Artificial Intelligence-Assisted Prediction of Bile Duct Bifurcation Site in Pure Laparoscopic Donor Right Hepatectomy: A Human-in-the-Loop Approach

doi:10.21203/rs.3.rs-4856893/v1

Download PDF

Article

Artificial Intelligence-Assisted Prediction of Bile Duct Bifurcation Site in Pure Laparoscopic Donor Right Hepatectomy: A Human-in-the-Loop Approach

https://doi.org/10.21203/rs.3.rs-4856893/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Accurate identification of the bile duct bifurcation site is crucial in pure laparoscopic donor right hepatectomy (PLDRH) for living donor liver transplantation. This study aimed to develop and evaluate an artificial intelligence (AI) model to predict the location of bile duct bifurcation during PLDRH procedures. In this single-institution retrospective feasibility study, we analyzed 55 PLDRH procedures performed between August 2021 and April 2022. We developed a deep learning model combining UNet with a MiT-B3 encoder, utilizing a human-in-the-loop approach. The model was trained on 150 manually annotated frames and refined using expert-reviewed pseudo-labels from an additional 901 frames. Model performance was evaluated using 5-fold cross-validation and an independent test set. The final model achieved 97% accuracy in clinical evaluation for 5-fold cross-validation and 93.3% accuracy on the independent test set. Quantitative metrics showed improvements from the initial to the final model, with mean Dice Similarity Coefficient increasing from 0.392 to 0.472 and Intersection over Union from 0.279 to 0.339. Sensitivity improved from 0.487 to 0.643, while specificity remained consistent at 0.993. This study demonstrates the potential of AI in accurately predicting the bile duct bifurcation site during PLDRH procedures. The human-in-the-loop approach proved effective in improving model performance and annotation efficiency. While challenges remain in bridging clinical accuracy and quantitative metrics, the high clinical accuracy suggests a promising step towards integrating AI into liver transplant surgery.

Health sciences/Gastroenterology/Hepatology/Liver

Biological sciences/Computational biology and bioinformatics/Machine learning

Artificial intelligence

Living donor liver transplantation

Laparoscopic hepatectomy

Bile duct

Computer vision

Liver transplantation is a life-saving procedure for patients with end-stage liver disease and hepatic malignancies^1,2. Nevertheless, the shortage of organs represents the most significant obstacle for the patients seeking to undergo this procedure³. To address this challenge, living donor liver transplantation (LDLT) has emerged as a vital solution, particularly in Eastern countries, such as South Korea⁴. Between 2019 and 2022, LDLT accounted for 5,716 cases (75.5%) out of a total of 7,570 liver transplantations performed in South Korea⁵.

However, LDLT presents a unique ethical and medical challenge: donors undergo a major surgical procedure without any direct health benefit, exposing themselves to potential risks⁶. Therefore, minimizing and managing these risks during donor surgery is of paramount importance⁷. Among the complications associated with donor hepatectomy, biliary leakage and strictures remain the most common and concerning^8,9. To mitigate these risks, significant efforts have been directed towards enhancing the understanding of biliary anatomy, such as magnetic resonance cholangiopancreatography (MRCP), intraoperative cholangiography (IOC), and indocyanine green (ICG) cholangiography^7,10.

A critical step in donor hepatectomy is the accurate prediction of the bile duct bifurcation point at the hepatic hilum. This anatomical landmark is crucial in determining the parenchymal transection plane during right or left donor hepatectomy¹¹. Inaccurate prediction of this point can lead to unnecessary manipulation during the bile duct dissection process, potentially increasing risks to the donor (Fig. 1). In recent years, computer vision (CV), a field of artificial intelligence (AI) that enables computational understanding of digital images and videos, has shown promising applications in minimally invasive surgical settings^12–14. Building on these advancements, this study aims to develop and evaluate an AI model that leverages CV techniques to predict the location of bile duct bifurcation at the hepatic hilum during pure laparoscopic donor right hepatectomy (PLDRH) procedures. Accurate localization of this critical anatomical landmark could potentially enhance surgical precision, reduce biliary complications, and improve donor safety outcomes.

Study design and patients

This single-institution retrospective feasibility study included 55 patients who underwent PLDRH between August 2021 and April 2022. The study was conducted in accordance with the Declaration of Helsinki and Istanbul, and was reviewed and approved by the Institutional Review Board of Samsung Medical Center (IRB, SMC-2022-07-149-001). Due to the retrospective nature of the study, the IRB waived the need of obtaining informed consent.

Data collection and preprocessing

High-quality surgical videos of the PLDRH procedures with a resolution of 1920 x 1080 pixels and a frame rate of 30 frames per second (fps) were reviewed. From each video, we extracted the segment showing the identification of the bile duct bifurcation at the hepatic hilum prior to parenchymal transection. Three representative frames were selected from each video, resulting in a total of 150 frames from 50 PLDRH videos. These frames were carefully chosen to ensure clear visibility of the hilum, free from obstruction by surgical instruments or fog. To establish the ground truth for the bile duct bifurcation site, we employed a reverse playback technique. By analyzing the video in reverse from the point of complete identification of the bifurcation site, we could accurately annotate the bifurcation site on the selected frames.

AI model development

This study utilized a deep learning architecture combining UNet with a MiT-B3 encoder, which has shown promising results in medical image segmentation tasks (see appendix for the detailed technical specifications of the AI model architecture, training process, and data augmentation techniques)^15,16. Our model development process incorporated a human-in-the-loop approach, which integrates expert surgical knowledge into the AI training process (Fig. 2). This approach involved initial model training on manually annotated frames, followed by pseudo-labeling of additional unlabeled data. These pseudo-labels were then reviewed by surgical experts, who discarded inappropriately labeled images and retained only those deemed accurate. This curated set of pseudo-labeled images, along with the original manually annotated data, was then used to retrain and refine the model, ensuring a high-quality dataset for the final model development.

Model evaluation

To assess the clinical utility of the AI model, we divided the hepatic hilum into five segments and evaluated the model's performance based on its ability to correctly predict the segment containing the bile duct bifurcation (Fig. 3). Additionally, we calculated quantitative metrics for segmentation tasks, including Dice Similarity Coefficient (DSC), Intersection over Union (IOU), sensitivity, and specificity¹⁷.

Clinical evaluation

The study performed a 5-fold cross-validation using 50 PLRDH surgical videos. From each video, three frames were extracted and defined chronologically as 1st, 2nd, and 3rd. The clinical evaluation revealed that the final model accurately predicted the bile duct bifurcation site in 145 out of 150 frames, achieving a 97% accuracy rate. In contrast, the initial model successfully predicted the bifurcation site in 133 out of 150 frames, demonstrating an 89% accuracy rate (Table 1).

Table 1

Clinical Accuracy Comparison of Initial and Final AI Models for Bile Duct Bifurcation Site Prediction. AI, artificial intelligence; SD, standard deviation.
	Initial model				Final model
	1st frame	2nd frame	3rd frame	Total(mean\(\:\pm\:\)SD)	1st frame	2nd frame	3rd frame	Total(mean\(\:\pm\:\varvec{S}\varvec{D}\))
1st fold	8/10 (0.8)	10/10 (1.0)	10/10 (1.0)	28/30 (0.93\(\:\pm\:\)0.12)	9/10 (0.9)	10/10 (1.0)	10/10 (1.0)	29/30 (0.97 \(\:\pm\:\)0.06)
2nd fold	8/10 (0.8)	9/10 (0.9)	10/10 (1.0)	27/30 (0.90\(\:\pm\:\)0.10)	10/10 (1.0)	10/10 (1.0)	10/10 (1.0)	30/30 (1.00\(\:\pm\:\)0.00)
3rd fold	9/10 (0.9)	6/10 (0.6)	8/10 (0.8)	23/30 (0.77\(\:\pm\:\)0.15)	9/10 (0.9)	8/10 (0.8)	9/10 (0.9)	26/30 (0.87\(\:\pm\:\)0.06)
4th fold	9/10 (0.9)	9/10 (0.9)	9/10 (0.9)	27/30 (0.90\(\:\pm\:\)0.00)	10/10 (1.0)	10/10 (1.0)	10/10 (1.0)	30/30 (1.00\(\:\pm\:\) 0.00)
5th fold	9/10 (0.9)	10/10 (1.0)	9/10 (0.9)	28/30 (0.93\(\:\pm\:\)0.06)	10/10 (1.0)	10/10 (1.0)	10/10 (1.0)	30/30 (1.00\(\:\pm\:\) 0.00)
Total	43/50 (0.86\(\:\pm\:\)0.05)	44/50 (0.88\(\:\pm\:\)0.16)	46/50 (0.92\(\:\pm\:\)0.08)	133/150 (0.89\(\:\pm\:\)0.10)	48/50 (0.96\(\:\pm\:\)0.05)	48/50 (0.96\(\:\pm\:\)0.09)	49/50 (0.98\(\:\pm\:\)0.04)	145/150 (0.97\(\:\pm\:\)0.06)

Subsequently, we applied the final model to the test set, which comprised 5 videos not used in the training process. Out of the 15 images from these videos, the model accurately predicted the bifurcation site in 14 cases (14/15, 93.3%) (Fig. 4).

Quantitative evaluation

The final model showed a mean DSC of 0.472 (SD 0.04), compared to 0.392 (SD 0.04) for the initial model. The IoU for the final model was 0.339 (SD 0.03), while the initial model had an IoU of 0.279 (SD 0.03). Sensitivity increased from 0.487 (SD 0.09) in the initial model to 0.643 (SD 0.08) in the final model. Specificity remained consistent at 0.993 (SD 0.001) for both models (Table 2).

Table 2

Quantitative Performance Metrics Comparison Between Initial and Final AI Models for Bile Duct Bifurcation Prediction. AI, artificial intelligence; SD, standard deviation; DSC, Dice similarity coefficient; IOU, intersection of union.
	Initial model						Final model
	1st fold	2nd fold	3rd fold	4th fold	5th fold	Mean (SD)	1st fold	2nd fold	3rd fold	4th fold	5th fold	Mean (SD)
DSC	0.44	0.336	0.386	0.44	0.357	0.392 (0.04)	0.508	0.418	0.425	0.508	0.504	0.472 (0.04)
IoU	0.315	0.246	0.278	0.308	0.249	0.279 (0.03)	0.366	0.288	0.307	0.365	0.371	0.339 (0.03)
Sensitivity	0.569	0.387	0.413	0.595	0.47	0.487 (0.09)	0.702	0.626	0.549	0.75	0.588	0.643 (0.08)
Specificity	0.993	0.994	0.996	0.992	0.992	0.993 (0.001)	0.993	0.992	0.994	0.992	0.995	0.993 (0.001)

This study demonstrated the potential of AI in accurately predicting the bile duct bifurcation site during PLDRH procedures. The proposed model achieved 97% accuracy in clinical evaluation for 5-fold cross-validation, and this high performance was maintained in the test set with 93.3% accuracy, suggesting robust performance on unseen data. This high accuracy in predicting the bile duct bifurcation site has significant clinical implications for PLDRH. Accurate identification of this anatomical landmark can potentially enhance surgical precision in determining the parenchymal transection plane, reduce the risk of unnecessary manipulation during bile duct dissection, and minimize the potential for biliary complications, a common concern in donor hepatectomy^8,9. Ultimately, this could lead to improved overall donor safety, which is paramount in living donor liver transplantation.

However, the quantitative evaluation showed relatively lower scores, with a mean DSC of 0.472 and IoU of 0.339, sensitivity at 0.643, and specificity at 0.993. The disparity between the high clinical accuracy and the modest quantitative metrics can be attributed to the unique challenges in identifying and labeling the bile duct bifurcation site. The common hepatic duct's entry point into the liver, while anatomically present, is often obscured by surrounding connective tissue. Even after dissection, this structure is rarely fully exposed during surgery. Our ground truth labeling method, which used circular annotations to indicate the predicted entry point, was designed to accommodate this clinical reality. However, this approach, while clinically relevant, may have led to lower quantitative scores. The circular annotation, necessary for practical identification, inherently introduces imprecision when evaluated against pixel-level segmentation metrics like DSC and IoU¹⁷.

Our human-in-the-loop approach, integrating expert surgical knowledge into the AI training process, proved highly effective in improving model performance. This method allowed for refinement of the model through expert review of pseudo-labels, elimination of inappropriately labeled images, and ensured a high-quality training dataset. The significant improvement in accuracy from the initial model (89%) to the final model (97%) underscores the value of this approach. Quantitatively, we observed improvements in key metrics, with the DSC increasing from 0.392 (SD 0.04) to 0.472 (SD 0.04), and the IOU rising from 0.279 (SD 0.03) to 0.339 (SD 0.03). This methodology not only enhanced model performance but also significantly improved the efficiency of the annotation process. While the initial pixel-level segmentation was performed on 150 frames, our approach enabled us to leverage the initial model to incorporate an additional 901 frames into the training process. This substantial increase in training data, achieved without the need for time-consuming manual annotation of each frame, demonstrates the scalability and efficiency of our human-in-the-loop approach¹⁸.

Despite the promising results, several limitations should be addressed in future studies. Firstly, the single-institution study design limits the generalizability of our findings. Secondly, the relatively small sample size, while sufficient for this feasibility study, calls for further validation with larger datasets. Additionally, the challenge of accurately representing the bile duct bifurcation site in ground truth labeling, as evidenced by the discrepancy between clinical and quantitative evaluations, requires further investigation and methodological refinement. Furthermore, the exploration of real-time application during surgery represents an exciting next step. This would involve integrating the AI model into the surgical workflow and assessing its impact on surgical decision-making and outcomes.

In conclusion, this study demonstrates the significant potential of AI in enhancing the precision and safety of PLDRH procedures. While challenges remain, particularly in bridging the gap between clinical accuracy and quantitative metrics, the high clinical accuracy of our model suggests a promising step toward integrating AI into liver transplant surgery. Future multi-center studies with larger datasets and refined evaluation methods will be crucial in validating and expanding upon these encouraging results.

Author Contribution

Conceptualization: N.Oh, J.Rhu, G-S.ChoiMethodology: N.Oh, B.KimSoftware: N.Oh, B.KimResources: J.Rhu, J.Kim, G-S.ChoiSupervision: J.Rhu, J.Kim, G-S.ChoiWriting - Original Draft: N.Oh, B.KimVisualization: N.Oh, B.KimProject Administration: G-S.ChoiFunding Acquisition: G-S.ChoiReview & Editing: All authors

Data Availability

The data that support the findings of this study are not openly available due to reasons of sensitivity and are available from the corresponding author upon reasonable request.

Financial support and sponsorship:

This study was supported by ‘Future Medicine 2030 Project’ of the Samsung Medical Center [SMX1230771].

Merion, R. M. et al. The survival benefit of liver transplantation. Am. J. Transplant. 5, 307–313. http://doi.org/10.1111/j.1600-6143.2004.00703.x (2005).
Oh, N. et al. Survival after living donor liver transplantation versus best supportive care in patients with end-stage liver disease with various MELD-Na scores: retrospective cohort study. BJS open 7, zrad127. http://doi.org/10.1093/bjsopen/zrad127 (2023).
Kwong, A. J. et al. OPTN/SRTR 2020 Annual Data Report: Liver. Am. J. Transplant. 22 Suppl 2, 204–309. http://doi.org/10.1111/ajt.16978 (2022).
Rela, M. & Rammohan, A. Why are there so many liver transplants from living donors in Asia and so few in Europe and the US? J. Hepatol. 75, 975–980. http://doi.org/10.1016/j.jhep.2021.05.036 (2021).
National Institute of Organ, Tissue and Blood Management. Annual Report of the transplant 2022. Seoul: The National Institute of Organ, Tissue and Blood Management; 2023. https://www.konos.go.kr/.
Malago, M. et al. Ethical considerations and rationale of adult-to-adult living donor liver transplantation. Liver Transpl. 7, 921–927. http://doi.org/10.1053/jlts.2001.28301 (2001).
Cherqui, D. et al. Expert Consensus Guidelines on Minimally Invasive Donor Hepatectomy for Living Donor Liver Transplantation From Innovation to Implementation: A Joint Initiative From the International Laparoscopic Liver Society (ILLS) and the Asian-Pacific Hepato-Pancreato-Biliary Association (A-PHPBA). Ann. Surg. 273, 96–108. http://doi.org/10.1097/SLA.0000000000004475 (2021).
Soubrane, O. et al. Minimally Invasive Donor Hepatectomy for Adult Living Donor Liver Transplantation: An International, Multi-institutional Evaluation of Safety, Efficacy and Early Outcomes. Ann. Surg. 275, 166–174. http://doi.org/10.1097/SLA.0000000000003852 (2022).
Rhu, J., Choi, G. S., Kim, J. M., Kwon, C. H. D. & Joh, J. W. Risk Factors Associated With Surgical Morbidities of Laparoscopic Living Liver Donors. Ann. Surg. 278, 96–102. http://doi.org/10.1097/SLA.0000000000005851 (2023).
Hong, S. K. et al. Optimal bile duct division using real-time indocyanine green near-infrared fluorescence cholangiography during laparoscopic donor hepatectomy. Liver Transpl. 23, 847–852. http://doi.org/10.1002/lt.24686 (2017).
Rhu, J., Choi, G. S., Kwon, C. H. D., Kim, J. M. & Joh, J. W. Learning curve of laparoscopic living donor right hepatectomy. Br. J. Surg. 107, 278–288. http://doi.org/10.1002/bjs.11350 (2020).
den Boer, R. B. et al. Computer-aided anatomy recognition in intrathoracic and -abdominal surgery: a systematic review. Surg. Endosc. 36, 8737–8752. http://doi.org/10.1007/s00464-022-09421-5 (2022).
Madani, A. et al. Artificial Intelligence for Intraoperative Guidance: Using Semantic Segmentation to Identify Surgical Anatomy During Laparoscopic Cholecystectomy. Ann. Surg. 276, 363–369. http://doi.org/10.1097/SLA.0000000000004594 (2022).
Oh, N. et al. Real-time segmentation of biliary structure in pure laparoscopic donor hepatectomy: an experimental study using deep learning for augmented surgery. 2024. https://www.researchsquare.com/article/rs-4611820/v1.
Xie, E. et al. SegFormer: Simple and efficient design for semantic segmentation with transformers. Advances in neural information processing systems. Adv. Neural Inf. Process. Syst. 34, 12077–12090 (2021).
Ronneberger, O., Fischer, P. & Brox, T. U-net: Convolutional networks for biomedical image segmentation. Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, October 5–9, 2015, proceedings, part III 18, 234–241 (Springer, 2015).
Taha, A. A. & Hanbury, A. Metrics for evaluating 3D medical image segmentation: analysis, selection, and tool. BMC Med. Imaging 15, 29. http://doi.org/10.1186/s12880-015-0068-x (2015).
Wu, X. et al. A survey of human-in-the-loop for machine learning. Future Gener. Comput. Syst. 135, 364–381. http://doi.org/10.1016/j.future.2022.05.014 (2022).

No competing interests reported.

appendix.docx

Download PDF

Editor invited by journal
14 Aug, 2024
Submission checks completed at journal
14 Aug, 2024
First submitted to journal
04 Aug, 2024

You are reading this latest preprint version

Artificial Intelligence-Assisted Prediction of Bile Duct Bifurcation Site in Pure Laparoscopic Donor Right Hepatectomy: A Human-in-the-Loop Approach

Status:

Version 1

Abstract

Figures

Introduction

Methods

Study design and patients

Data collection and preprocessing

AI model development

Model evaluation

Results

Clinical evaluation

Quantitative evaluation

Discussion

Declarations

Author Contribution

Data Availability

Financial support and sponsorship:

References

Additional Declarations

Supplementary Files

Status:

Version 1