Dental X-Ray Segmentation Model Comparison: Why Does UNet++ Come Out on Top?

Automatic segmentation of dental X-ray images has long been a challenging problem in medical image analysis. Low contrast, noise interference, and inconsistent imaging quality combine to make even experienced radiologists prone to interpretation discrepancies. In recent years, deep learning-based semantic segmentation methods have been widely introduced to this field, but a practical problem persists: different studies use different experimental setups, making fair comparisons between models nearly impossible.

A recent study addressed this gap by systematically comparing four mainstream segmentation models—PSPNet, DeepLabV3, UNet, and UNet++—under unified experimental conditions. This analysis breaks down their experimental design, core findings, and the practical implications of these results for dental imaging AI applications.

Why Is Dental X-Ray Segmentation So Difficult?

Dental radiographic images are fundamentally different from natural images. The grayscale differences between teeth, alveolar bone, and periodontal tissues under X-ray are extremely subtle, and the boundaries between lesion areas and normal tissue are often blurred. Add to that variations in exposure parameters across different devices, differences in patients' oral structures, and artifacts from metal fillings, and image quality fluctuates considerably.

These challenges directly affect the performance ceiling of segmentation models. The study points out that precisely because of these complexities, directly comparing model performance from existing literature becomes difficult—different studies use inconsistent image sizes, batch sizes, and preprocessing pipelines, making conclusions hard to cross-validate.

Experimental Design: A Fair Showdown Under Unified Conditions

Dataset and Preprocessing

The study used 1,269 dental X-ray images and applied a series of preprocessing operations before training. This step is critical because the noise patterns and contrast distributions of dental images differ significantly from general-purpose datasets, and directly transferring ImageNet pretrained weights without preprocessing often yields poor results.

The Four Compared Models

• PSPNet: Captures multi-scale contextual information through a pyramid pooling module
• DeepLabV3: Uses atrous convolution and ASPP structure to handle multi-scale targets
• UNet: Classic encoder-decoder architecture with skip connections at its core
• UNet++: Adds nested dense skip connections on top of UNet to enhance feature fusion

The encoder-decoder architecture is the foundational paradigm of the UNet family of models. The encoder progressively reduces the spatial resolution of feature maps through a series of convolution and pooling operations to extract high-level semantic information; the decoder then restores low-resolution features to the original image size through upsampling (transposed convolution or interpolation), thereby outputting pixel-wise segmentation results. Skip connections are the key design of this architecture: they pass high-resolution features from each encoder level directly to the corresponding decoder level, compensating for spatial details lost during downsampling. For fine structures like tooth contours and alveolar bone boundaries, skip connections can significantly improve segmentation edge accuracy—this is also the core reason UNet is widely adopted in the medical imaging field.

Atrous convolution (also known as dilated convolution) is the core component of DeepLabV3. Unlike standard convolution where kernel elements are spatially contiguous, atrous convolution inserts gaps (i.e., dilation rates) between kernel elements, thereby expanding the receptive field without increasing the number of parameters. This is particularly important in semantic segmentation: segmentation tasks require both a sufficient receptive field to understand global context and high spatial resolution to precisely locate boundaries. DeepLabV3 uses the ASPP (Atrous Spatial Pyramid Pooling) structure to extract multi-scale features in parallel at different dilation rates, effectively addressing the problem of scale variation in segmentation targets.

The Pyramid Pooling Module (PPM) is the core design of PSPNet. It pools feature maps into regions of multiple different sizes (e.g., 1×1, 2×2, 3×3, 6×6), then upsamples and concatenates the pooling results from each scale. This approach allows the network to simultaneously capture local details and global contextual information, making it particularly suitable for segmentation tasks involving targets of varying sizes. In dental X-rays, tooth sizes vary considerably, and PPM's multi-scale aggregation capability has a certain degree of adaptability.

Evaluation Metrics

The study used the Dice coefficient (DICE), IoU, accuracy, precision, and training/validation loss to comprehensively evaluate the models. Among these, IoU and DICE are the most critical metrics for segmentation tasks—the former measures the degree of overlap between predicted and ground truth regions, while the latter is more robust to class imbalance.

The Dice coefficient (Dice Similarity Coefficient) and IoU (Intersection over Union) are the two most commonly used overlap metrics in semantic segmentation. IoU is calculated as the intersection area of the predicted region and the ground truth annotation divided by the union area, with values ranging from 0 to 1—the higher the value, the more accurate the prediction. The Dice coefficient is defined as twice the intersection area divided by the sum of the two region areas, and is mathematically highly correlated with IoU, but Dice is more robust to class imbalance—when the target region is much smaller than the background region, IoU values are diluted by the large number of background pixels, whereas Dice is more sensitive to foreground targets. In dental X-ray segmentation, lesions or specific tooth regions often occupy only a small portion of the entire image, so Dice is generally more suitable than IoU as the primary evaluation criterion. Reporting both metrics simultaneously provides a more comprehensive picture of model performance across different evaluation dimensions.

Variable Control

Experiments were repeated at different image sizes and batch sizes, allowing researchers to observe the specific impact of hyperparameters on model performance rather than simply reporting the best results under a single configuration.

What Makes UNet++ Stand Out?

Under the configuration of 300×300 image size and batch size 16, UNet++ achieved the highest IoU of 0.6719 and DICE of 0.7869, performing best among all compared models.

This result deserves closer examination. The core improvement of UNet++ over UNet lies in its nested skip connection structure—rather than simply building a single pathway between the encoder and decoder, it redesigns feature fusion paths at every level. For tasks like dental X-ray segmentation with blurred boundaries and a need for precise localization, this dense connectivity allows the network to more fully utilize feature information at different scales.

Another interesting phenomenon is the impact of hyperparameters. When image size increased from smaller values to 300×300, model performance changed noticeably, indicating that detail information in dental X-rays has a certain dependence on resolution. The batch size of 16 was also determined as a relatively optimal configuration after multiple experiments.

Nested dense skip connections are the core innovation of UNet++ relative to UNet. Traditional UNet's skip connections are a straight pathway from encoder directly to decoder, with no lateral interaction between different levels. UNet++ instead constructs dense intermediate nodes between each level of encoder and decoder, with each node receiving feature inputs from the encoder at the same level as well as from all lower-level intermediate nodes. This design shortens the semantic gap between encoder and decoder, enabling fine-grained spatial information from shallow layers and abstract semantic information from deep layers to be fully fused across multiple scales. For the blurred boundaries between teeth and alveolar bone regions in dental X-rays, this multi-level, multi-path feature aggregation mechanism allows the network to more accurately locate and organize segmentation boundaries, thereby achieving optimal results in the experiments.

Three Takeaways for Dental AI Applications

First, fair comparison matters more than chasing SOTA. The value of this study is not in proposing a new model, but in providing a reproducible comparison baseline under controlled conditions. For teams looking to deploy AI in dental imaging, this type of work offers a reliable reference for model selection.

Second, UNet++'s structural advantages have been repeatedly validated in medical imaging. From cell segmentation to dental X-rays, nested skip connections have consistently delivered stable results for fine structure segmentation. If you want to quickly build a baseline system for dental scenarios, UNet++ is an architecture worth trying first.

Third, an IoU of 0.67 means there is still significant room for improvement. This number indicates that even the best-performing model is far from reaching a level of full clinical automation for dental X-ray segmentation. Data augmentation, domain adaptation, and the construction of larger-scale annotated datasets remain key directions for future research.

This study provides a solid starting point for automated analysis of dental radiographic images. The performance differences among the four models under unified conditions, along with data on the impact of image size and batch size, serve as benchmarks that future research can directly reference. The potential of deep learning in dental imaging is undeniable, but from experimental metrics to clinical deployment, there is still a great deal of engineering and validation work to be done.

One-line Summary

Under unified experimental conditions comparing four segmentation models, UNet++ wins in dental X-ray segmentation with an IoU of 0.6719.

Paragraph Summary

This study systematically compared four mainstream semantic segmentation models—PSPNet, DeepLabV3, UNet, and UNet++—on dental X-ray images under unified experimental conditions. Using 1,269 dental X-ray images, the study evaluated Dice coefficient, IoU, accuracy, and other metrics, and examined the effects of hyperparameters such as image size and batch size. Results showed that UNet++ achieved the highest IoU of 0.6719 and DICE of 0.7869 under a 300×300 image size and batch size of 16, with its nested dense skip connection structure being particularly effective for fine segmentation tasks with blurred boundaries in dental X-rays. However, the current best model's IoU of only about 0.67 indicates that dental X-ray automatic segmentation still has a significant gap from full clinical automation, with data augmentation, domain adaptation, and large-scale annotated data construction being key future directions.