X-YOLO: A method for detecting wrist fractures in children based on dynamic feature enhancement and lightweight design

Objectives:

Our goal here is to introduce a nifty, multi-task wrist fracture detection technique with a leaner footprint, leveraging an upgraded YOLO11s setup. This approach tackles the tricky balance between accuracy and efficiency in real-world medical settings.

Methods:

We have got our hands on the HGNetV2 module, which gets our backbone network trim and tight by cutting down on parameters and computational bloat. The DySample module joins the neck network to up our game in pinpointing tiny fractures and bone issues, all thanks to its dynamic sampling strategy that boosts multi-scale feature mapping. We swapped out the regular Conv blocks for GSConv in the neck network, which not only slashes the computational load but also keeps the info flowing smoothly. And to top it off, we have concocted the Focaler-CIoU loss function, a tweak on the Focaler-IoU that gives priority to different samples and sharpens the model’s learning across various scales.

Results:

We have tested our system on GRAZPEDWRI-DX and FracAtlas datasets, and the improvements are nothing short of spectacular. Our model now clocks in at a mere 7.0M parameters, a 25.5% shrink from the original YOLO11s. Ablation studies show that our X-YOLO model slashes computational costs by a hearty 21.5% without missing a beat, hitting [email protected] and [email protected]:0.95 scores of 65% and 41.8%, respectively. This confirms that our lightweight design and dynamic feature-boosting strategies are the real deal.

Conclusion:

In summary, we have crafted a detection framework that is a perfect fit for pediatric wrist X-rays, thanks to the clever combo of HGNetV2, DySample, GSConv, and Focaler-CIoU. Our model has a keen eye for microfractures and bone lesions, all while keeping the computational footprint small and the latency low.