Related Experiment Video
Updated: Dec 14, 2025
10:23
Author Spotlight: Three-Dimensional Cephalometric Landmark Annotation Demonstration on Human Cone Beam Computed Tomography Scans
Published on: September 8, 2023
3.5K
LandmarkNet: a 2D digital radiograph landmark estimator for registration
Zhen Wang1, Cong Liu2, Longhua Ma2
1College of Control Science and Engineering, Zhejiang University, Yugu Road, Hangzhou, 310013, China.
BMC Medical Informatics and Decision Making
|July 23, 2020
Summary
This study introduces LandmarkNet, a novel deep learning model for accurate landmark detection in blurred medical images. This advancement improves image-guided radiotherapy (IGRT) precision for better cancer treatment outcomes.
Area of Science:
- Medical Imaging
- Radiotherapy
- Artificial Intelligence
Background:
- Radiation therapy demands high precision to target tumors while sparing healthy tissues.
- Image guidance is crucial for modern radiotherapy, but registering low-quality megavoltage digital radiographs (MV-DRRs) with kilovoltage digital reconstructed radiographs (KV-DRRs) is challenging.
- Registering based on anatomical landmarks offers a simplified approach to overcome image quality limitations.
Purpose of the Study:
- To develop an accurate and automated method for landmark detection in medical images for improved image-guided radiotherapy (IGRT).
- To address the difficulties in registering low-quality MV-DRRs by focusing on precise landmark identification.
Main Methods:
- Proposed LandmarkNet, a novel keypoint estimation architecture utilizing a dual Feature Pyramid Network (FPN) for enhanced feature extraction and landmark localization.
- Employed intermediate supervision to ensure stable parameter updates during training.
- Generated heatmaps for approximate landmark locations, refined by non-maximum suppression (NMS) for precise estimation.
Main Results:
- Achieved high accuracy in landmark estimation across multiple anatomical points: spinous process (81.24% PCK), tracheal bifurcation (98.95% PCK), and Louis angle (85.61% PCK).
- Demonstrated minimal mean deviation between predicted and ground truth landmark locations (2.38, 0.98, and 2.64 pixels, respectively).
- Validated the model's performance on a dataset from multiple cancer hospitals.
Conclusions:
- LandmarkNet provides highly accurate landmark estimation for various anatomical features, particularly excelling with the distinct tracheal bifurcation.
- The model's robust performance in both quantity and position estimation for the spinous process is notable.
- This method significantly aids doctors in image-guided radiotherapy (IGRT), enabling more precise cancer treatments.
