Related Experiment Videos
Domain adaptive detection framework for multi-center bone tumor detection on radiographs
Bing Li1, Danyang Xu2, Hongxin Lin3
1Medical AI Lab, School of Biomedical Engineering, Medical School, Shenzhen University, Shenzhen, China; Medical Imaging Department, The First Affiliated Hospital of Guangdong Pharmaceutical University, China.
Summary
A new domain-adaptive framework improves automatic bone tumor detection in radiographs by bridging domain gaps between different imaging centers. This enhances model generalization and accuracy for better cancer diagnosis.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oncology
Background:
- Automatic bone tumor detection on radiographs is vital for reducing bone cancer mortality.
- Model performance degrades across different imaging centers due to domain shift.
- Acquiring large annotated datasets for diverse domains is clinically challenging.
Purpose of the Study:
- To propose a domain-adaptive (DA) detection framework to address the domain gap in bone tumor radiographs.
- To improve the generalization and performance of automatic bone tumor detection models across multiple centers.
Main Methods:
- A four-part DA framework: multilevel feature alignment (MFAM), Wasserstein distance critic (WDC), instance feature alignment (IFAM), and consistency regularization (CRM).
- Incorporation of a domain discriminator with an attention mechanism to further enhance model performance.
Main Results:
- The proposed framework improved average precision (AP) at an intersection over union threshold of 0.2 (AP@20) by 1% on the source domain and 8.9% on the target domain.
- The addition of an attention-based domain discriminator further boosted AP@20 by 2% on the source and 10.7% on the target domain.
Conclusions:
- The DA framework effectively bridges the domain gap in bone tumor radiographs.
- The model demonstrates significant improvements in generalization and performance across different clinical centers, addressing a key challenge in AI-assisted diagnostics.