By combining YOLO detection, particle filtering and monocular depth estimation, researchers have identified 82 per cent of occlusions and estimated depth to within one centimetre.
In minimally invasive surgery, the instruments regularly pass in front of the area of interest and hide it from the surgeon's screen. A team led by Anuska Sarkar and Subhamoy Mandal proposes a hybrid framework to continue to monitor this area despite occultations, in real time.
Detection, Filtering, and Depth
The system combines an improved version of the YOLO detector to locate instruments, a particle filter to maintain tracking of the target region, and the MiDaS network to estimate depth from a single camera. An occultation is reported as soon as the distance between an instrument and the area of interest falls below a threshold. Instruments are detected with a confidence score of 0.85, and the mean absolute depth error reaches 0.34 cm (0.65 cm during independent phantom validation compared to stereo measurements).
The device detects 81.95% of occultations while rotating at 38.7 frames per second. Initial feedback from surgeons considers it clinically relevant, but it remains to be tested in real operating conditions before considering an intraoperative navigation aid.
Read the full article on the original source
Go to the source →