A Deep Learning Model for Blood Volume Estimation From Multi-modal Ultrasound

1. Background \& Rationale: Accurate assessment of a patient's blood volume (BV) status before surgery is critical for preventing perioperative complications. However, there is currently no clinically feasible, accurate, and non-invasive method for direct BV quantification. We hypothesize that dynamic ultrasound videos of major blood vessels contain rich, sub-visual spatiotemporal information about vascular compliance and filling that can be leveraged to estimate BV. 2. Objective: To develop and validate a deep learning model that integrates multi-modal ultrasound video data to achieve non-invasive, quantitative estimation of preoperative blood volume. 3. Study Design: A prospective, single-center, observational study. 4. Methods: Participants: Adult patients scheduled for surgery. Data Acquisition: Input (Features): Preoperative ultrasound video clips will be recorded in standardized views of four key vessels: the Internal Jugular Vein (IJV), Subclavian Vein (SCV), Inferior Vena Cava (IVC), and Common Carotid Artery (CA). Target (Label): The true Blood Volume (BV) will be calculated for each patient using the acute normovolemic hemodilution (ANH) method. The change in hemoglobin concentration before and after this process is used to calculate the total blood volume with high clinical reliability. Model Development: A hybrid deep learning architecture (e.g., CNN + LSTM/Transformer) will be trained to extract features from the ultrasound videos and learn the complex, non-linear mapping to the BV value derived from ANH. The model will be trained and internally validated using a k-fold cross-validation approach. 5. Expected Outcome \& Significance: We anticipate the development of a novel, end-to-end deep learning model capable of providing a quantitative BV estimate from routine ultrasound scans. This technology has the potential to revolutionize perioperative fluid management by offering a rapid, non-invasive, and accurate tool for ob