentation network, 3D OrSeg, is a modified architecture based on 3D U-Net. Figure 2 illustrates the overall network structure. The model parameters θ of the learning model F_θare trained with input data DL={(Xi,Yi)}i=1N. The network architecture consists of Convolution, Residual ASPP and Average Pooling layers in the input stage represented by the CRA block, and UpConvolution, Attention Gate, and two Convolution layers in the output stage represented by the UAtC block. D_L passes through four CRA blocks for feature extraction of images. Subsequently, it traverses two latent layers, and the image is gradually restored through four stages of UAtC blocks. The output of each layer of the decoder can be formulated as follows:

pi+1=UAtCi(pi),i=1,2,…,n[1]

Y^i=Conv(pi+1),i=1,2,…,n[2]

Figure 2 Proposed the 3D OrSeg network structure. The RASPP module has been added, which significantly improves model performance. All convolution layers except the last convolution layer are followed by Batch Normalization and Mish, activation function. 3D OrSeg, 3-Dimensional Outlier-robust Segmentation; ASPP, Atrous Spatial Pyramid Pooling; RASPP, Residual Atrous Spatial Pyramid Pooling.

where p is the predicted image from the previous layer and i represents each layer which has n layers. And the output of each block enters respective Convolution layers. As shown in Figure 2, the RASPP module, which can amplify performance, is added to the network. Furthermore, Attention Gate is used instead of skip connection. By utilizing the trained F_θ(D_L), {Y^i}i=14 can be obtained. Ground truth (GT) in D_Land predicted results are of the same size, and the loss function can be formalized through Focal Tversky Loss (24) as follows:

Ltotal=∑i=14Lft(GT,Y^i)[3]

where L_totalis total loss function. Thus, with the last layer, the overall features of the decoder can be learned.

Algorithm to remove outliers

To identify specific pixels of a clavicle object in an image, the breadth-first search (BFS) graph traversal algorithm was employed. Starting from the initial pixel of the clavicle object, BFS was utilized to visit and mark all pixels belonging to the object. The process of finding the initial pixel using the process connected components (PCC) procedure is introduced in Figure 3. This process helps to remove external outliers from the boundary of the object.

Figure 3 Initialization procedure for detecting initial points in 3D image volumes. A four-step algorithm that identifies the z-slice with maximum area, performs connected component analysis to detect objects, identifies the largest object, and selects its centroid as the initial point. 3D, 3-dimensional.

Additionally, a morphological algorithm utilizing the Closing operation was applied to remove internal outliers. The closing operation follows the process of dilation and erosion as follows: y^4⋅f=(y^4⊕f) ⊖f, where f represents a filter image of size 3×3×3, ⊕ denotes the dilation operation and ⊖ denotes the erosion operation. By utilizing the Closing operation and BFS algorithm based on the predicted result y^4, the clavicle object with outliers removed can be obtained.D_L

Pseudo labeling

In Figure 1, our proposed three-step 3D OrSeg framework is presented. In the first step, D_Lis utilized for training, resulting in F_θobtaining the learned parameters θ. The size of the data is denoted as N=|D_L|, M=|D_U|. At each iteration, θ is initialized using the Xavier initialization method (25) as follows:

θijk∼U(−6nin+nout,6nin+nout)[4]

where θ_ijkrepresents an element corresponding to the i-th row, j-th column, and k-th depth. The distribution U(a, b) is truncated, where a is the lower bound and b is the upper bound. In the previous layer, n_in denotes the number of neurons, representing the input size, while n_outdenotes the number of neurons in the current layer, representing the output size. The local minimum problem can be avoided through the initialization. The second step involves inference using the trained F_θ on the unlabeled dataset is denoted as DU={(Ui,0)}i=1M, and F_θ(D_L) yields P^. The third step involves utilizing P^ to generate D_P and updating D_L and D_U. P^′ can be formalized through Morphology algorithm Morp and Breadth-First Search algorithm BFS as follows:

P^′=BFS∘Morp(P^)[5]

Where P^′ represents the result with outliers and noise removed. The extent of removal varies depending on P^. By calculating the difference between P^ and P^′ , the uncertainty of outlier removal can be determined. To compute the extent of outlier removal, L_ftis utilized. Through loss function calculation, L_u= {L₁, …, L_M} is obtained. Subsequently, it can be formalized through ascending order as follows: Lusort={L1,…,LM},(L1≤,…,≤LM). Sorting reveals that the front part of Lusort has lower values, indicating that outliers are relatively less removed. Conversely, the rear part has higher values, implying that outliers are relatively more removed. Our network F requires a dataset with fewer outliers for the training of θ. The front part of Lusort provides a sufficient amount of such data. By selecting a subset τ from all M data, overfitting can be prevented. Finally, automatically masked pseudo-labeled dataset can be obtained as follows: DP={(DUi,Y^i)}i=1τ. The reason for using Y^ as additional data is that if Y^′ is used, the learning process from Y^ to Y^′ does not occur, potentially leading to overfitting. Subsequently, D_Land D_LU are updated as follows: D_L = D_L + D_P, D_U = D_U – D_P. Then, the process restarts from step 1 and iterates until M ≤ τ is satisfied.

Datasets

The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was reviewed by the Institutional Review Board (IRB) of Pusan National University Yangsan Hospital and was granted an exemption from ethical review (IRB No. 05-2022-265), as it was a minimal-risk retrospective study using anonymized imaging data. The requirement for informed consent was waived by the IRB. We collected a total of 420 axial CT datasets at Yangsan Busan University Hospital. These datasets underwent preprocessing, adjusting intensities to the range of [450 to 850 Hounsfield unit (HU)]. Among them, annotation was conducted for 20 datasets, while the remaining 400 datasets were left unannotated. Each labeled dataset comprised volumetric CT scans of size 96×512×512, corresponding to 96 axial slices with an in-plane resolution of 512×512 pixels. To obtain left-side regions of interest, the right half of each volume was cropped, and a horizontal flip was applied so that the anatomical orientation corresponded to the left side. Through this left–right separation and augmentation procedure, the datasets were ultimately converted into tensors with dimensions of 40×96×512×256. Subsequently, for training efficiency, cropping was applied to obtain datasets of size 40×96×288×256. The train dataset is a 60% random dataset, the validation dataset is a 20% random dataset and the test dataset is a 20% random dataset from 40 datasets. The unannotated datasets underwent the same preprocessing steps, resulting in datasets sized 400×96×288×256.

In contrast, the comparative 2D image based model operates on two-dimensional image inputs rather than voxel volumes. To accommodate this architecture, the same volumetric data were decomposed along the depth axis and flattened into a total of 3,840 individual 2D images (3,840×288×256). Each image slice was treated as an independent input sample during training. As a result, inter-slice spatial relationships were not explicitly modeled in the 2D approach, whereas the 3D model was designed to capture structural information at the voxel level.

Implementation details

Our experiments are conducted on an NVIDIA RTX A6000 48G in the PyTorch framework. The AdaFactor (26) optimizer and Checkpoint were used in our model for memory efficiency. The learning speed was increased using GradScaler. An epoch of 60 was sufficient and the learning rate was set to r. The performance of the network can be finely adjusted through α, β = (1–α), and γ in the loss function L_ft. When tested on the clavicle dataset, α=0.5 and γ=1 showed the most effective results. The τ value for pseudo labeling was set to 20. Therefore, a dataset is added to D_L 20 times in total, with τ added per cycle from D_U.

Evaluation

The segmentation performance was quantitatively evaluated using four commonly adopted metrics: Intersection over Union (IoU), Dice similarity coefficient (DSC), Recall, and Precision.

These metrics were computed based on the numbers of true positive (TP), false positive (FP), false negative (FN), and true negative (TN) pixels. IoU measures the overlap between the predicted segmentation and the ground truth. It is defined as:

IoU=TPTP+FP+FN[6]

DSC=2×TP2×TP+FP+FN[7]

Recall=TPTP+FN[8]

Precision=TPTP+FP[9]

These metrics are widely used in medical image segmentation tasks and provide complementary perspectives on model performance. The Hausdorff distance (HD) measures the maximum distance between the boundary points of the predicted segmentation and the ground truth, defined as:

HD(P,G)=max{sup{p∈P}inf{g∈G}d(p,g),sup{g∈G}inf{p∈P}d(g,p)}[10]

where d(⋅,⋅) denotes the Euclidean distance. However, the standard HD is highly sensitive to outliers. Therefore, the 95th percentile Hausdorff distance (95HD) is commonly adopted in medical image analysis. The 95HD computes the distance at the 95th percentile of all boundary point distances, effectively reducing the influence of extreme outliers while still capturing boundary discrepancies. And The model’s training time and size were also evaluated.

For the comparative evaluation, four state-of-the-art segmentation models were employed, including 2D U-Net, 3D U-Net, 3D V-Net, and 2D U²-Net, which are widely used as baseline architectures in medical image segmentation tasks. The architectural characteristics and general principles of 2D U-Net and 2D U²-Net were previously introduced in the Introduction section. In addition, the 3D U-Net was adopted to exploit volumetric contextual information by extending the encoder-decoder structure to three-dimensional convolutions. To ensure a fair and valid comparison, all models were trained and evaluated using the same dataset, identical data preprocessing procedures, and consistent experimental settings. Specifically, the same training, validation, and test splits were applied across all models. Furthermore, identical optimization strategies, learning rate schedules, batch sizes, and stopping criteria were employed, unless otherwise specified. The 3D U-Net and 3D V-Net models were additionally evaluated after incorporating outlier removal algorithms under the same experimental conditions as the proposed method.

Ablation study was conducted to enhance the performance of the model. The first condition pertains to the model size. The Large model has a structure where the channel size doubles as it passes through the encoder layers, starting from 64 channels in the second layer, while the decoder layers have a structure where the channel size halves as it progresses. The input and output layers of the large model have channel sizes of 1. The half depth model has a channel size of 32 in the second layer, with the rest of the structure being identical to that of the full depth model. The second condition involves random sampling from D_L in the proposed self-training method. This condition evaluates the results when random sampling is not performed from the augmented dataset D_Pin D_L. The third condition involves selecting from L_M–τ to L_M based on Lusort in the proposed method. Such selection allows the model to be trained on naive data. This study was conducted based on the 3D U-Net baseline to demonstrate performance under different conditions. Since the full depth with shuffle experiment showed the best performance, the results applying large model, random sampling, and sophisticated selection to our model are presented together. To analyze the robustness and reliability of the proposed framework, statistical analysis was performed based on repeated experimental runs. All ablation experiments were conducted ten times by randomly reshuffling the training data, and segmentation performance was summarized using the mean and standard deviation of the DSC. Variance reduction across repeated runs was used as an indicator of training stability and robustness to stochastic factors inherent in self-training frameworks. This analysis enables a statistical characterization of performance consistency under identical experimental conditions.

In self-training frameworks employing iterative pseudo-labeling, stochastic factors such as random sampling and data reshuffling are known to exert substantial influence on model behavior. To ensure a fair and controlled comparison, these factors were systematically regulated in the ablation study, allowing the contribution of each component—model size, random sampling strategy, and label selection method—to be independently assessed under identical training conditions. Although multi-seed experiments could not be conducted due to the high computational cost associated with 3D volumetric training, all experiments were performed using the same backbone architecture, training schedule, data split, and hyperparameter settings so that uncontrolled variability was minimized.

Results

Comparison study

Figure 4 presents the clavicle segmentation results of the models. As summarized in Table 1, the proposed 3D OrSeg model achieved the highest segmentation performance across most quantitative metrics, including an IoU of 0.939 and a DSC of 0.968, outperforming both 3D U-Net (DSC: 0.962) and 3D V-Net (DSC: 0.957). In particular, improvements were observed in precision and recall, reaching 0.973 and 0.965, respectively, indicating superior boundary delineation and robustness to subtle structural variations near edges. Regarding geometric accuracy, the proposed method achieved a competitive 95HD of 2.795 mm, comparable to other 3D-based models. Although the training time and model size of the proposed approach were larger, these factors are expected to impose minimal practical burden considering recent advancements in hardware capabilities. Overall, the results demonstrate that the proposed method provides improved segmentation accuracy while maintaining acceptable computational efficiency.

Figure 4 Quantitative comparison of semantic segmentation models in different cases. 3D OrSeg shows relatively better performance compared to other state-of-the-art methods. The red box guides the region of interest. 2D, 2-dimensional; 3D OrSeg, 3-Dimensional Outlier-robust Segmentation.

Ablation study

In Figure 5, performance was observed under each condition using the same self-training method. Up to the initial 24 cycles, no clear performance differences were observed across the models. Among the semi-supervised learning configurations without half-depth and data shuffling, the lowest performance was obtained. In contrast, the configuration employing full-depth with data shuffling and the naive strategy exhibited superior performance during the early cycles. As the training cycles progressed, performance improvements were observed in most models. Except for the first cycle, the proposed model consistently demonstrated superior performance across all cycles, achieving a final performance of 96.8% at the 20th cycle. In the case of the 3D U-Net model, the highest performance was achieved when full-depth training with data shuffling was applied.

Figure 5 Ablation study. The performance was evaluated by utilizing the same self- training method while varying the model size and the randomness of the dataset. 3D, 3-dimensional; 3D OrSeg, 3-Dimensional Outlier-robust Segmentation.

The second study evaluated the performance of the dataset. As summarized in Table 2, the effects of dataset expansion, model architecture, and post-processing modules were quantitatively evaluated using 3D U-Net and 3D OrSeg as baselines. The Baseline model and our network were compared using the dataset D_P' generated from the results of the previous ablation study, the full depth with shuffle. Additionally, the impact of Morphology and BFS was primarily reflected in the reduction of boundary errors. For the 3D U-Net trained with D_L, the 95HD value was reduced from 453.25 mm to 305.27 mm after applying these post-processing steps. Similar trends were observed for models trained with D_L + D_P', where consistent reductions in 95HD were achieved while maintaining comparable Dice scores. Notably, the proposed 3D OrSeg model combined with Morphology and BFS achieved the lowest 95HD value of 273.99 mm, demonstrating improved boundary accuracy. The performance improvement of 1.12% reported in the abstract corresponds to the results presented in Table 2, where the proposed model with Morphology and BFS is compared against 3D U-Net with D_L and without Morphology and BFS. Similarly, the 0.384% value refers to the standard deviation of the DSC for the same comparison.

Discussion

The OrSeg framework was employed as a self-training method capable of generating a substantial amount of data from a limited initial dataset. Although only 20 samples were provided at the outset, a total of 400 additional labeled instances were successfully produced by the framework. Under this experimental setting, a performance gap of 12.1% in DSC was observed between the 2D baseline model (0.847) and the proposed model (0.968), as shown in Table 1. This difference can be attributed to the fundamental discrepancy between learning from 3,840 independent two-dimensional image slices and learning from 40 three-dimensional voxel volumes. In the proposed approach, depth information is processed simultaneously using 3×3×3 convolutional filters, resulting in learning dynamics that are inherently different from those of conventional 2D image-based training. Such volumetric learning is therefore considered particularly effective in scenarios where only a small number of annotated samples are available.

Although pseudo-labeling inherently carries the risk of error propagation, the performance trend shown in Figure 5 suggests that pseudo-labels generated in the early stages are more likely to contain errors. Nevertheless, most of these incorrect labels are effectively filtered out by the outlier-removal algorithm, thereby minimizing the likelihood of cumulative error propagation as the training cycles are repeated.

Despite these advantages, a limitation of the proposed framework lies in the absence of explicit interpretability mechanisms for validating the uncertainty-based pseudo-label selection process. While confidence estimation was employed to filter unreliable predictions, recent studies on feature-preserving segmentation networks (27) have emphasized the importance of understanding feature propagation within encoder-decoder architectures, particularly those incorporating attention gates. Furthermore, advances in explainable artificial intelligence (XAI) for clinical decision support suggest that uncertainty measures alone may be insufficient to explain why certain regions are deemed reliable or unreliable. Accordingly, future work will incorporate XAI techniques, such as attribution maps and uncertainty heatmaps, to visualize unstable prediction regions and to more rigorously assess the credibility of pseudo-labels.

From a practical perspective, the proposed framework offers several notable advantages. Owing to its model-agnostic design, it can be readily integrated into a wide range of segmentation architectures and enables efficient exploitation of large-scale unlabeled datasets. The method demonstrated favorable performance in the segmentation of small regions of interest and anatomically thin structures, where prediction accuracy is particularly sensitive to anomalous variations. This characteristic enhances its applicability in clinically challenging scenarios.

Although the training process may appear computationally demanding, this limitation must be considered in the context of the substantial cost associated with manual annotation in medical imaging. In practice, annotating a single case requires approximately three hours of expert labor; consequently, constructing a dataset of 400 fully annotated cases would require more than two months of continuous effort. In contrast, large-scale pseudo-labeled datasets can be automatically generated using the proposed framework, significantly reducing human workload and offering considerable economic benefits. From a clinical standpoint, this capability is particularly valuable in medical institutions facing shortages of radiology staff or rapidly increasing imaging demands.

While this study focused on clavicle segmentation, the framework is readily extensible to other anatomies characterized by complex geometries or slender structures, such as the scapula, wrist bones, and small joint components. By simultaneously reducing annotation burden and improving data reliability, the proposed approach provides a scalable and practical pathway toward the development of robust and clinically deployable medical image segmentation models.

Conclusions

In this study, a semi-supervised learning framework based on OrSeg was proposed to automatically transform unlabeled data into pseudo-labeled data in a progressive manner. Given that unlabeled datasets inherently contain noisy and unreliable samples, an uncertainty-aware filtering strategy integrating morphological operations and a BFS algorithm was introduced to regulate the self-training process. Rather than explicitly enforcing robustness by exposing the model to corrupted labels, the proposed framework was designed to stabilize pseudo-label generation by selectively excluding highly uncertain samples during iterative training. Through this mechanism, the self-training process was gradually guided by increasingly reliable pseudo-labeled data, resulting in improved learning stability and enhanced segmentation performance. Experimental results demonstrated that the proposed approach effectively mitigates the adverse impact of outliers in semi-supervised learning scenarios.

Furthermore, it was observed that uncertainty estimation operates more effectively within semantically coherent contexts than in naïve or highly noisy environments. This finding suggests that maintaining a stable semantic structure is critical for reliable outlier filtering and for preventing error accumulation during self-training. Nevertheless, the proposed approach is subject to certain limitations, as it is primarily tailored to semantic-level segmentation tasks and may not generalize directly to instance-level segmentation problems that require explicit object separation.

Acknowledgments

None.

Footnote

Data Sharing Statement: Available at https://jmai.amegroups.com/article/view/10.21037/jmai-2025-217/dss

Peer Review File: Available at https://jmai.amegroups.com/article/view/10.21037/jmai-2025-217/prf

Funding: This study was conducted with the support of the Ministry of SMEs and Startups’ Smart Manufacturing Innovation Technology Development Project (SE230035).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://jmai.amegroups.com/article/view/10.21037/jmai-2025-217/coif). This study was conducted with the support of the Ministry of SMEs and Startups’ Smart Manufacturing Innovation Technology Development Project (SE230035). The authors have no other conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. This study was reviewed by the Institutional Review Board (IRB) of Pusan National University Yangsan Hospital and was granted an exemption from ethical review (IRB No. 05-2022-265), as it was a minimal-risk retrospective study using anonymized imaging data. The requirement for informed consent was waived by the IRB.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Ronneberger O, Fischer P, Brox T. U-Net: convolutional networks for biomedical image segmentation. In: Navab N, Hornegger J, Wells WM, Frangi AF, editors. Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015. Lecture Notes in Computer Science. Vol 9351. Cham: Springer; 2015. p. 234-41.
2. Çiçek Ö, Abdulkadir A, Lienkamp SS, et al. 3D U-Net: learning dense volumetric segmentation from sparse annotation. In: Ourselin S, Joskowicz L, Sabuncu MR, Unal G, Wells W, editors. Medical Image Computing and Computer-Assisted Intervention – MICCAI 2016. Lecture Notes in Computer Science. Cham: Springer; 2016. p. 424-32.
3. Qin X, Zhang Z, Huang C, et al. U2-Net: going deeper with nested U-structure for salient object detection. Pattern Recognit 2020;106:107404.
4. Zhou Z, Siddiquee MMR, Tajbakhsh N, et al. UNet++: A Nested U-Net Architecture for Medical Image Segmentation. Deep Learn Med Image Anal Multimodal Learn Clin Decis Support 2018;2018:3-11. [Crossref] [PubMed]
5. Liu H, Shen X, Shang F, et al. CU-Net: cascaded U-Net with loss weighted sampling for brain tumor segmentation. In: Zhu D, Yan J, Huang H, et al., editors. Multimodal Brain Image Analysis and Mathematical Foundations of Computational Anatomy. Cham: Springer; 2019. p. 102-11.
6. Qiu X. U-Net-ASPP: U-Net based on atrous spatial pyramid pooling model for medical image segmentation in COVID-19. J Appl Sci Eng 2022;25:1167-76.
7. Milletari F, Navab N, Ahmadi SA. V-Net: fully convolutional neural networks for volumetric medical image segmentation. In: Proceedings of the Fourth International Conference on 3D Vision (3DV). Piscataway (NJ): IEEE; 2016. p. 565-71.
8. Wakamatsu Y, Kamiya N, Zhou X, et al. Automatic segmentation of supraspinatus muscle via bone-based localization in torso computed tomography images using U-Net. IEEE Access 2021;9:155555-63.
9. Liu P, Han H, Du Y, et al. Deep learning to segment pelvic bones: large-scale CT datasets and baseline models. Int J Comput Assist Radiol Surg 2021;16:749-56. [Crossref] [PubMed]
10. González Sánchez JC, Magnusson M, Sandborg M, et al. Segmentation of bones in medical dual-energy computed tomography volumes using the 3D U-Net. Phys Med 2020;69:241-7. [Crossref] [PubMed]
11. Möller M, Wolf O, Bergdahl C, et al. The Swedish Fracture Register—ten years of experience and 600,000 fractures collected in a national quality register. BMC Musculoskelet Disord 2022;23:141. [Crossref] [PubMed]
12. Robinson CM. Fractures of the clavicle in the adult: epidemiology and classification. J Bone Joint Surg Br 1998;80:476-84. [Crossref] [PubMed]
13. Friedli L, Kloukos D, Kanavakis G, et al. The effect of threshold level on bone segmentation of cranial base structures from CT and CBCT images. Sci Rep 2020;10:7361. [Crossref] [PubMed]
14. Deng Y, Wang L, Zhao C, et al. A deep learning-based approach to automatic proximal femur segmentation in quantitative CT images. Med Biol Eng Comput 2022;60:1417-29. [Crossref] [PubMed]
15. Lim HK, Jung SK, Kim SH, et al. Deep semi-supervised learning for automatic segmentation of inferior alveolar nerve using a convolutional neural network. BMC Oral Health 2021;21:630. [Crossref] [PubMed]
16. Hogeweg L, Sánchez CI, de Jong PA, et al. Clavicle segmentation in chest radiographs. Med Image Anal 2012;16:1490-502. [Crossref] [PubMed]
17. Yoo D, Kweon IS. Learning loss for active learning. In: 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR); Long Beach, CA, USA, 2019, pp. 93-102.
18. Wu X, Chen C, Zhong M, et al. COVID-AL: The diagnosis of COVID-19 with deep active learning. Med Image Anal 2021;68:101913. [Crossref] [PubMed]
19. Han K, Sheng VS, Song Y, et al. Deep semi-supervised learning for medical image segmentation: a review. Expert Syst Appl 2024;245:123052.
20. Kim JN, Kim J, Dharejo FA, et al. Lightweight Mamba model for 3D tumor segmentation in automated breast ultrasounds. Mathematics 2025;13:2553.
21. Karamcheti S, Krishna R, Fei-Fei L, et al. Mind your outliers! Investigating the negative impact of outliers on active learning for visual question answering. In: Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing; 2021 Aug. p. 7265-81.
22. Smiti A. A critical overview of outlier detection methods. Comput Sci Rev 2020;38:100306.
23. Chen LC, Papandreou G, Kokkinos I, et al. DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs. IEEE Trans Pattern Anal Mach Intell 2018;40:834-48. [Crossref] [PubMed]
24. Abraham N, Khan NM. A novel focal Tversky loss function with improved attention U-Net for lesion segmentation. In: Proceedings of the IEEE 16th International Symposium on Biomedical Imaging (ISBI 2019); 2019. p. 683-7.
25. Glorot X, Bengio Y. Understanding the difficulty of training deep feedforward neural networks. In: Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics; 2010 May 13–15; Chia Laguna Resort, Sardinia, Italy. Proceedings of Machine Learning Research. 2010;9:249-56.
26. Shazeer N, Stern M. Adafactor: Adaptive learning rates with sublinear memory cost. In: Proceedings of the 35th International Conference on Machine Learning; 2018 Jul 10–15; Stockholm, Sweden. Proceedings of Machine Learning Research. 2018;80:4596-604.
27. Abidin ZU, Naqvi RA, Kim HS, et al. Optimizing optic cup and optic disc delineation: Introducing the efficient feature preservation segmentation network. Eng Appl Artif Intell 2025;144:110038.