Enhanced Blood Loss Quantification Using a Novel Multi-Wavelength Imaging Approach

Current clinical methods for quantifying blood loss have limitations, often leading to inaccurate assessments, such as overestimations caused by non-blood fluids mixing with blood. Given blood’s ability to absorb Near Infrared (NIR) light, this wavelength is widely utilized in blood spectroscopy and advanced techniques incorporating computer vision and machine learning. This study then investigates a multi-wavelength approach that extracts and integrates features from images captured under visible (Vis) light, NIR light, and a combination of both (Vis-NIR), which leverages unique information provided by individual wavelengths. The findings reveal that using color features from the multi-wavelength method significantly reduced errors by 56% compared to relying on a single wavelength. Further incorporating spectral and spatial features alongside hyperparameter tuning decreased errors by an additional 23%. A mean absolute percentage error (MAPE) of 8.475% was achieved, falling well within the acceptable error threshold of 20% in clinical blood loss estimations. Lastly, an individualized interpretation for each surgical test gauze is provided, a useful tool to understand how the machine learning model predicts the absorbed blood volume.