Traumatic Abdominal Injury Detection on CT Scans with Ensemble Learning on Transformer-Based Models

Traumatic abdominal injuries are challenging to detect on CT scans due to complex and variable injury patterns. Transformer-based models have shown remarkable contributions in medical image analysis; however, their diversity in architectures and modes makes selecting and fine-tuning the most suitable model a tedious and non-trivial task. To address this challenge, ensemble learning offers a robust strategy by fusing the predictions of multiple transformer models into a unified framework, thereby mitigating the limitations of relying on a single architecture. In this study, we apply ensemble learning strategies to leading transformer-based models-ViT, CvT, and Swin Transformer-on the RSNA Abdominal Traumatic Injury CT dataset. Our results show that ensemble methods consistently outperform individual models' performance. While the Swin Transformer achieved the highest single-model F1 score ($\mathbf{0. 7 9 4}$) for solid organ injury detection, stacked ensembles, particularly with a Random Forest meta-learner, attained a superior F1 score of 0.861 on the same test set. These findings demonstrate that ensemble techniques can effectively harness the strengths of diverse transformer architectures, advancing automated trauma detection.