A simple and effective approach for body part recognition on CT scans based on projection estimation
Franko Hrzic1,2, Mohammadreza Movahhedi3, Ophelie Lavoie-Gagne3
1Musculoskeletal Digital Innovation and Informatics (MDI²) Program, Department of Orthopaedic and Sports Medicine, Boston Children's Hospital, Harvard Medical School, Boston MA, USA. franko.hrzic@uniri.hr.
Scientific Reports
|August 28, 2025
Summary
This study introduces a novel 2D X-ray-like estimation method for identifying body regions in 3D Computed Tomography (CT) scans. This approach significantly improves medical dataset annotation accuracy, outperforming existing 2.5D, 3D, and foundation models.
Area of Science:
- Medical Imaging
- Machine Learning
- Data Science
Background:
- Machine learning models require extensive annotated data for optimal performance.
- Labeling Computed Tomography (CT) data is challenging due to its 3D nature and often incomplete metadata.
- Accurate body region identification is crucial for high-quality medical datasets.
Purpose of the Study:
- To propose a simple and effective method for body region identification in 3D CT scans using 2D X-ray-like estimations.
- To enhance the construction of high-quality medical datasets by accurately identifying 14 distinct body regions.
- To compare the proposed method's effectiveness against existing 2.5D, 3D, and foundation model approaches.
Main Methods:
- A novel approach utilizing 2D X-ray-like estimations derived from 3D CT scans was developed.
- The method was employed to identify 14 distinct anatomical body regions.
- Performance was evaluated against 2.5D DenseNet-161, 3D VoxCNN, and MI2 foundation model using an F1-Score metric.
Main Results:
- The proposed 2D estimation method significantly outperformed 2.5D, 3D, and foundation models in body region identification.
- The best-performing model, EffNet-B0, achieved a statistically significant F1-Score of 0.980 ± 0.016.
- This represents a substantial improvement over the F1-Scores of 0.840 ± 0.114 (2.5D), 0.854 ± 0.096 (3D), and 0.852 ± 0.104 (MI2).
Conclusions:
- The 2D X-ray-like estimation technique offers a highly effective and accurate solution for body region identification in CT scans.
- This method provides valuable information for creating robust and high-quality medical imaging datasets.
- The approach demonstrates superior performance and statistical significance compared to existing methods, paving the way for improved medical AI applications.
