Real-time and accurate estimation of surgical hemoglobin loss using deep learning-based medical sponges image analysis

Kai Li, Zexin Cheng, Jun‐Jie Zeng, Ying Shu, Xiaobo He, Hui Peng, Yongbin ZhengScientific 2023 · DOI

Abstract

Real-time and accurate estimation of surgical hemoglobin (Hb) loss is essential for fluid resuscitation management and evaluation of surgical techniques. In this study, we aimed to explore a novel surgical Hb loss estimation method using deep learning-based medical sponges image analysis. Whole blood samples of pre-measured Hb concentration were collected and diluted to simulate varying levels of Hb concentration across 851 blood-soaked sponges. A deep learning model based on the YOLOv5 network was used for target region extraction, and three models (Feature extraction, ResNet-50, and SE-ResNet50) were trained to predict surgical Hb loss. The SE-ResNet50 model achieved the best performance ($R^2 = 0.99$, MAE = 11.09 mg, MAPE = 8.6%), with Bland-Altman analysis showing a bias of 1.343 mg. This method proves feasible for real-time estimation of surgical Hb loss, aiding clinical decisions and technical evaluation.

Introduction

Accurate assessment of blood loss during surgery is a cornerstone of patient safety, directly influencing decisions regarding fluid resuscitation and blood transfusions. Traditionally, clinicians have relied on visual estimation—which is notoriously subjective and prone to error—or gravimetric analysis. While weighing sponges is more objective, it is time-consuming and often inaccurate because sponges absorb non-blood fluids like saline, irrigation, or ascites.

There is a critical clinical need for a method that measures actual hemoglobin mass rather than total fluid volume in real-time. This study proposes a computer-vision-based approach to automate this process using deep learning to analyze images of discarded medical sponges.

Methodology

The researchers developed a multi-stage pipeline for hemoglobin estimation. First, the YOLOv5 (You Only Look Once) object detection framework was employed to identify and crop the sponge from the background, ensuring the predictive models focused solely on the blood-soaked region.

To predict the hemoglobin mass, three architectures were compared. The most advanced model, SE-ResNet50, integrated Squeeze-and-Excitation (SE) blocks into a standard ResNet-50 backbone. These SE blocks act as an attention mechanism, allowing the network to perform dynamic channel-wise feature recalibration. This helps the model prioritize the most informative visual features related to hemoglobin concentration while ignoring noise from lighting or sponge texture.

The Squeeze-and-Excitation (SE) block architecture used to enhance feature recalibration within the residual learning framework.

Dataset and Training

A controlled dataset of 851 blood-soaked sponges was created using whole blood samples (Hb 50–170 g/L) diluted with normal saline. The data was split 7:3 into training (n = 595) and testing (n = 256) groups. Five-fold cross-validation was used during training to ensure model robustness.

Results and Performance

The SE-ResNet50 model significantly outperformed both traditional feature extraction methods and the standard ResNet-50. It achieved a near-perfect correlation with actual hemoglobin levels ($R^2 = 0.99$). The Mean Absolute Error (MAE) was 11.09 mg, representing a Mean Absolute Percentage Error (MAPE) of only 8.6%.

Bland-Altman analysis further confirmed the model's clinical reliability. The analysis revealed a minimal mean bias of 1.343 mg with narrow limits of agreement (-29.81 to 32.5 mg), indicating that the model remains consistent across a wide range of blood saturation levels.

Bland-Altman plot for the SE-ResNet50 model demonstrating high agreement between predicted and actual hemoglobin loss with minimal bias.

Beyond accuracy, the system demonstrated high computational efficiency. With an average processing time of 0.027 seconds per image (approximately 37 frames per second), the model is well-suited for real-time integration into surgical workflows.

Discussion and Conclusion

This study demonstrates that deep learning, particularly architectures utilizing attention mechanisms like SE-ResNet50, can effectively model the complex, non-linear relationship between the visual appearance of a blood-soaked sponge and its actual hemoglobin content.

While the results are promising, future work should expand the dataset to include a wider variety of sponge materials and real-world surgical contaminants such as bone fragments or smoke debris. Nonetheless, the high accuracy and real-time processing speed suggest that this technology could be integrated into "smart" operating rooms to provide surgeons with immediate, objective data on patient blood loss, ultimately improving surgical outcomes and resource management.

Summary updated

Abstract

Real-time and accurate estimation of surgical hemoglobin (Hb) loss is essential for fluid resuscitation management and evaluation of surgical techniques. In this study, we aimed to explore a novel surgical Hb loss estimation method using deep learning-based medical sponges image analysis. Whole blood samples of pre-measured Hb concentration were collected, and normal saline was added to simulate varying levels of Hb concentration. These blood samples were distributed across blank medical sponges to generate blood-soaked sponges. Eight hundred fifty-one blood-soaked sponges representing a wide range of blood dilutions were randomly divided 7:3 into a training group (n = 595) and a testing group (n = 256). A deep learning model based on the YOLOv5 network was used as the target region extraction and detection, and the three models (Feature extraction technology, ResNet-50, and SE-ResNet50) were trained to predict surgical Hb loss. Mean absolute error (MAE), mean absolute percentage error (MAPE), coefficient (R<sup>2</sup>) value, and the Bland-Altman analysis were calculated to evaluate the predictive performance in the testing group. The deep learning model based on SE-ResNet50 could predict surgical Hb loss with the best performance (R<sup>2</sup> = 0.99, MAE = 11.09 mg, MAPE = 8.6%) compared with other predictive models, and Bland-Altman analysis also showed a bias of 1.343 mg with narrow limits of agreement (- 29.81 to 32.5 mg) between predictive and actual Hb loss. The interactive interface was also designed to display the real-time prediction of surgical Hb loss more intuitively. Thus, it is feasible for real-time estimation of surgical Hb loss using deep learning-based medical sponges image analysis, which was helpful for clinical decisions and technical evaluation.
BibTeX
@inproceedings{li2023real,
  author = {Kai Li and Zexin Cheng and Jun‐Jie Zeng and Ying Shu and Xiaobo He and Hui Peng and Yongbin Zheng},
  title = {{Real-time and accurate estimation of surgical hemoglobin loss using deep learning-based medical sponges image analysis}},
  booktitle = {Scientific},
  year = {2023},
  doi = {10.1038/s41598-023-42572-6},
}
Figure