A paradigm of hybrid-supervision for annotation-scarce periapical film analysis

Abstract

Background

Accurate segmentation of anatomical and pathological structures in periapical radiographs is essential for digital dentistry, yet obtaining sufficient manual annotations remains a major challenge. This study aims to develop a hybrid self-supervised and semi-supervised learning framework to address the data annotation bottleneck in multi-class structure segmentation of dental periapical films.

Method

We propose a two-stage approach combining self-supervised learning with semi-supervised fine-tuning. First, we employ an Intensity-Gradient-Local Contrast based Masked Autoencoder (IGLC-MAE) for self-supervised learning on 74,292 unlabeled periapical films, utilizing structured adaptive masking specifically designed for dental radiography characteristics. The pre-trained model then generates pseudo-labels for 6,259 unlabeled images, which are combined with 229 manual annotations for semi-supervised fine-tuning using Mask2Former. To optimize this process, we systematically evaluated loss weight configurations and introduced an adaptive weighting mechanism, which together improved the quality of pseudo-labels and further enhanced segmentation performance.

Results

Compared to traditional supervised learning methods, our self-supervised learning approach achieved significant improvements in oral structure segmentation, with a Dice score of 73.17%, surpassing the best supervised learning configuration by 10.52%. Subsequently, the semi-supervised learning strategy with pseudo-labels further enhanced performance, reaching the highest Dice score of 74.12% at an optimal weight ratio of 0.85:0.15 for the manual-to-pseudo-label loss. The adaptive semi-supervised strategy delivered an additional 1.23% improvement in Dice score by effectively suppressing low-confidence pseudo-label noise.

Conclusion

The proposed hybrid self-supervised and semi-supervised framework. effectively addresses the challenge of annotation data scarcity and provides a new technical approach for dental image analysis. Our method achieves superior segmentation performance in periapical films while minimizing dependency on manual annotations, offering a clinically viable solution for medical imaging.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12903-026-09157-x.

Introduction

Periapical radiographs serve as a fundamental imaging tool in dental practice for evaluating complex anatomical structures and pathological changes in teeth and the periapical regions [1]. In recent years, deep neural networks have demonstrated remarkable progress across broad domains [2, 3]. They have been particularly successful in dental image analysis tasks, including tooth structure segmentation [4, 5], caries diagnosis [6, 7], and periapical lesion detection [8]. However, the success of such methods heavily relies on the availability of large-scale, high-quality annotated datasets [9]. Currently, publicly available high-quality dental datasets remain exceedingly sparse, notable in dentistry [10]. Researchers often need to establish their own datasets, while professional annotation is expensive, laborious, and requires skillful domain expertise. Pursuing large-scale annotated datasets entails substantial time and financial investments. Moreover, insufficient labeled data often leads to underfitting and poor generalization in supervised learning models. This dilemma makes annotation scarcity a critical bottleneck hindering the application of deep learning technologies in dentistry.

To reduce the dependency on annotated data, self-supervised and semi-supervised learning approaches have gained increasing attention. These methods leverage readily available unlabeled data to facilitate effective learning under limited supervision. Promising results have been achieved in both computer vision [11–14] and medical image analysis [15, 16]. Self-supervised learning designs pretext tasks, such as masked image reconstruction [17] or contrastive learning [18]. These tasks enable models to learn transferable visual representations from large amounts of unlabeled images. Consequently, reliance on manual labeling is substantially reduced. Among these, masked autoencoding (MAE), proposed by Kaiming He in 2021 [19], which involves randomly masking image regions and reconstructing missing pixels, has proven effective in learning robust representations and has been validated in multiple medical imaging applications [20, 21]. To better adapt MAE to the characteristics of periapical radiographs, this study introduces a structured patch-wise adaptive masking strategy based on intensity, gradient, and local contrast (IGLC-MAE), optimizing the standard MAE framework for dental X-ray imaging. IGLC-MAE is novel in its domain-adaptive masking strategy rather than merely altering the MAE reconstruction objective. Unlike standard MAE with random masking [19] or recent variants such as Scale-MAE [17] and Hi-End-MAE [20] that focus on scale-aware or hierarchical design, IGLC-MAE leverages three radiographic priors—intensity, gradient magnitude, and local contras—to gauge the diagnostic informativeness of each patch. In periapical radiographs, these cues correspond to clinically relevant structures such as enamel-dentin boundaries, pulp contours, restorative margins, carious defects, and periapical radiolucencies. By masking these high-information regions, the reconstruction task becomes more challenging and clinically meaningful. The encoder is therefore driven to learn representations that are more transferable to downstream fine-grained segmentation.

However, even with self-supervised pre-training, models may still suffer from insufficient representation transfer when labeled data are extremely limited, ultimately compromising performance. Pseudo-labeling is a popular semi-supervised technique. It assigns high-confidence predictions to unlabeled samples to augment the training set. This strategy has proven effective for improving model generalization [22, 23]. Nonetheless, the effectiveness of pseudo-labeling is highly dependent on the quality of the initial model [24]. Under severe annotation scarcity, prediction biases from a poorly-initialized model can be amplified over iterative training, leading to performance degradation.

To address these limitations, a novel paradigm that integrates self-supervised pre-training with semi-supervised fine-tuning has recently emerged [15, 25, 26]. This approach uses self-supervised learning as a robust feature initialization step, followed by pseudo-labeling for task-specific adaptation, enabling more stable performance under limited annotation. Although this combined strategy has shown potential in general computer vision, its systematic exploration and validation in the context of fine-grained periapical film segmentation remain unexplored.

To tackle the challenge of limited annotations in dental radiography, this study introduces a hybrid learning framework that systematically combines self-supervised pre-training with adaptive semi-supervised fine-tuning. Our approach begins with domain-adaptive pre-training via an enhanced Masked Autoencoder on extensive unlabeled periapical films, yielding a feature-rich foundational model. This model then produces pseudo-labels for unlabeled images, which are merged with scarce manual annotations under a confidence-aware re-weighting mechanism during fine-tuning. We conducted an extensive evaluation on a multi-class segmentation task using periapical films. The task encompasses seven critical structures: tooth, pulp, crown, fillings, root canal fillings, caries, and periapical lesions. Results show that the proposed framework attains markedly superior segmentation accuracy even with limited labeled data. This work offers a systematic and practical solution to address annotation scarcity in periapical films analysis. It paves a viable way for deploying accurate AI-assisted diagnostic tools in resource-constrained dental clinical settings.

Methods

This study was approved by the Ethics Committee of Peking University School of Stomatology (PKUSSIRB-2025107012). All periapical films included in the study were taken during routine diagnosis and treatment at the Department of the Fourth Clinical Division, Peking University School and Hospital of Stomatology between 2012 and 2024, with patient informed consent for research purposes. The research process strictly adhered to medical ethical guidelines, ensuring patient privacy and data security.

Dataset

This study utilized a dataset comprising 74,292 unlabeled periapical radiographs and 285 annotated images, all of which were previously introduced in a previous study [27]. All images were acquired using Planmeca (Helsinki, Finland) periapical X-ray machines operating at 220–240 V, 60 kVp, 2 mA, with an exposure time of 0.05 s. Each image contained up to 3–4 teeth, representing clinically routine periapical films.

To prevent information leakage, the unlabeled pre-training set (74,292 periapical radiographs) and the labeled downstream set (285 images) were strictly separated at the patient level with non-overlapping cohorts. No unlabeled images from downstream test patients were included in self-supervised pre-training.

1. unlabeled dataset

   The unlabeled dataset consisted of 6,259 high-resolution images (612 × 474 pixels) and 68,033 thumbnails (120 × 100 pixels). Due to limitations in the database system when exporting high-resolution images in batch, most images could only be obtained as thumbnails. This multi-resolution data combination allowed us to explore the effectiveness of multi-scale self-supervised learning strategies.

2. labeled dataset

Consistent with the objective of this study, establishing and validating an annotation-efficient paradigm for the structures most prevalent in routine periapical films, the labeled set was deliberately confined to routine films. For the annotated data, during the data screening process, we excluded samples with poor image quality, obvious artifacts, and those containing non-routine dental structures (such as implants, bridges, supernumerary teeth, deciduous teeth, residual roots, and developmental anomalies). Poor-quality images were excluded because artifacts and projection distortions obscure the radiographic boundaries required for pixel-wise annotation. Complex restorations (implants, fixed bridges) were excluded because they lack natural tooth anatomy or span multiple units, making them incompatible with our single-tooth-based seven-class scheme. Similarly, teeth with abnormal morphology, atypical development, or physiological root resorption (supernumerary, deciduous, residual roots, developmental anomalies) deviate from the normal permanent dentition assumptions underlying our annotation protocol. The descriptive statistics of the annotated dataset include a mean age of 35.6 years (SD: 15.2, range: 16—80), with a gender distribution of 51% male and 49% female.

A rigorous pixel-level annotation protocol was applied, including dental structures (tooth, pulp), restorative materials (crown, fillings, root canal fillings), and common pathologies (caries, periapical lesions). Annotation was performed by a dental specialist with over 10 years of clinical experience and verified by a second equally qualified expert to ensure accuracy and consistency. The labeling process was conducted using LabelMe, an open-source tool also adopted in our prior work. All annotators underwent professional training and calibration to standardize segmentation procedures. To evaluate inter-annotator agreement, three experienced dentists independently annotated 20 randomly selected radiographs, achieving a Fleiss’ Kappa of 0.948.

The dataset was ultimately divided into training and test sets (in an 8:2 ratio), comprising 229 training images and 56 test images. To evaluate the performance of different self-supervised training strategies, all comparative experiments were conducted under identical conditions. The downstream segmentation model was configured with 7 categories (tooth, pulp, crown, fillings, root canal fillings, caries, and periapical lesions). A per-class summary of the annotated dataset is provided in Supplementary Table A.1.

Study design

Stage 1: self-supervised pre-training

This study adopted the self-supervised learning paradigm of Masked Autoencoder (MAE) to address problems in oral X-ray image analysis. The core principle of MAE is to learn the intrinsic representation of images by reconstructing randomly masked image regions, allowing training on large amounts of unlabeled data without manual annotation. We designed an improved algorithm called structured block-based adaptive mask MAE, specifically optimized for the characteristics of oral X-rays. Unlike traditional MAE random masking strategies, we proposed an adaptive masking mechanism based on Intensity, Gradient, and Local Contrast (IGLC-MAE). As shown in Fig. 1(a), this mechanism first calculates a comprehensive score for each image patch based on three image features, with the scoring function expressed as:

where S(p) represents the comprehensive score of image patch p, , , and represent the normalized intensity variation, gradient magnitude, and local contrast values, respectively, and 、, and are the corresponding weight coefficients. Subsequently, the algorithm implements a structured masking strategy based on patch scores—dividing the original image into a regular macro-block grid, each containing multiple adjacent basic image patches. The visual comparison between random MAE masking and the proposed IGLC-MAE masking strategy is shown in Fig. 2. Additional masking comparisons across diverse periapical cases are provided in Supplementary Fig. A.1. Random MAE masking scatters masked patches across the image without considering anatomical structure, whereas IGLC-MAE preferentially masks high-information regions, such as the main tooth region, root canal-related structures, and adjacent diagnostically relevant areas.

Fig. 1.

Overview of the proposed framework. Stage 1: self-supervised pre-training. The system calculates a combined score S(p) for each image patch based on Intensity (I), Gradient (G), and Local Contrast (C). High-score macro-blocks are masked, and a ViT encoder–decoder reconstructs the image with a contrastive objective to learn domain-adaptive features. Stage 2: Semi-supervised fine-tuning. The pre-trained ViT Encoder weights are transferred to the Mask2Former segmentation framework for multi-structure fine-tuning

Fig. 2.

Visual explanation of the IGLC-MAE masking strategy

The rationale for prioritizing high IGLC-score macro-blocks is that these regions carry more task-relevant structural uncertainty than homogeneous background or uniformly radiopaque areas. Masking such regions forces the encoder to infer missing boundaries and tissue transitions from context, promoting learning of local morphology, spatial continuity, and radiographic contrast. The structured macro-block design prevents scattered isolated patches, instead creating clinically meaningful missing regions that align the pretext task more closely with downstream delineation of connected oral structures and lesions. The macro-block score is determined by the average score of the internal basic blocks, and the system prioritizes masking high-scoring macro-blocks. This macro-block structured masking strategy is superior to random masking because it preserves the local spatial continuity of oral tissues, conforming to the morphological characteristics of structures such as teeth and periodontal tissues. The network architecture adopts a Vision Transformer encoder-decoder architecture, where the encoder processes unmasked regions and extracts feature representations, and the decoder reconstructs the masked regions based on these features. This self-supervised learning method enables the model to extract morphological features and spatial relationship representations of key oral structures from unlabeled data, thereby obtaining a domain-adaptive feature extractor. After completing self-supervised pre-training, the model was transferred to the oral structure segmentation task. As shown in Fig. 1 (b), the study adopted the Mask2Former framework to implement end-to-end segmentation, which transforms the segmentation problem into query-based mask prediction. The downstream task was configured to identify 7 classes of oral structures and lesions: tooth, pulp, crown, fillings, root canal fillings, caries, and periapical lesions. Training used pre-trained weights for initialization and fine-tuning on 229 annotated images. Mask2Former accurately delineates complex oral structure boundaries through multi-scale feature fusion and pixel-level mask prediction, providing technical support for computer-assisted diagnosis.

Stage 2: Semi-supervised fine-tuning

The overall pipeline of the adaptive semi-supervised fine-tuning stage is illustrated in Fig. 3. This stage contains two complementary components: a static pseudo-labeling baseline and an online EMA teacher–student strategy with confidence-adaptive pseudo-label weighting.

Fig. 3.

Pipeline of the adaptive semi-supervised learning framework with confidence-weighted pseudo-labeling

Static pseudo-labeling baseline: After self-supervised pre-training, the IGLC-MAE pre-trained backbone is transferred to the Mask2Former segmentation framework and fine-tuned on the 229 manually annotated periapical radiographs to obtain an initial segmentation model. This model is then used to perform one-time inference on the 6,259 unlabeled high-resolution images, generating fixed pseudo-labels. We jointly train the segmentation model with the manually annotated images and the pseudo-labeled images by adjusting the loss weight ratio .

Adaptive pseudo-labeling with online EMA Teacher: To improve pseudo-label stability, we adopt a teacher–student paradigm. The student model is optimized by gradient descent, while the teacher model is updated as an exponential moving average (EMA) of the student parameters:

where and denote the teacher and student parameters, respectively, and α is the EMA decay. At each training iteration, the current EMA teacher re-infers the unlabeled images to generate dynamic pseudo-labels and confidence maps, ensuring that pseudo-label quality improves progressively as the student refines and avoiding noise accumulation from fixed labels.

Confidence-guided adaptive pseudo-label weighting: Given the fuzzy boundaries and low-contrast structures in periapical radiographs, pseudo-labels may contain noise. Therefore, instead of treating all pseudo-label pixels equally, we introduce a confidence-adaptive weighting mechanism. For each unlabeled image, the teacher predicts a pixel-wise confidence map . The pseudo-label loss contribution is modulated by a Sigmoid-based weight:

where is the confidence center threshold and τ controls the smoothness of the weighting transition. All pseudo-label pixels participate in training with continuously modulated weights; no hard thresholding is applied, preserving boundary information from medium-confidence regions that would otherwise be discarded. Pixels with low confidence are softly down-weighted, while high-confidence regions receive stronger supervision. This avoids abrupt hard filtering and yields more stable optimization.

Semi-supervised Optimization Objective: The total training objective combines supervised loss on labeled images and adaptively weighted pseudo-label loss on unlabeled images:

where is computed using manual annotations (including Mask2Former’s classification and mask losses), is computed using the teacher-generated pseudo-labels, and , are balancing coefficients. We systematically explored different manual-to-pseudo loss ratios and adopted the configuration yielding the best validation performance.

Through this stage, the model continuously absorbs reliable supervision from unlabeled data while suppressing noisy pseudo-label regions, leading to improved generalization in multi-class dental structure segmentation.

Experiment

To enhance model generalization capability, we implemented data augmentation strategies including random cropping, horizontal flipping, and standardization. All oral X-ray images were processed to a uniform size.

In stage one, the self-supervised pre-training phase used a dataset containing 74,292 unlabeled oral X-ray images. All images were resized to 224 × 224 pixels before being fed into the network to ensure a unified input resolution and patch scale. We adopted a ViT-Base (patch size 16 × 16) structure, initialized with ImageNet pre-trained weights, and configured with a structured macro-block size of 2 × 2 and a masking rate of 0.75. For the IGLC-MAE adaptive masking mechanism, we set the intensity weight to 0.15, the gradient magnitude weight to 0.35, and the local contrast weight to 0.5 to optimize the mask distribution for oral tissue structures. Training used an AdamW optimizer with a base learning rate of 1.5 × 10⁻⁴, weight decay coefficient of 0.05, batch size of 128, for 500 epochs, while applying pixel-level normalization loss function to ensure training stability.

The downstream segmentation task fine-tuning phase was based on the Mask2Former framework, with a learning rate of 2 × 10⁻⁴ and maximum iterations of 50,000. The model was configured to recognize seven oral categories (tooth, pulp, crown, fillings, root canal fillings, caries, and periapical lesions). We used 285 images annotated by professional dentists. To ensure statistical reliability under limited labeled data, we adopted a Monte Carlo cross-validation strategy with five independent patient-level random train–test splits. In each split, 229 annotated images were used for training, 56 annotated images were used for testing, with no patient overlap between the training and test sets. The final results were reported as the mean ± standard deviation across the five independent splits. Because the unlabeled and annotated image pools were separated at the patient level before model training, no downstream test patients appeared in the unlabeled pre-training pool or in the pseudo-label generation process. All comparative experiments were conducted under this consistent evaluation protocol.

In stage two, we conducted semi-supervised fine-tuning for multi-class periapical structure segmentation with an online adaptive pseudo-label weighting strategy. Specifically, 229 labeled images () were used, while 6,259 unlabeled images () were assigned initial pseudo-labels generated by the stage-one IGLC-MAE pre-trained model. A MixedSampler was employed to construct each mini-batch with a balanced labeled-to-unlabeled ratio of :=1:1. Both teacher () and student () networks followed the Mask2Former architecture initialized from stage-one pre-trained weights, which was updated by standard backpropagation and by EMA with a decay rate α = 0.999. To determine the adaptive weighting parameters, the labeled dataset (=229) was randomly split into 90% training and 10% validation. We performed grid search over T_c ∈ {0.7, 0.8, 0.9} and τ ∈ {0.05, 0.10, 0.20}, and selected the configuration maximizing validation Dice. To suppress noisy pseudo-labels, we applied a Sigmoid-based confidence-adaptive weighting function with Tc = 0.8 and τ = 0.10 to modulate the pseudo-label loss, and set the global loss weight = 1.0 for balancing supervised and weighted pseudo-supervised objectives.

The experimental environment for this study was built on the Ubuntu 18.04 operating system, paired with CUDA 11.7 and using the PyTorch 2.0.0 deep learning framework. Python was used as the primary programming language for model development and training. For computing hardware, we used an NVIDIA RTX 3090 GPU with 24 GB of video memory for model training and evaluation.

Model evaluation and statistical analysis

We evaluated segmentation performance using pixel-level Precision, Recall, and Dice coefficient. For each class, pixels were categorized as true positive (TP, correctly predicted as the target class), true negative (TN, correctly predicted as non-target), false positive (FP, incorrectly predicted as the target class), and false negative (FN, incorrectly predicted as non-target). Precision, Recall, Dice, and IoU were calculated as , , , and , respectively. Precision reflects the accuracy of predictions, representing the proportion of actual positive examples among those predicted as positive; Recall reflects the completeness of predictions, representing the proportion of correctly identified positive examples among all actual positive examples; the Dice coefficient and mIoU evaluates class-specific segmentation quality by calculating the overlap ratio between the predicted region and the ground truth. We calculated these metrics separately for each category (tooth, pulp, crown, fillings, root canal fillings, caries, and periapical lesions) to evaluate the model's performance differences in recognizing different oral structures and lesions.

To ensure statistical reliability and robust generalization under the limited labeled dataset (=229), we adopted a patient-level Monte Carlo repeated random sub-sampling protocol. Specifically, five independent random 8:2 splits were performed at the patient level on the 285 annotated images, yielding 229 training images and 56 test images. For each split, the model was trained from scratch—including IGLC-MAE initialization and adaptive semi-supervised fine-tuning on the 229 training images—and evaluated on the corresponding 56 test images. The final Precision, Recall, and Dice were reported as the mean and standard deviation across the five independent test sets, reflecting the model's stability across different patient compositions. Test-set patients in all splits were strictly excluded from the 74,292 unlabeled images used for IGLC-MAE pre-training and from the 6,259 unlabeled images used for pseudo-label generation.

Results

Comparison of learning paradigms and feature initialization

This study systematically evaluated the performance of different learning paradigms and model configurations on dental periapical film segmentation tasks, with detailed comparative results shown in Table 1. In the supervised learning paradigm, traditional convolutional neural network models using ImageNet pre-trained ResNet50 backbone networks (including FPN[28], LinkNet[29], U-Net[30], and DeepLabV3 + [31]) achieved Dice scores ranging from 33.53% to 41.25%. In contrast, under the same supervised learning paradigm, using the Mask2Former architecture with ViT-base backbone and ImageNet pre-trained weights demonstrated significantly better performance, with a Dice score of 62.75%. This confirms the potential of Transformer-based architectures in this task.

Table 1.

Performance comparison of different backbone configurations on periapical film

Learning Paradigm	Model Architecture	Backbone	Backbone Training	Metrics
				Sensitivity (%)	Precision (%)	mDice (%)	mIoU (%)
Supervised Learning	Linknet	ResNet50	ImageNet Pre-trained	28.52 ± 3.10	43.29 ± 2.59	33.53 ± 2.97	23.78 ± 1.17
	FPN			32.89 ± 2.80	50.32 ± 3.04	38.34 ± 2.37	28.01 ± 1.10
	DeepLabV3 +			32.61 ± 3.50	48.84 ± 3.28	39.76 ± 2.18	28.78 ± 1.03
	Unet			37.19 ± 3.00	46.10 ± 2.80	41.25 ± 1.34	28.72 ± 1.03
	Mask2Former	ViT-base		55.19 ± 1.60	73.70 ± 1.25	62.75 ± 1.41	49.16 ± 1.18
Self-supervised Learning	Mask2Former	ViT-base	ImageNet Pre-trained	64.64 ± 1.30	77.84 ± 1.05	67.94 ± 1.15	54.38 ± 1.09
			BEiTv3 (ImageNet → Dental)	60.43 ± 2.47	74.57 ± 3.20	65.17 ± 1.77	52.36 ± 1.29
			MoCov3 (ImageNet → Dental)	61.27 ± 2.49	77.18 ± 1.48	65.92 ± 1.68	53.17 ± 1.27
			MAE (ImageNet → Dental)	68.44 ± 1.20	79.33 ± 0.95	72.15 ± 1.05	59.82 ± 1.09
			IGLC-MAE (ImageNet → Dental)	69.48 ± 1.10	79.78 ± 0.85	73.17 ± 0.95	60.91 ± 1.03
			DINOv2 (LVD-142M → Dental)	71.64 ± 2.15	81.42 ± 2.35	74.67 ± 1.45	62.82 ± 1.23
Hybrid-supervised Learning (ours)	Mask2Former	ViT-base	Static Pseudo-labeling	70.50 ± 1.00	78.50 ± 0.90	74.12 ± 0.90	62.03 ± 0.93
			Adaptive Weighting	72.50 ± 0.80	78.50 ± 0.75	75.35 ± 0.70	63.76 ± 0.82

To further validate the impact of different self-supervised pre-training strategies on fine-grained periapical film segmentation, we systematically compared ImageNet direct fine-tuning, contrastive learning-based BEiTv3 and MoCov3, reconstruction-based standard MAE, discriminative pre-training DINOv2, and our proposed IGLC-MAE under the identical Mask2Former architecture. As shown in Table 1, ImageNet direct fine-tuning achieved a Dice of 67.94%, whereas BEiTv3 (65.17%) and MoCov3 (65.92%) exhibited limited gains, suggesting a domain gap between generic visual contrastive objectives and fine-grained dental segmentation. Standard MAE significantly improved performance to 72.15% through masked reconstruction pre-training, confirming the suitability of generative pre-training for medical imaging. Notably, DINOv2 achieved a strong result of 74.67%, benefiting from large-scale LVD-142 M pre-training; however, this relies on external generic datasets. In contrast, our IGLC-MAE reached 73.17% using only the 74,292 in-domain periapical films collected in this study, demonstrating comparable performance under purely domain-specific data constraints while providing a more stable feature foundation for subsequent semi-supervised optimization. This result improved by 1.02 percentage points over standard MAE and by 10.42 percentage points over the best supervised learning configuration. Regarding other key metrics, IGLC-MAE achieved a sensitivity of 69.48% and precision of 79.78%, substantially exceeding the 55.19% sensitivity and 73.70% precision of the best supervised approach, comprehensively demonstrating the advantages of the proposed self-supervised learning strategy combined with an advanced model architecture.

To intuitively illustrate the category-wise differences among pre-training strategies, Fig. 4 reports per-class Dice of representative methods grouped by tooth structures, restorative materials, and lesion classes. Both IGLC-MAE and our final adaptive semi-supervised method outperform the fully supervised Mask2Former and the best CNN (U-Net) across all seven categories, with particularly pronounced gains in fine-grained or low-contrast classes such as pulp, fillings, caries, and periapical lesions, indicating that the improvements from self-supervised pre-training are consistent across anatomical categories and especially beneficial for boundary-sensitive structures. The lesion-related categories remained the most challenging due to severe class imbalance and spatial sparsity: caries and periapical lesions accounted for only 1.16% and 0.88% of annotated foreground pixels, respectively. However, they appeared in 36.1% and 26.7% of images. Nevertheless, the proposed adaptive semi-supervised method significantly improved lesion-class performance compared with the self-supervised Mask2Former baseline, with gains of + 12.29% for caries (p = 0.001) and + 3.88% for periapical lesions (p = 0.018), as detailed in Supplementary Tables A.2 and A.3 and Supplementary Fig. A.2.

Fig. 4.

Per-class Dice performance across representative methods. The shaded areas group tooth structures, restorative materials, and lesion classes

Figure 5 provides a qualitative comparison of segmentation results across different architectures and learning paradigms. The rows include four supervised CNN baselines (LinkNet, FPN, DeepLabV3 +, and U-Net), supervised Mask2Former, Mask2Former models fine-tuned from different self-supervised pre-trained weights, including BEiTv3, MoCov3, DINOv2, and IGLC-MAE, and the final adaptive semi-supervised model. The selected periapical samples cover typical clinical challenges, including complex root apex morphology, multi-material restorations, thin root canal fillings, and subtle low-contrast lesion regions. Overall, conventional CNN-based methods provide a visual baseline but show evident limitations. Their predictions are often fragmented around thin or small structures, such as fillings and root canal fillings, and they may miss low-contrast internal regions such as pulp or periapical lesions. In cases with ambiguous grayscale overlap, CNN outputs also show boundary leakage or class mixing, reducing anatomical plausibility.

Fig. 5.

Qualitative comparison of periapical film segmentation across models. Rows show the input, ground truth, and predictions from our adaptive Mask2Former, self-/supervised Mask2Former, and CNN baselines; colors denote tooth (light sky blue), pulp (violet), caries (green), fillings (gray), root canal fillings (orange), periapical lesions (dark blue), and crown (brick red)

Compared with CNN baselines, Mask2Former-based methods generally produce more coherent anatomical structures and smoother boundaries. Among the self-supervised baselines, BEiTv3 and MoCov3 improve structural completeness to some extent but remain less stable in fine boundaries and lesion-related regions. DINOv2 shows strong overall segmentation quality, consistent with its large-scale pre-training advantage. IGLC-MAE achieves comparable qualitative performance to DINOv2 in several examples, suggesting that domain-specific adaptive masking can provide competitive representations for periapical radiographs despite using a much smaller in-domain pre-training corpus. The final adaptive semi-supervised model shows the most consistent overall delineation among the compared methods. It better preserves continuity between adjacent dental structures and improves the segmentation of fine or low-contrast regions, including pulp, root canal fillings, caries, and periapical lesions. These qualitative findings are consistent with the quantitative results, where the adaptive semi-supervised model achieved the highest overall Dice score. Together, the visual and numerical comparisons suggest that domain-specific self-supervised pre-training and confidence-adaptive pseudo-label learning provide complementary benefits for periapical film segmentation.

Effect of unlabeled data size in self-supervised learning

We further investigated the impact of the amount of unlabeled data on model performance. As summarized in Table 2, using only ImageNet initialization without self-supervised pre-training yielded a Dice score of 67.94%. Introducing low-resolution unlabeled data consistently improved performance, with Dice rising to 70.40% at 10 k images and gradually increasing to 72.00% at 60 k images, showing a clear positive trend with mild saturation. Pre-training on the full unlabeled set (6,259 high-resolution + 68,033 low-resolution; 74,292 total) achieved the best result, reaching a Dice score of 73.17%. These findings indicate that increasing unlabeled data volume strengthens representation learning and that mixed-resolution pre-training may provide complementary benefits. Although low-resolution thumbnails were upsampled before pre-training, the performance trend suggests that their distributional diversity and global structural information still contributed positively to representation learning.

Table 2.

Effect of different training data volumes on segmentation performance

Subset Name	High Resolution	Low Resolution	Total Images	Dice (%)
imagenet	0	0	0	67.94 ± 1.15
subset_highres_only	6,259	0	6,259	68.78 ± 1.05
subset_lowres_10k	0	10,000	10,000	70.40 ± 1.32
subset_lowres_20k	0	20,000	20,000	71.06 ± 0.90
subset_lowres_30k	0	30,000	30,000	71.15 ± 0.88
subset_lowres_40k	0	40,000	40,000	71.39 ± 0.85
subset_lowres_50k	0	50,000	50,000	71.71 ± 0.84
subset_lowres_60k	0	60,000	60,000	72.00 ± 0.75
All_data	6,259	68,033	74,292	73.17 ± 0.95

Semi-supervised strategy with static pseudo-labeling

In stage two, we explored semi-supervised learning strategies to further improve model performance. We first used the trained self-supervised model to generate pseudo-labels for 6,259 high-resolution unlabeled images, then jointly trained the segmentation model with these pseudo-labeled data and 229 manually annotated images. Table 3 shows the performance variations under different loss weight configurations. When training with only 229 manually annotated images, the baseline Dice score was 73.17%. By adjusting the loss weight ratio between manual annotations and pseudo-labeled data, we found that when the manual annotation weight was 0.85 and the pseudo-label weight was 0.15, the Dice score reached the highest value of 74.12%, an improvement of 0.95 percentage points compared to using only manually annotated data. As the pseudo-label weight increased or decreased, model performance declined, such as a Dice value of 73.81% at a weight ratio of 0.8:0.2, and a Dice value of 73.73% at a ratio of 0.9:0.1. Overall, these configurations showed a modest improvement over the baseline result (73.17%) obtained using only manually annotated data, indicating that introducing pseudo-labeled data can be beneficial within an appropriate weight range, while an optimal balance point is still required.

Table 3.

Performance Variations Under Different Weight Configurations

Dataset	Loss Weight Ratio	Dice (%)
6,259 pseudo-labeled + 229 manual-labeled	0:1.0	73.17 ± 0.95
	0.1:0.9	73.73 ± 1.05
	0.15:0.85	74.12 ± 0.85
	0.2:0.8	73.81 ± 0.88
	0.25:0.75	73.68 ± 0.92
	0.3:0.7	73.24 ± 0.95

Optimization via adaptive semi-supervised weighting

The results from Table 3 demonstrated that while static pseudo-labeling is effective, the performance is limited by the fixed loss weight ratio, which cannot discriminate between high-quality and noisy pseudo-labeled pixels. To overcome this limitation and maximize the potential of the unlabeled data, we introduced an online confidence-adaptive pseudo-label weighting strategy based on the Teacher–Student framework with exponential moving average (EMA).

In this framework, the teacher model generates pseudo-labels and confidence maps for unlabeled images, while the student model is optimized using both supervised loss from labeled images and confidence-weighted pseudo-label loss from unlabeled images. The key component of this strategy is the dynamic scaling of the pseudo-label loss using a pixel-wise weight , which is calculated from the teacher model’s confidence C via a temperature-controlled sigmoid function. As shown in Fig. 6, this adaptive weighting function assigns lower weights to low-confidence pixels and progressively increases the contribution of high-confidence pseudo-labels. This design reduces the influence of unreliable pseudo-supervision while preserving useful training signals from confident regions.

Fig. 6.

Comparison of confidence-adaptive pseudo-label loss weighting strategies. The curves show how the pixel-wise pseudo-label loss weight Ŵ changes with teacher confidence C under different Tc and τ settings. The horizontal line at W = 0.5 indicates the sigmoid midpoint. Configuration B ( = 0.8, τ = 0.10) achieved the best balance between down-weighting uncertain pseudo-labels and retaining useful supervision from high-confidence regions

We performed an ablation study on the two core self-adaptive parameters: the confidence center and the temperature coefficient τ, using the best static pseudo-labeling configuration as the baseline. The detailed comparison is presented in Table 4. The results in Table 4 confirm the superiority of the adaptive approach. Our optimal configuration, configuration B ( = 0.8, τ = 0.10), achieved the highest Dice score of 75.35%. This represents a 1.23 percentage point improvement over the best static semi-supervised result of 74.12%. The comparison across different adaptive configurations further clarifies the importance of balancing noise suppression and the effective use of pseudo-labels. Configuration A ( = 0.8, τ = 0.20) used a smoother weighting curve, which may retain more uncertain pseudo-label signals and therefore provided only moderate improvement. Configuration C ( = 0.9, τ = 0.10) applied a stricter confidence center, which more aggressively suppressed pseudo-labeled pixels but may also discard informative medium-confidence regions around ambiguous anatomical or lesion boundaries. Configuration D ( = 0.8, τ = 0.05) used the smallest temperature coefficient, producing a very sharp transition close to hard thresholding, which may reduce optimization stability. In contrast, configuration B provided a moderately sharp but continuous transition and achieved the highest Dice score. These results suggest that the best performance was obtained not by retaining the largest amount of pseudo-label supervision or by applying the strictest filtering, but by achieving an appropriate balance between suppressing uncertain pseudo-labels and preserving useful supervision from reliable regions.

Table 4.

Performance variations under Adaptive Loss Weighting Configurations

Configuration		τ	Dice (%)
Baseline (Static Weighting)	N/A	N/A	74.12 ± 0.90
A (Smooth Filter)	0.8	0.20	74.90 ± 0.82
B (Optimal Adaptive)	0.8	0.10	75.35 ± 0.70
C (Strict Weighting)	0.9	0.10	74.50 ± 0.85
D (Sharp Transition)	0.8	0.05	74.33 ± 0.88

Discussion

In this study, we propose a hybrid self-supervised and semi-supervised learning framework and demonstrate its superior performance in periapical film segmentation. Traditional supervised learning methods heavily rely on high-quality annotated data. Such data are particularly valuable and costly to acquire in the medical imaging field. Dental radiographs exhibit complex structures, blurred boundaries, and low contrast, making precise annotation extremely challenging and requiring specialized knowledge of oral radiology. In contrast, our self-supervised IGLC-MAE pre-training exploits large-scale unlabeled periapical data. It learns robust oral tissue representations from these data. This approach reduces reliance on scarce annotations. It also provides a strong initialization for downstream segmentation. Building upon this foundation, we further applied semi-supervised strategies. From the initial improvement gained by static pseudo-labeling, we advanced to an online adaptive loss weighting mechanism. This dynamic strategy enabled the model to filter pseudo-label quality. It ensures that learning is guided solely by high-confidence signals. Consequently, it successfully suppresses noise and achieves the highest segmentation accuracy.

Building upon the self-supervised learning phase, we introduced IGLC-MAE. It enhances the traditional MAE framework through an innovative structured block-based adaptive masking strategy. Traditional MAE adopts a random masking strategy [19], which may not adequately focus on key diagnostic areas in medical images. In contrast, IGLC-MAE, through an adaptive masking mechanism based on Intensity, Gradient, and Local Contrast, can intelligently identify key information areas in images. This mechanism first calculates a comprehensive score for each image patch, then prioritizes masking high-scoring areas, forcing the model to learn feature representations of these information-rich regions. In oral X-rays, different anatomical structures (such as tooth, pulp) and lesion areas (such as caries, periapical lesions) have different image characteristics. By specifically masking these areas, the model can learn more meaningful feature representations. This leads to better performance in downstream segmentation tasks. Importantly, this improvement should not be interpreted merely as an effect of using more unlabeled data. Standard MAE and IGLC-MAE were trained on the same unlabeled dataset, but IGLC-MAE defines a more diagnostically aligned reconstruction task by preferentially masking regions with stronger contrast transitions and boundary information. This design reduces the mismatch between self-supervised pre-training and downstream segmentation, particularly for small or low-contrast structures. Therefore, the key contribution is the combination of large-scale unlabeled dental radiographs with a radiographic-semantics-guided masking policy.

The choice of Mask2Former further supports effective utilization of the representations learned in stage one. Periapical film segmentation is inherently multi-class and fine-grained. It requires simultaneous delineation of adjacent anatomical structures and small lesions. Mask2Former formulates segmentation as query-based mask prediction with multi-scale feature fusion, which promotes clearer class separation and smoother, more complete boundaries under low-contrast and blurred-edge conditions [32]. This is particularly beneficial for visually similar or tightly coupled structures, such as pulp versus fillings and subtle periapical lesions near the root apex. Accordingly, our downstream model pairs an IGLC-MAE–pretrained ViT backbone with a Mask2Former segmentation head. Conventional CNN-based decoders often struggle to maintain fine-grained boundaries under limited annotations. In contrast, the Mask2Former head can more effectively translate domain-adaptive representations into accurate multi-class masks. These representations were learned from large unlabeled data. As a result, the framework preserves structural continuity across roots, pulp, restorations, and lesion regions, yielding robust segmentation in challenging periapical radiographs.

Regarding the data strategy for self-supervised learning, we systematically explored the relationship between unlabeled data volume and pre-training effect. The results are summarized in Table 2. The results show that as the number of unlabeled low-resolution images increased from 10,000 to 60,000, the Dice steadily improved from 70.40% to 72.00%. This positive correlation confirms the positive impact of large-scale unlabeled data on self-supervised learning performance: more diverse training samples allow the model to encounter a wider range of data patterns, thereby learning more robust feature representations. However, we also observed that this relationship has a certain saturation effect; when the data volume reaches a certain threshold, performance improvement tends to plateau. It is also important to consider the resolution heterogeneity of the unlabeled dataset. All images were resized to 224 × 224 pixels before MAE pre-training, which ensured a unified input size and patch scale. Nevertheless, upsampled low-resolution thumbnails cannot recover true high-frequency anatomical details and may introduce interpolation blur. Therefore, the contribution of low-resolution thumbnails should be interpreted mainly as providing distributional diversity and global structural context, rather than replacing the fine anatomical detail available in native high-resolution radiographs. In comparison, the smaller high-resolution subset provides finer boundary and texture information. Furthermore, it is worth noting that pre-training with only a large amount of low-resolution data (e.g., 60 k, Dice = 72.00%) far outperforms using only a small amount of high-resolution data (Dice = 68.78%). This indicates that for self-supervised learning tasks, data "quantity" may be more critical than "quality" (resolution). This is especially true under our framework, which focuses on reconstructing image structures and contextual information. We further discuss the balance relationship between data quality and quantity. When we combined 6,259 high-resolution images with 68,033 low-resolution images, the Dice score reached its highest value of 73.17%. High-resolution images are relatively few in number. Nevertheless, they provide more detailed visual information. When combined with a large number of low-resolution images, the model can learn both fine-grained visual features and diverse data distributions. This finding provides guidance for data collection strategies in the medical imaging field: under limited resources, prioritize obtaining high-quality data while supplementing with a large amount of more easily obtainable ordinary quality data to achieve a better performance balance.

In the second phase of our framework, we further analyzed the effectiveness of the semi-supervised pseudo-labeling strategy. The experimental results show clear improvement. By introducing 6,259 pseudo-labeled images and optimizing loss weight configurations, the model's Dice score improved from 73.17% to 74.12%. This confirms the value of this strategy in oral medical image segmentation tasks. Although pseudo-labels may contain certain errors, they encompass structural information and feature patterns from a large amount of unlabeled data. By integrating this information, the model gains stronger generalization capabilities. It achieves high segmentation accuracy even with limited annotation resources. This has important implications for the oral medical imaging field, where annotation costs are high and professional personnel resources are tight. Experimental data shows that the loss weight configuration has a significant impact on model performance. The manual annotation to pseudo-label weight ratio of 0.85:0.15 achieved the best results; under this configuration, the model can fully utilize the precise boundary information provided by high-quality manual annotations while learning broader data distribution features from numerous pseudo-labels. However, when the pseudo-label weight is too high (e.g., ≥ 0.2), it leads to performance degradation, as the inaccurate information contained may interfere with the model's learning of accurate annotations, thereby affecting final performance.

To address the limitation of static pseudo-labeling, we further introduced an adaptive semi-supervised strategy based on confidence-guided loss re-weighting. Although the best static pseudo-label setting improved performance, it still assigned a uniform weight to all pseudo-labeled pixels. It therefore could not distinguish reliable pseudo-labels from uncertain predictions. In contrast, the proposed teacher–student framework uses the EMA teacher to generate online pseudo-labels and confidence estimates. This allows the pseudo-label loss to be adjusted at the pixel level. As a result, pseudo-label supervision shifts from a quantity-driven strategy to a quality-aware strategy. High-confidence regions contribute more strongly, while uncertain pixels are down-weighted. The ablation study of the confidence center and temperature coefficient further shows that the effectiveness of adaptive weighting depends on achieving an appropriate balance rather than simply retaining more pseudo-label supervision or applying stricter filtering. A smoother weighting curve may preserve more uncertain pseudo-label signals, whereas an overly strict or overly sharp transition may discard informative medium-confidence pixels or lead to unstable optimization. The optimal setting provided a moderately sharp but continuous transition, which retained useful supervision from reliable regions while reducing the influence of noisy pseudo-labels. This balance is particularly important for periapical radiographs, where blurred boundaries, overlapping anatomical structures, early caries, and subtle periapical lesions often produce medium-confidence predictions rather than clearly correct or incorrect pseudo-labels. Therefore, IGLC-MAE and confidence-adaptive semi-supervised learning play complementary roles: the former provides a domain-adapted feature initialization, whereas the latter improves the reliability of pseudo-label supervision during downstream training.

Beyond technical metrics, it is important to consider how these segmentation outputs translate to clinical practice. From a clinical perspective, this study should be interpreted as a pre-clinical methodological validation rather than direct evidence of improved patient outcomes. The primary contribution is to address annotation scarcity in multi-structure periapical film segmentation; the current experiments, therefore, focus on algorithmic accuracy rather than prospective clinical decision testing. Nevertheless, the seven-class output was designed to correspond to clinically meaningful tasks. Tooth, pulp, and crown segmentation provide anatomical references for endodontic treatment planning, including pulp-space morphology and root-canal trajectory assessment. Fillings and root canal fillings support retrospective evaluation of restorative or endodontic treatment status, such as identifying secondary caries risk and estimating the relationship between root filling material and the radiographic apex. Caries and periapical lesion masks directly correspond to common diagnostic targets; their segmented areas may help quantify lesion extent and support longitudinal comparison during follow-up (e.g., monitoring whether a periapical radiolucency heals or progresses after treatment). Thus, the segmentation output may serve as a structured visual aid and quantitative reference. It can support treatment planning, lesion monitoring, restoration assessment, and image-based follow-up. From an engineering standpoint, end-to-end inference for a single periapical radiograph takes approximately 0.4 s on a single NVIDIA RTX 3090 GPU (Supplementary Table A.4), indicating that the current implementation is compatible with offline batch analysis or server-assisted clinical reading workflows without introducing prohibitive latency. However, it is intended to assist rather than replace dentists' clinical judgment. Prospective reader studies, external multi-center validation, and workflow-based clinical evaluation are required before deployment.

While our method achieves encouraging results, several limitations need to be addressed in future work. First, the dataset composition presented challenges: only 6,259 high-resolution images were available, while the remaining 68,033 low-resolution thumbnails required upsampling, a process that may compromise fine details. We note that these artifacts are confined to the pre-training phase, as downstream segmentation operates on native high-resolution inputs; furthermore, MAE’s patch-level reconstruction objective exhibits inherent tolerance to such interpolation defects. Nevertheless, rare micro-structures (e.g., early periapical radiolucencies) may be under-represented, and future work could explore super-resolution preprocessing or multi-scale patch embedding to mitigate this resolution heterogeneity. Second, the pathological diversity of the annotated cohort was constrained. Consistent with the objective of this study, the labeled dataset was confined to routine periapical films and deliberately excluded rare and complex presentations, such as implants, bridges, supernumerary or deciduous teeth, residual roots, developmental anomalies, and severe atypical pathologies. Moreover, even among the included classes, clinically critical lesions were under-represented: within the annotated cohort, caries and periapical lesions accounted for only 1.16% and 0.88% of annotated foreground pixels, respectively, despite instance-level prevalences of 6.56% and 3.94% (Supplementary Table A.1). These lesions are thus spatially small minority classes, underscoring the difficulty of segmenting fine, low-contrast pathological structures under severe class imbalance. Consequently, although the model demonstrates utility for routine structures, restorations, caries, and periapical lesions, its behavior on atypical or severe clinical presentations remains to be established. As the data were also drawn from a single center and a predominantly young-adult population, future work will prioritize multi-center collection together with the inclusion of atypical and pathological-variant samples, in order to broaden demographic and pathological coverage and to support wider clinical generalizability. Third, the current research relies exclusively on 2D periapical radiographs, which cannot fully capture the three-dimensional features of dental structures. Future work could incorporate 3D imaging modalities such as cone-beam computed tomography (CBCT) to develop multi-modal learning frameworks. Fourth, although the proposed hybrid-supervised learning reduces dependence on annotated data, the scale of 229 manually annotated images is still relatively limited; expanding the annotated dataset, especially including more atypical and pathological variant samples, is expected to further improve model performance in complex situations. Additionally, the current model mainly focuses on segmentation tasks; future work could develop multi-task learning frameworks that provide both segmentation results and diagnostic suggestions to increase its clinical utility. Finally, clinical validation remains incomplete: this study did not include a formal model-versus-expert comparison, and prospective clinical trials are needed to assess the actual impact on diagnostic accuracy, physician work efficiency, and patient prognosis. In summary, our research provides valuable technical advancements for automated analysis of oral X-rays, particularly innovations in self-supervised learning and pseudo-labeling, offering new pathways to improve the accuracy and efficiency of oral medical image analysis. Future work will further refine these methods, address current limitations, and ultimately achieve the widespread application of AI-assisted oral diagnostic systems in clinical practice.

Conclusion

In conclusion, this study presents a hybrid self-supervised and semi-supervised learning framework tailored for multi-class periapical film segmentation. The core contribution lies in two synergistic designs: an intensity–gradient–local-contrast-guided adaptive masking strategy that directs MAE pre-training toward diagnostically informative anatomical regions, and a confidence-adaptive pseudo-label weighting mechanism that enables robust semi-supervised learning by softly suppressing low-certainty pixels while preserving boundary supervision. By combining these innovations, the framework achieves accurate segmentation of tooth, pulp, crown, fillings, root canal fillings, caries, and periapical lesions using only 229 manually annotated images, while leveraging 74,292 unlabeled radiographs to learn domain-adaptive representations. This substantially reduces reliance on costly expert annotation, offering a practical pathway for developing dental AI systems in clinical settings where labeled data are scarce.

Authors’ contributions

Qianli Zhang: Writing—original draft, Conceptualization, Methodology, Data Curation and Annotation, Resources and Ethics Approval Meiyu Hu: Writing—original draft, Validation, Methodology, Formal analysis, Conceptualization Mu Yuan: Data Curation and Annotation Pingyi Jia: Data Curation and Annotation Huajie Yu: Data Curation, Supervision Feng Chen: Validation, Supervision Xucheng Yin: Supervision, Data curation Junran Peng: Writing—review & editing, Project administration, Conceptualization.

Funding

This work was supported in part by Guangdong Basic and Applied Basic Research Foundation (No. 2024A1515110065).

Data availability

The clinical imaging data generated and analyzed during the current study are not publicly available due to patient privacy protection and institutional ethical restrictions at Peking University School of Stomatology. However, the code is publicly available on GitHub (https://github.com/Molly720-my/Hybrid-Supervision-Periapical-Film.).

Declarations

Ethics approval and consent to participate

The study was conducted following the Declaration of Helsinki and approved by the Ethics Committee of Peking University School of Stomatology (PKUSSIRB-2025107012). All periapical films included in the study were taken during routine diagnosis and treatment at Peking University School of Stomatology and affiliated hospitals between 2012 and 2024, with patient informed consent for research purposes. The research process strictly adhered to medical ethical guidelines, ensuring patient privacy and data security.

Consent for publication

The manuscript contains periapical films from participants, has obtained the consent of the patients for publication.

Competing interests

The authors declare no competing interests.