Source-linked AI summary
Automatic segmenting teeth in X-ray images: Trends, a novel data set, benchmarking and future perspectives
Gil Jader, Luciano Oliveira, Matheus Pithon
TL;DR
Dental X-ray segmentation lacks sufficiently varied evidence for evaluating automated tooth isolation, particularly for complex extra-oral images. The paper reviews and categorizes prior methods, introduces a 1,500-image panoramic dataset, and benchmarks ten segmentation methods. Niblack achieved the strongest reported tooth-segmentation performance, while the study identifies unresolved difficulty and future scope for learning-based methods.
Problem
Variations in teeth, restorative artifacts, noise, low contrast, missing teeth, and surrounding facial structures make automated dental X-ray analysis difficult.
Method
The paper reviews dental X-ray segmentation literature, categorizes methods and image types, and evaluates ten methods on 1,500 annotated panoramic images.
Results
Niblack reached approximately 83% recall and approximately 82% specificity, with superior overall performance for tooth segmentation among the evaluated methods.
Takeaways & Limitations
The review provides a broader performance assessment for panoramic tooth segmentation and identifies local thresholding with Niblack as the strongest evaluated approach.
Takeaways & Limitations
Reviewed datasets generally had small variation, and panoramic X-ray characteristics made complete tooth isolation difficult; learning-based methods were identified as a future direction.
Abstract
from arXiv · showhide
This review presents an in-depth study of the literature on segmentation methods applied in dental imaging. Ten segmentation methods were studied and categorized according to the type of the segmentation method (region-based, threshold-based, cluster-based, boundary-based or watershed-based), type of X-ray images used (intra-oral or extra-oral) and characteristics of the dataset used to evaluate the methods in the state-of-the-art works. We found that the literature has primarily focused on threshold-based segmentation methods (54%). 80% of the reviewed papers have used intra-oral X-ray images in their experiments, demonstrating preference to perform segmentation on images of already isolated parts of the teeth, rather than using extra-oral X-rays, which show tooth structure of the mouth and bones of the face. To fill a scientific gap in the field, a novel data set based on extra-oral X-ray images are proposed here. A statistical comparison of the results found with the 10 image segmentation methods over our proposed data set comprised of 1,500 images is also carried out, providing a more comprehensive source of performance assessment. Discussion on limitations of the methods conceived over the past year as well as future perspectives on exploiting learning-based segmentation methods to improve performance are also provided.
1. Introduction
Dental X-ray images support diagnosis but remain difficult to analyze automatically, especially extra-oral images containing teeth alongside surrounding anatomical structures. This review addresses these challenges by surveying segmentation methods and benchmarking ten methods on a diverse dataset of 1,500 annotated panoramic images.
- Dental X-rays aid diagnosis by revealing the condition of teeth, gums, jaws, and mouth bones.
- Patient variation, restorative artifacts, noise, low contrast, missing teeth, and acquisition limits hinder automated dental X-ray analysis.
- The review seeks segmentation methods that isolate teeth from jaws, facial structures, and gums to facilitate automatic X-ray analysis.
- Existing studies generally used small datasets, while the only study exceeding 1,000 images was not publicly available or varied mainly by tooth number.
- 1,500 annotated panoramic X-ray images were organized into 10 categories covering structural and clinical variation, including restorations, implants, appliances, and tooth counts.
2. Research methodology
The review selected and categorized relevant dental X-ray segmentation studies from four digital libraries, then benchmarked the reviewed methods and synthesized publication patterns. The selection process ended with 41 articles and showed increasing publication activity in recent years.
- The review selected articles from IEEE Xplore, ScienceDirect, Google Scholar, and Scopus, then classified them by segmentation method, X-ray type, and dataset characteristics.
- 94 articles were initially identified, and the selection process reduced the corpus to 41 articles.
- 49% of selected relevant articles came from ScienceDirect, while IEEE Xplore contributed 34 articles, or 36%.
- 66% of the articles found were published in the last five years, indicating an increasing trend in recent publications.
- Table 1 reports the total number of studies found by digital library.
- Table 2 distributes articles by digital library and year of publication.
3. Taxonomy of the relevant works
The reviewed works are classified by segmentation approach, X-ray type, and dataset characteristics. Threshold-based methods and intra-oral images predominate, while dataset sizes are generally small and variable.
- Segmentation methods: Region-based methods divide images according to pixel-intensity discontinuities, whereas cluster-based methods group data by similarity with the number of groups specified initially.
- Segmentation methods: 54% of relevant papers use threshold-based segmentation, making it the most common category.Thresholding assigns pixels to regions according to whether their intensity exceeds a selected threshold.
- Segmentation methods: Histogram-based thresholding can fail under substantial contrast or illumination variation, motivating shading-function compensation for non-uniform intensities.
- Segmentation methods: Boundary-based methods appear in 34% of relevant papers and search for image-intensity discontinuities such as points and edges.
- X-ray image types: Approximately 80% of reviewed papers use intra-oral X-ray images, and only three use extra-oral panoramic images.
- Dataset characteristics: 61% of relevant papers use datasets containing 1–100 X-ray images, while reviewed datasets generally show limited variation beyond tooth number.
4. Evaluation of the segmentation methods
The study evaluates ten segmentation methods on a manually categorized, annotated panoramic X-ray dataset, using image-region statistics and standard performance metrics. Results show substantial differences across metrics, with Niblack providing the strongest recall and local thresholding outperforming global thresholding.
- Dataset construction: 1,500 panoramic X-ray images were manually categorized into 10 structural categories and annotated for teeth and oral regions.The dataset includes variations in tooth structure, tooth number, restorations, implants, appliances, supernumerary teeth, and mouth or jaw size.
- Evaluation methodology: The evaluation determined image statistics within each region of interest and used those statistics as parameters for the segmentation algorithms.Only pixels inside the image ROIs were considered when calculating evaluation metrics.
- Evaluation methodology: The methods were assessed using accuracy, specificity, precision, recall, and F-score.These metrics were applied to the image ROIs rather than the full panoramic images.
- Results: Splitting and merging, Sobel, fuzzy C-means, and Canny produced high specificity but poor recall, indicating predominance of true negatives.Splitting and merging and Sobel achieved almost perfect specificity, whereas fuzzy C-means and Canny exceeded 90% specificity.
- Results: Niblack achieved the highest recall at approximately 83% and approximately 82% specificity, while marker-controlled watershed was the only other method exceeding 80% recall.Active contour without edges and level set segmentation achieved less than 70% recall.
- Results: Local thresholding with Niblack significantly improved panoramic tooth segmentation compared with basic global thresholding.Niblack presented superior performance among the evaluated segmentation methods.
5. Discussion and conclusions
The review finds that dental X-ray segmentation research has favored threshold-based methods and intra-oral images, while extra-oral panoramic images remain difficult to segment reliably. It addresses this gap with a diverse 1,500-image dataset and benchmarking of ten methods, then points toward learning-based approaches.
- Discussion: 80% of reviewed articles used intra-oral X-rays, favoring images showing isolated tooth regions over extra-oral images.Extra-oral radiographs also contain jaw, temporomandibular, and other facial structures that complicate analysis.
- Discussion: Threshold-based segmentation was the predominant category in the reviewed literature.
- Conclusions: 1,500 panoramic X-ray images formed a diverse dataset spanning 10 defined categories for evaluating segmentation performance.The study measured accuracy, specificity, precision, recall, and F-score across ten segmentation methods.
- Conclusions: None of the ten evaluated algorithms completely isolated teeth, failing mainly because bone parts were included.
- Future perspectives: Machine learning and energy-minimization operations are proposed as future directions for highlighting tooth regions in orthopantomography images.The key challenge is overlap between tooth characteristics and the jaw and skull in extra-oral X-rays.