Let’s Stop Bleeding! Precise Bleeding Data Estimation & Visualization Methods for Laparoscopic Surgeries
Keywords: Medical imaging, surgical image, GAN, image to image translation, segmentation
TL;DR: SBAM (Selective Bleeding Alert Map) is a GAN-based framework for real-time detection of bleeding sources during surgery. Using the SBAM framework, trained on a silicone-derived dataset only, SBAM can detect clinical bleeding regions with source.
Abstract: Intraoperative bleeding remains a significant challenge in modern surgery, necessitating rapid and accurate localization of bleeding sources to ensure effective hemostasis. Proactive detection and timely intervention are critical for minimizing blood loss, reducing operative time, preventing complications, and decreasing the need for intensive postoperative care. In this research, we introduce Selective Bleeding Alert Map (SBAM), a novel GAN-based framework designed for precise real-time detection of bleeding origins during surgery. Building upon our earlier BAM framework, SBAM shifts from broad, area-wide alerts to a focused approach that highlights only the exact bleeding areas, enhancing visual accuracy and potentially improving surgeon focus and visibility—particularly beneficial in cases of minor bleeding where excessive alerts could interfere with the surgical process. To achieve this, we developed advanced image-to-image translation and segmentation models, custom thresholding techniques, and trajectory detection algorithms to pinpoint bleeding sources with high precision. Utilizing our developed mimic organ system for ethically sourced, realistic datasets—alongside synthetic data generated from the orGAN system and Large Mask Inpainting (LaMa)—we created a dedicated dataset specifically for SBAM training, including over 1,000 manually annotated images capturing both bleeding and non-bleeding regions within marked bleeding areas. Our instance segmentation model achieved a precision of 92.5%, an accuracy of 98% and a mask mean Average Precision of 85% at an IoU threshold of 0.5 (mAP@50). Additionally, the SBAM model demonstrated high accuracy in detecting bleeding points within real surgical videos from the Hamlyn dataset, underscoring its potential for practical surgical applications.Powered by core algorithms and uniquely developed datasets, SBAM represents a pivotal advancement in AI-assisted surgery, demonstrating superior performance in detecting bleeding regions with high precision during critical scenarios.
Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 8498
Loading