Development of a Deep-Learning Model for Automated Detection and Quantification of Bleeding in Unilateral Biportal Endoscopic Spine Surgery

TL;DR

This paper introduces a deep-learning model that detects and measures bleeding during spine surgery, helping assess visual field impairment and support surgical decisions.

Contribution

A novel deep-learning model is developed for automated detection and quantification of intraoperative bleeding in endoscopic spine surgery.

Findings

The fine-tuned model achieved 86% accuracy in detecting bleeding presence in images.

Pixel-level segmentation showed a median Dice score of 0.79 and IoU of 0.65 for ground-truth-positive images.

Model performance exceeded 0.80 Dice score in cases with strong inter-surgeon agreement and substantial bleeding.

Abstract

Objectives: To develop and validate a deep-learning model capable of detecting and quantifying intraoperative bleeding to objectively evaluate visual field impairment in unilateral biportal endoscopic spine surgery (UBE). Methods: Overall, 223,568 still images were extracted from 20 UBE videos and used to train a U-Net++ segmentation model based on the red masks generated using hue, saturation, and value (HSV) thresholding. The model was fine-tuned using 350 manually annotated images that differentiated clinically relevant bleeding (red masks) from non-bleeding red regions (zero masks). The model performance was evaluated against 180 ground-truth images annotated by three spine surgeons, which were extracted from videos that were separate from those used for training and fine-tuning. Dice and intersection-over-union (IoU) scores were calculated, and correlation analyses were performed based on inter-annotator agreement. Results: The HSV-based model reproduced the red regions with high fidelity; however, it showed limited agreement with the ground-truth bleeding regions (median Dice = 0.57, IoU = 0.40). The fine-tuned model improved substantially. For image-wise binary classification of bleeding presence, the model achieved an accuracy of 86%, with a sensitivity of 93% and a specificity of 60%. For pixel-level segmentation performance, the model achieved a median Dice score of 0.79 and a median IoU of 0.65 on ground-truth-positive images. Dice performance exceeded 0.80 in cases with strong inter-surgeon ground-truth concordance (≥0.80) and substantial bleeding area (>20%). Conclusions: This deep-learning model can accurately detect clinically meaningful intraoperative bleeding in UBE and quantify visual field impairments in still images and surgical videos. Future applications include the evaluation of hemostatic techniques, postoperative image-based assessment of surgical quality, and real-time intraoperative bleeding alerts to support surgical decision-making.

Figures