Poster R13 - Using Deep Learning for Gunshot Wound Bullet Fragment Detection in Emergency Medicine

Introduction: Gun violence is a public health issue in the United States, with over 48,000 deaths per year from related injuries [1]. For patients with gunshot wounds (GSW), computed tomography (CT) scans are crucial for identifying injuries sustained from ballistic trauma, including solid organ injuries, acute hemorrhage, penetrating fractures, and retained bullet fragments. However, a rapid assessment of multiple GSWs in emergency settings is clinically difficult and risks treatment delays and complications. This work aims to build and assess the capabilities of machine learning pipelines to identify bullet fragments in CT scans quickly, accurately, and precisely to facilitate trajectory determination and injury identification.

Materials and

Methods: Leveraging CT scout view images and manual annotations from 1335 patients treated for GSW injuries at Penn Medicine, an urban Level 1 trauma center, from 2019 to 2022, three pipelines built upon pre-existing architectures were trained for bullet detection and segmentation (Figure 1). Pipeline 1 employed a Mask R-CNN model [2] trained on bullet fragment annotations to detect and segment bullet fragments. Pipeline 2 employed a Mask R-CNN [2] model trained on high-density annotations to detect high-density ROIs then classify and segment bullet fragments. Pipeline 3 used a Faster R-CNN model [3] to detect high-density ROIs, two ResNet50 models [4] to classify images as coronal or sagittal and determine if they contain the patient’s head, three ResNet50 models [4] to classify ROIs as bullet fragments, and a FCN model [5] for segmentation. All models were trained from scratch with no pretrained weights using 80% of the dataset and evaluated with the remaining 20% of the dataset. Non-maximum suppression with intersection over union thresholds ranging across 0 to 0.5 was applied on model predictions of the locations of bullet fragments to obtain the final set of predictions used to evaluate the pipelines.

Results, Conclusions, and Discussions: Pipeline 3 identified the most bullet fragments at 89-91% prediction intersection of bullet annotations across varying non-maximum suppression intersection over union thresholds (76-79% for pipeline 1, 62-66% for pipeline 2), yielded 88-91% accuracy in ROI classification (63%-71% for pipeline 1, 48-59% for pipeline 2), and also provided the greatest overlap of predicted masks and annotations with a Dice coefficient of 0.488 (0.336 for pipeline 1, 0.247 for pipeline 2). Intermediate models in pipeline 3 classified images as coronal vs sagittal with 98.8% accuracy and containing the head or not with 99.3% accuracy. Notably, while pipelines 1 and 2 both utilize the same base architecture, it is clear from these metrics that they result in different performance from the inclusion of a high-density non-bullet ROI class. Additionally, the outperformance of pipeline 3 over 1 and 2 highlights the indispensability of tailored machine learning models within larger pipelines for precise bullet fragment detection and segmentation. Exploration into specific image categories where Mask R-CNN exhibits varying performance may offer insights for optimizing model efficacy and generalizability. Further fine-tuning could enhance bullet fragment detection, potentially matching the performance of radiologists and facilitating the calculation of additional metrics, such as bullet movement over time relative to major arteries, to improve clinical decision-making and expedite treatment in emergency settings.

References:
1. Pew Research Center. What the data says about gun deaths in the U.S. Pew Research Center. Apr 26, 2023 [cited Apr 10, 2024]. Available from: https://www.pewresearch.org/short-reads/2023/04/26/what-the-data-says-about-gun-deaths-in-the-u-s/.
2. He K, Gkioxari G, Dollár P, Girshick R. Mask R-CNN. Proceedings of the IEEE International Conference on Computer Vision. 2017; p.2961-2969. https://arxiv.org/abs/1703.06870.
3. Ren S, He K, Girshick R, Sun J. Faster R-CNN: Towards real-time object detection with region proposal networks. Advances in Neural Information Processing Systems. 2015; p.91-99. https://arxiv.org/abs/1506.01497.
4. He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2016; p.770-778. https://arxiv.org/abs/1512.03385.
5. Long J, Shelhamer E, Darrell T. Fully convolutional networks for semantic segmentation. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2015; p.3431-3440. https://arxiv.org/abs/1411.4038.

Acknowledgements (Optional):