A robust dataset for surgical instrument detection to aid autonomous robotic surgery

Robot-assisted surgery (RAS) is transforming modern healthcare by enhancing precision, reducing human error, and improving patient outcomes. A crucial step toward fully autonomous robotic surgery is the accurate and real-time recognition of surgical instruments. In this work, we present a comprehensive surgical instrument dataset named as SID-RAS which comprises of 6000 high resolution images categorized into nine distinct classes: cotton, episiotomy scissors, forceps, gloves, hemostats, mayo, scalpel, stitch scissors, and syringe. To ensure dataset's diversity and simulate real world surgical scenarios, multiple augmentations were applied, including motion blur, varying lighting conditions (low light and high brightness), simulated blood stains, and 360-degree rotation. The dataset was evaluated using YOLOv10 (nano, small, medium) and YOLOv11 (nano, small, medium) object detection models, aiming to assess their effectiveness in recognizing and localizing surgical instruments in real-time. On an average, the models have attained 99.3% of mean Average Precision (mAP) and 99.2% F1-score, demonstrating the quality of SID-RAS dataset for surgical tool detection. These findings contribute to the preliminary development of AI-driven robotic surgical assistance systems, which can be extended to various types of surgeries.