Abstract: Abstract: Dynamic and precise estimation of blood loss is important for perioperative management, while the measurement of intraoperative blood loss is a difficult task, especially when the blood is absorbed by medical gauze. In view of the above situation, a deep multitask Gaussian process regression (DMGPR) method based on DenseNet is proposed to estimate the intraoperative blood loss and hemoglobin loss by taking the blood-soaked medical gauze images as the research object. DMGPR consists of two parts: DenseNet, a dense connected convolutional network for automatic feature extraction, and a multi task Gaussian regression process (MGPR) for estimating blood loss and hemoglobin loss. Under normal lighting conditions in the operating room, 569 images of blood-soaked gauze were collected to construct an experimental dataset. Taking coefficient of determination (R2), mean squared error (MSE) and mean absolute error (MAE) as performance indicators, the DMGPR method was evaluated and compared. In terms of blood loss estimation, the R2, MSE, and MAE of the DMGPR method are 0.971, 0.080, and 0.151, respectively. And in terms of estimating hemoglobin loss, the corresponding results of the DMGPR method are 0.950, 0.217, and 0.292, respectively. The experimental results indicate that DMGPR can dynamically and accurately estimate intraoperative blood loss and hemoglobin loss, and has better performance than other compared methods, making it more suitable for surgeries that mainly use medical gauze and small to moderate blood loss.