An Enhanced YOLOv8 model for Needle Tip Recognition in Ultrasound Images for Prostate Brachytherapy

Robotic-assisted medical procedures require precise needle tip detection. This is especially important in treatments like prostate brachytherapy. Traditional ultrasound-based manual detection methods have several drawbacks. These include low-resolution imaging, challenges posed by complex tissue conditions, and heavy reliance on operator skill for accurate recognition. We developed an improved needle detection system based on YOLOv8 to solve these problems. Our system combines three key components with the YOLOv8 framework. These include Depthwise Separable Convolution, Coordinate Attention, and BoTNet-Transformer modules. These additional modules enhance the model’s detection capabilities and processing speed. We test the system using a dual robotic arm setup. One arm controls the needle while the other moves an ultrasound device to scan tissue. We conducted tests using both phantom datasets and real-time experiments. The results show our modified model performs better than standard YOLOv8 in all key metrics. These improvements are consistent in both static and dynamic conditions. The system maintains accuracy despite common challenges like tissue movement and limited scanning areas.