Robust and efficient blood loss estimation using color features and gradient boosting trees

Traditional visual methods for estimating intraoperative blood loss are often inaccurate, posing risks to patient safety. While promising, deep learning solutions are often resource-intensive and lack robust real-world validation. This paper presents an efficient and robust framework for automated blood loss estimation from surgical sponge images using color features and gradient boosting trees. We leverage a large dataset (Chiang Mai University) of 88,353 surgical sponge images, including gauze (4x4 inch; 0.5-10 ml) and swabs (15x15 inch; 5-100 ml). Our framework incorporates robust pre-processing to handle images captured under diverse imaging perspectives (e.g., top-down, high-angle), camera types (e.g. multiple resolutions, aspect ratios), and natural noise prevalent in operating rooms (e.g., feet, trays). To ensure data quality and generalizability, we implemented a filtering process, removing outliers using the Interquartile Range (IQR) method and DINO embeddings. Our final evaluation method was designed to prevent data leakage, with distinct training and test splits. The framework utilizes precise sponge segmentation followed by comprehensive color moments from multiple color spaces (RGB, HSV, Lab, Luv). Evaluated on held-out test sets, our framework demonstrates superior robustness and efficiency. For gauze sponges, we achieved a mean squared error (MSE) of 0.39, mean absolute error (MAE) of 0.45, and mean absolute percentage error (MAPE) of 10.90%. For swab sponges, results were MSE of 24.51, MAE of 3.41, and MAPE of 7.86%. Our framework is approximately 39 times faster than state-of-the-art deep learning models while maintaining comparable accuracy. This practical, highly accurate, and deployable solution offers potential for enhancing clinical decision-making and patient safety.