Speed and accuracy in Tandem: Deep Learning-Powered Millisecond-Level pulmonary embolism detection in CTA

Abstract

Background and Objectives

Pulmonary embolism (PE) is a critical medical condition that requires a rapid and accurate diagnosis. Traditional methods, although highly precise, have focused primarily on accuracy, neglecting the urgency of speed required in emergency settings. This study aims to develop a deep learning model that not only maintains high accuracy but also achieves millisecond-level PE detection speed.

Materials and Methods

This study employed an internal dataset comprising 160 patients from Tianjin Medical University General Hospital, and an external RSNA dataset for validation. Our model, built upon the YOLOv5 framework, was enhanced with Partial Convolution, a C2f module, and decoupled head structure.

Results

The internal test set achieved a recall of 82.5 %, precision of 84.2 %, and mean average precision (mAP) of 87.2 %, significantly outperforming the other leading models. Notably, our model provided an inference time of just 1.6 ms per image, setting a new benchmark for real-time PE detection, which was faster than YOLOv5 (2.9 ms), YOLOv6 (4.0 ms), and YOLOv8 (3.2 ms). Furthermore, our model demonstrated faster convergence and consistently lower loss values during training, achieving perfect precision at a significantly lower confidence threshold than other YOLO variants, highlighting its superior optimization and generalization capabilities.

Conclusion

This study successfully developed a deep learning model capable of millisecond-level PE detection without compromising the accuracy. Its performance underscores its potential to revolutionize PE diagnosis in emergency clinical settings, enabling timely and reliable intervention.