HemoSet: The First Blood Segmentation Dataset for Automation of Hemostasis Management

TL;DR

HemoSet addresses the lack of annotated data for automated hemostasis by introducing the first surgically grounded blood segmentation dataset collected from live porcine thyroidectomy. The authors benchmark five state-of-the-art segmentation models, finding that although UNet++ performs best, all models fall short of human labeling consistency and struggle with the irregular pooling geometry of blood pools. The dataset comprises 11 videos with 102,616 frames and 857 labeled frames at 640×480/30 FPS, with STAPLE-based labeling quality around 0.933 precision and 0.996 specificity. This resource aims to enable development of autonomous suction tools, blood loss estimation, and surgical guidance, with future work extending to depth imaging and greater pool diversity to enhance generalization in real-world operating rooms.

Abstract

Hemorrhaging occurs in surgeries of all types, forcing surgeons to quickly adapt to the visual interference that results from blood rapidly filling the surgical field. Introducing automation into the crucial surgical task of hemostasis management would offload mental and physical tasks from the surgeon and surgical assistants while simultaneously increasing the efficiency and safety of the operation. The first step in automation of hemostasis management is detection of blood in the surgical field. To propel the development of blood detection algorithms in surgeries, we present HemoSet, the first blood segmentation dataset based on bleeding during a live animal robotic surgery. Our dataset features vessel hemorrhage scenarios where turbulent flow leads to abnormal pooling geometries in surgical fields. These pools are formed in conditions endemic to surgical procedures -- uneven heterogeneous tissue, under glossy lighting conditions and rapid tool movement. We benchmark several state-of-the-art segmentation models and provide insight into the difficulties specific to blood detection. We intend for HemoSet to spur development of autonomous blood suction tools by providing a platform for training and refining blood segmentation models, addressing the precision needed for such robotics.

Figures (6)

• Figure 1: HemoSet is collected from a thyroidectomy performed on a porcine model. The dataset is collected by identifying blood vessels to induce hemorrhaging during the procedure. A layout of the procedure (top) and example labeled images from the dataset (bottom).
• Figure 2: The percentage of blood coverage over labeled frames alongside heatmaps of blood across trials 6, 10, and 11 (without augmentation). Our dataset is distinct from previous works in the high variation of blood coverage. The controlled environment of previous datasets leads to poor generalization when moving towards real-world applicability. Our dataset has the variety in blood geometry necessary to train a practical blood-detection algorithm.
• Figure 3: Our image provided to labelers for the annotation comparison. We point out areas to include (green, white) and areas to avoid from blood stains (black, pink) and the suction tool (yellow). With our guidelines, we achieve a STAPLE precision of 0.933 and a specificity of 0.996.
• Figure 4: IoU of each model over trial 9, with a 10-wide running mean window. We also include sample images from each section of the trial, noting the first and third quartiles of the percentages of blood coverage. Although UNet++ performs the best, it is still with difficulty comparable to the other models. We note the performance variance over the dataset, correlated with blood coverage, thus demonstrating the necessity for more nuanced architectures.
• Figure 5: From top to bottom: Image w/GT, UNet, UNet++, DeepLabV3+, MANet, Segformer. On images from the 3rd (left) and 9th (right) trials in our dataset.