Cephalometric Landmark Detection in Lateral X-ray Images

Cephalometric analysis is a fundamental examination which is routinely used in fields of orthodontics and orthognathics. The key operation in the analysis is marking craniofacial landmarks from lateral cephalograms. These landmarks serve as the datum of the succeeding qualitative assessment of angles and distances, which provide diagnosis information of the craniofacial condition of a patient and affect treatment planning decision. However, reliable landmark annotations often require experienced doctors, and even for seasoned orthodontists, manually identifying these landmarks can be a time-consuming and labor-intensive process. Hence, fully automatic and accurate landmark localization has been a long-standing area with a great deal of need.
 
In this challenge, we aim to promote the development of universal cephalometric landmark detection in lateral X-ray images. This is an extension of CL-Detection2023 challenge. In CL-Detection2023, the challenge task is to detect 38 landmarks. In CL-Detection2024, we will add the cervical vertebrae and ruler landmark detection task, which can use to examine the skeletal growth of a patient and assess the magnification errors during the scanning process, respectively. Specifically, the detection algorithm is expected to accurately locate 53 landmarks (13 soft tissue-related landmarks, 6 tooth-related landmarks, 19 skull-related landmarks, 13 cervical spine-related landmarks, 2 calibration ruler landmarks) in lateral X-ray images. Our challenge aims to provide a comprehensive benchmark for cephalometric landmark detection methods.
 
Compared to the dataset in CL-Detection2023 (600 lateral X-ray images), in CL-Detection2024, we further increase the dataset with an inclusion of data from a new medical center, resulting in a total of to 700 lateral X-ray images. The all data from this new center will serve as a test set to explore the algorithm’s domain generalization. To the best knowledge, this will be the most diverse and most landmark annotated public dataset for cephalometric landmark detection. 
 
In addition, based on the results in CL-Detection2023, we found that detection models cannot achieve a good tradeoff between segmentation accuracy and efficiency.

Thus, motivated by the FLARE 2022&2023 challenge, in CL-Detection2024, the challenge evaluation criteria are not limited to detection accuracy, but also include two new evaluation metrics: runtime and GPU memory consumption, which provide a comprehensive evaluation of detection accuracy and efficiency.
 
In summary, the CL-Detection2024 challenge has three main features:
 (1) Task: this is the first challenge for cervical vertebrae landmark detection in lateral X-ray images.
 (2) Dataset: we provide the most diverse and most meticulous lateral X-ray dataset, including 700 2D X-ray images from 4 medical centers.
 (3) Evaluation: we not only focus on detection accuracy but also detection efficiency, which are in concordance with real clinical practice and requirements.