Towards Better Cephalometric Landmark Detection with Diffusion Data Generation

TL;DR

This work tackles data scarcity and annotation burden in cephalometric landmark detection by introducing an Anatomy-Informed Cephalometric X-ray Generation (AICG) pipeline that jointly generates X-ray images and landmark annotations. It combines the MIRA augmentation, Anatomy-Informed Topology (AIT), and a Prompt Description Generator (PDG) to condition diffusion-based generators, plus a new Prompt-CX dataset linking real X-rays with medical text prompts. Training large-scale detectors on abundant synthetic data yields substantial improvements in SDR within 2 mm (reaching 82.2%) and reduces mean radial error, demonstrating the effectiveness and generalization of diffusion-driven data augmentation for medical landmark detection. The approach enables broader deployment of accurate cephalometric analysis across diverse scanners and patient populations, with extensive ablation and expert validation supporting its robustness and clinical relevance.

Abstract

Cephalometric landmark detection is essential for orthodontic diagnostics and treatment planning. Nevertheless, the scarcity of samples in data collection and the extensive effort required for manual annotation have significantly impeded the availability of diverse datasets. This limitation has restricted the effectiveness of deep learning-based detection methods, particularly those based on large-scale vision models. To address these challenges, we have developed an innovative data generation method capable of producing diverse cephalometric X-ray images along with corresponding annotations without human intervention. To achieve this, our approach initiates by constructing new cephalometric landmark annotations using anatomical priors. Then, we employ a diffusion-based generator to create realistic X-ray images that correspond closely with these annotations. To achieve precise control in producing samples with different attributes, we introduce a novel prompt cephalometric X-ray image dataset. This dataset includes real cephalometric X-ray images and detailed medical text prompts describing the images. By leveraging these detailed prompts, our method improves the generation process to control different styles and attributes. Facilitated by the large, diverse generated data, we introduce large-scale vision detection models into the cephalometric landmark detection task to improve accuracy. Experimental results demonstrate that training with the generated data substantially enhances the performance. Compared to methods without using the generated data, our approach improves the Success Detection Rate (SDR) by 6.5%, attaining a notable 82.2%. All code and data are available at: https://um-lab.github.io/cepha-generation

Figures (10)

• Figure 1: The Comparison of Traditional and Our Method. The lack of samples and complex labeling process limit the data scale and diversity. To address this limitation, we propose a conditional data generation method that can simultaneously generate X-ray images and corresponding annotations. This approach enables incorporating large-scale detection models into this area without introducing overfitting.
• Figure 2: Overview of the Anatomy-Informed Cephalometric X-ray Generation (AICG) Framework. Three primary stages include Condition Generation (highlighted in blue), Image Generation (marked in yellow), and Landmark Detection (stroked in purple), delineating the pipeline from condition preparation through image synthesis to landmark detection.
• Figure 3: The Proposed MIRA Module. This figure depicts the transformation of real landmark labels through Global Augmentation and the rules of Anatomy-Informed Augmentation to generate a diverse and anatomically accurate position of cephalometric landmarks.
• Figure 4: The Anatomy-Informed Topology (AIT) Module: (a) Shows 38 cephalometric landmarks with their positions and names. (b) Describes the AIT module's process, highlighting the construction of a graph with critical landmarks and employing Distance-Based Coloring for intuitive anatomical relationship representation.
• Figure 5: The Cephalometric Prompt Description Generator. (a) A word cloud depicts the commonality of keywords annotated by medical experts. (b) The Prompt Description Generator module's schematic creates diverse image descriptions across three categories.