A simple and effective approach for body part recognition on CT scans based on projection estimation

Abstract: It is well known that machine learning models require a high amount of annotated data to obtain optimal performance. Labelling Computed Tomography (CT) data can be a particularly challenging task due to its volumetric nature and often missing and$/$or incomplete associated meta-data. Even inspecting one CT scan requires additional computer software, or in the case of programming languages $-$ additional programming libraries. This study proposes a simple, yet effective approach based on 2D X-ray-like estimation of 3D CT scans for body region identification. Although body region is commonly associated with the CT scan, it often describes only the focused major body region neglecting other anatomical regions present in the observed CT. In the proposed approach, estimated 2D images were utilized to identify 14 distinct body regions, providing valuable information for constructing a high-quality medical dataset. To evaluate the effectiveness of the proposed method, it was compared against 2.5D, 3D and foundation model (MI2) based approaches. Our approach outperformed the others, where it came on top with statistical significance and F1-Score for the best-performing model EffNet-B0 of 0.980 $\pm$ 0.016 in comparison to the 0.840 $\pm$ 0.114 (2.5D DenseNet-161), 0.854 $\pm$ 0.096 (3D VoxCNN), and 0.852 $\pm$ 0.104 (MI2 foundation model). The utilized dataset comprised three different clinical centers and counted 15,622 CT scans (44,135 labels).

Body Part Recognition on CT Scans Using 2D Projection Estimation

The paper, titled "A Simple and Effective Approach for Body Part Recognition on CT Scans Based on Projection Estimation," presents a novel approach in the field of medical imaging by introducing a method for body region identification in computed tomography (CT) scans. This study fundamentally addresses the challenge posed by the volumetric nature of CT data, which traditionally complicates data annotation due to missing or incomplete metadata.

Methodology Overview

Researchers propose a systematic approach focusing on the estimation of 2D X-ray-like images from 3D CT volumes, subsequently classifying these into 14 distinct anatomical regions. The study contrasts this method with existing approaches such as 2.5D, 3D, and those employing foundation models like MI2, demonstrating superior performance with the 2D projection method. This experimentation involved several neural networks including DenseNet-161, EfficientNet-B0, ResNet-50, VoxCNN, R3D-18, ViT-3D, and a model based on MI2 foundation model embeddings.

Key Results

EfficientNet-B0 emerged as the best-performing model within the proposed 2D approach, achieving an F1-Score of 0.980 ± 0.016, a noteworthy improvement over its closest competitors DenseNet-161 and VoxCNN (3D approach) with F1-Scores of 0.840 ± 0.114 and 0.854 ± 0.096, respectively. The MI2 foundation model, while showing comparable performance to 2.5D and 3D methods, fell short against the EfficientNet-B0 model in statistical tests across all body regions evaluated.

Implications

The computational efficiency of the 2D approach is underscored by its significantly reduced hardware requirements, necessitating only 124.58MB RAM compared to the larger memory footprint of 2103.25MB RAM required by 3D approaches like VoxCNN. This implies a facilitation of model deployment in resource-constrained environments. Additionally, the rapid prediction times further enhance its suitability for clinical applications.

From a theoretical perspective, the study challenges the preconceived notion that more sophisticated deep learning models provide substantially better outcomes in all scenarios. The utility of the proposed simpler models indicates potential reconsideration of algorithmic complexity in certain medical imaging tasks.

Limitations and Future Directions

While promising, the proposed method has a few limitations. The scope is restricted to 14 body regions, necessitating model retraining with the addition of new classes. The reliance on bone-focused regions might limit application in scenarios where soft tissue characterization is critical. Moving forward, deeper integration with diverse clinical datasets is suggested to increase robustness and generalizability. Furthermore, adjusting ROI boundaries to incorporate broader anatomical features or soft tissue regions could enhance the method's versatility.

Conclusion

The research contributes meaningfully to the evolution of CT scan interpretation through machine learning. The demonstrated effectiveness of the 2D projection approach not only paves the way for more domain-specific models but also reflects a strategic advancement in machine learning applications for medical imaging, emphasizing simplicity and efficiency in algorithm design. Future work should focus on extending the range of identifiable regions and optimizing model retraining processes in response to evolving clinical needs, as well as exploring integration possibilities with other emerging foundation models to refine predictive capabilities across diverse datasets.