Medical Student University of Illinois College of Medicine
Introduction: : Endoscopic spinal surgery (ESS) is an approach for minimally invasive surgery to treat spinal pathologies. The approach favors fast recovery and fewer perioperative complications. ESS is conducted in a fluid-filled cavity, where minor bleeds can obstruct visualization of the surgical field. This limitation of ESS slows the procedure compared to its open equivalent. The increase in surgical time and continuous visual obstruction makes adoption of the unilateral biportal endoscopic (UBE) technique more difficult for neurosurgeons. This study details a machine-learning model developed for bleeding point detection in ESS, for the purpose of assisting surgeons in quickly identifying and resolving bleeding points before they obscure the visual field.
Methods: The machine learning model was trained using 121 frames from 5 recordings of ESS at a major academic medical center. We performed an analysis of 20 images sourced from 3 endoscopic surgeries with the novel deep learning and convoluted neural network (DL-CNN) model. A one-sided t-test was conducted to analyze the distance of the true bleeding point to the detected bleeding point with a significance level of 0.05.
Results: Of the 20 test images, 80% detected the bleeding point within 5 mm, and the average distance was 1.1312 mm (p value 9.7456e-09). The model achieved 95.56% accuracy in determining the background from the bleed point.
Conclusion : The CNN correctly identified the majority of the bleeding points with statistical significance, indicating that our model is able to accurately identify bleeding points in ESS. Further efficacy studies used intraoperatively will be needed to determine clinical efficacy of this methodology.
.jpg)
