Deep learning implementation for extrahepatic bile duct detection during indocyanine green fluorescence-guided laparoscopic cholecystectomy: pilot study

Abstract

Background

A real-time deep learning system was developed to identify the extrahepatic bile ducts during indocyanine green fluorescence-guided laparoscopic cholecystectomy.

Methods

Two expert surgeons annotated surgical videos from 113 patients and six class structures. YOLOv7, a real-time object detection model that enhances speed and accuracy in identifying and localizing objects within images, was trained for structures identification. To evaluate the model's performance, single-frame and short video clip validations were used. The primary outcomes were average precision and mean average precision in single-frame validation. Secondary outcomes were accuracy and other metrics in short video clip validations. An intraoperative prototype was developed for the verification experiments.

Results

A total of 3993 images were extracted to train the YOLOv7 model. In single-frame validation, all classes’ mean average precision was 0.846, and average precision for the common bile duct and cystic duct was 0.864 and 0.698 respectively. The model was trained to detect six different classes of objects and exhibited the best overall performance, with an accuracy of 94.39% for the common bile duct and 84.97% for the cystic duct in video clip validation.

Conclusion

This model could potentially assist surgeons in identifying the critical landmarks during laparoscopic cholecystectomy, thereby minimizing the risk of bile duct injuries.

Introduction

Bile duct injury (BDI) remains a serious complication, occurring in 0.2–1.5% cases^1,2 during laparoscopic cholecystectomy (LC). Misinterpretation of biliary anatomy contributes to 71–97% of BDIs^3,4. Intraoperative extrahepatic bile duct visualization techniques have been implemented, such as intraoperative ultrasonography and cholangiography, but its use is often limited to selected cases due to the prolonged surgery duration, requirement for additional medical resources, and need for specific surgical expertise^5,6.

Fluorescence-guided surgery with indocyanine green (ICG) has emerged as a promising technique for enhancing visualization of biliary structures, and the visualization rates of the common bile duct (CBD) can reach 66% before the dissection of Calot’s triangle and 94% afterwards⁷. Enhanced contrast and delineation of biliary structures may enable surgeons to avoid variability in fluorescence intensity.

The integration of artificial intelligence (AI) into surgical laparoscopy has demonstrated considerable promise, particularly in LC. Several studies have focused on using semantic segmentation to identify ‘go’ and ‘no-go’ zones^8,9 as well as to automatically assess the critical view of safety (CVS) during LC^10–12. A study reported low average precision (AP) scores of 0.320 for the CBD and 0.074 for the cystic duct¹³. The challenges stem from factors like background noise, variable image quality, and the presence of multiple objects in the field, intra-abdominal fat and fibrous tissues, complicating accurate demarcation.

The aim of this study was to combine the benefits of fluorescence-guided surgery and deep learning by training YOLOv7 models to accurately identify anatomical landmarks during LC, with a specific focus on extrahepatic bile duct identification. The study also aimed to assess the performance of these models under varying surgical difficulties while developing a prototype landmark indication system for use in the operating room.

Methods

Data set preparation

This study was approved by the Institutional Review Board of Chang Gung Medical Foundation (registration date: 9 June 2022; registration number: 202200847B0). Videos were collected from patients who underwent ICG fluorescence-guided LC between June 2022 and April 2023. Patients who had subtotal cholecystectomy or required conversion to open surgery were excluded. All patients received 2 ml of intravenous ICG 15 min before the operation. Four surgeons, following a standardized surgical protocol, performed the procedures. The Nassar operative difficulty scale¹⁴, based on operative findings of gallbladder, cystic pedicle, and associated adhesions, was recorded for each patient.

From each LC video, frames were extracted at a rate of 1 frame per second from the moment the gallbladder was grasped before the dissection of the hepatocystic triangle, continuing until the transection of the cystic structures was complete. Subsequently, 35–40 frames were selected from each patient by two expert surgeons. Frames where the Calot’s triangle was completely out of view, or when the image was obscured by factors such as mist, blood, or poor focus, images during adhesiolysis or gallbladder take-down, were excluded to avoid introducing noise. The data set was randomly stratified at the patient level into training, validation, and testing sets, with the distribution of frames approximately following an 8:1:1 ratio.

Data annotations and augmentation

Six intraoperative structures, including liver, gallbladder, CBD, instrument, cystic artery, and cystic duct, were annotated by two experienced surgeons with over 200 LC procedures. A senior surgeon, with over 500 LC procedures, validated the annotations. During the YOLOv7 training process, data augmentation techniques, including flipping, cropping, rotation, and adjustments to saturation and brightness, were applied to ensure diverse representation of surgical scenarios.

Deep learning model training and performance validation

A modern deep learning model, the reliable object detector You Only Look Once (YOLOv7), was applied. This model was trained using a dataset comprising LC images. YOLOv7 was selected over several competing object detectors because of its notable advancements in speed and accuracy, mainly attributed to its newly designed trainable bag-of-freebies methods, re-parameterized module, and dynamic label assignment strategy¹⁵. Its architecture facilitates rapid detection of complex geometric relationships between various anatomical structures, rendering it ideal for real-time surgical applications. Three models were trained to determine the most effective CBD identification method (the complete workflow is illustrated in Fig. 1).

Evaluation of model performance

Two validation methods, single-frame validation and short-video clip validation, were employed to evaluate the YOLOv7 model’s performance on still images and videos.

Single-frame validation

To evaluate the performance of the YOLOv7 model, predictions for the presence and localization of anatomical structures in test images were generated using bounding boxes and compared to the ‘ground truth’ annotations provided by hepato-pancreato-biliary (HPB) experts. The model's performance was assessed using precision, sensitivity, AP, mean average precision (mAP), and the F1-score, all of which were automatically computed by the model.

Precision, also referred to as positive predictive value (PPV) in medical statistics, measures the percentage of correctly identified anatomical structures (true positives) out of all the structures the model predicts as positive. Sensitivity, also known as recall, assesses how well the model detects all actual anatomical structures present in the image, capturing the proportion of true positives among all relevant instances. The F1-score provides a balanced measure by combining both precision and sensitivity, offering insight into the model’s ability to minimize false positives and false negatives.

For each structure, AP summarizes the model’s accuracy across different thresholds, representing the ability to correctly identify a specific structure (for example CBD). mAP is the AP across all structures, providing an overall evaluation of the model’s performance. The mAP is calculated with an overlap criterion (IoU threshold of 0.5), meaning the model’s predicted boundaries must overlap with the ground truth by at least 50% for a correct match (detailed description in the Supplementary Methods section).

Short-video clip validation

In practical surgical applications, the YOLOv7 model's ability to accurately detect the duct in portions of a video above a particular threshold can provide valuable information for image recognition during surgery. The model's performance was evaluated in real-world applications from a different perspective than single-frame validation. In this context, the 6-class model was applied for short video clip validation.

The evaluation methodology was as follows: Each patient's surgical video was edited into a 2–3 min key video and segmented into 5-second clips using computer vision techniques. The 6-class YOLOv7 model performed inferences on these clips, with only the bounding boxes for the CBD and cystic duct retained for analysis. The number of positive detections for each extrahepatic bile duct was evaluated. If detections exceeded a threshold of 10 per clip, the visibility of the extrahepatic bile duct was marked as ‘positive,’ otherwise as ‘negative.’ Two HPB surgeons validated these findings, and the model’s performance was assessed using accuracy, sensitivity, specificity, and F1-score.

Results

This study analysed ICG fluorescence-guided LC videos from 113 patients. Indications for LC included gallbladder polyps (4.42%, n = 5), gallbladder stones with biliary colic (28.32%, n = 32), acute cholecystitis (17.7%, n = 20), chronic cholecystitis (29.2%, n = 33), choledocholithiasis (19.47%, n = 22), and biliary pancreatitis (0.88% n = 1). Laparoscopic common bile duct exploration was required in four patients (3.54%, n = 4) due to unsuccessful preoperative endoscopic stone extraction. Postoperative complications included wound infections (3.53%, n = 4), intra-abdominal abscesses (2.65%, n = 3), and bile leakage (1.76% n = 2), with no CBD injuries. Nassar grading indicated 15.92% (n = 18), 55.7% (n = 63), 22.1% (n = 25), and 6.1% (n = 7) of patients as grades I, II, III, and IV respectively. Patient characteristics are shown in Table 1.

The performance of the three evaluated models using the single-frame validation method was presented by AP bar chart (Fig. 2 and Table 2). The first model was trained with all six classes: liver, gallbladder, CBD, surgical instruments, cystic artery, and cystic duct. The second model excluded the cystic duct and cystic artery to address the imbalance in size and frequency of these structures. The third model focused solely on the extrahepatic bile ducts, using only CBD and cystic duct for training.

The first model achieved a mAP of 0.846. Within this model, the AP for the CBD reached 0.864, and the cystic duct had an acceptable AP of 0.698. The AP for gallbladder was 0.936, for cystic artery 0.704, for instruments 0.944, and for liver 0.929. The second model resulted in a decreased overall mAP (0.83) and a decreased AP for the CBD (0.664). The AP values for the other surgical landmarks in the second model remained similar to those in the first model: gallbladder, 0.97; instruments, 0.92; and liver, 0.865. The third model also showed a significant decline in AP for CBD at 0.59 and the cystic duct at 0.235 (the comparison is illustrated in Fig. 3).

The first model demonstrated superior performance in recognising the CBD in patients with symptomatic gallbladder stones and gallbladder polyps but exhibited lower accuracy in detecting the cystic duct in the cholecystitis group (Table S1, Fig. 4 and Fig. S1).

The mean accuracy for the CBD and cystic duct using the 6 class YOLOv7 was 94.39% and 84.97%, the sensitivity was 99.36% and 87.98%, the specificity was 19.35% and 88.89%, and F1-score 97.08% and 85.92% respectively (Table 3). The model demonstrated commendable proficiency in detecting CBD across various Nassar grades, maintaining high accuracy and sensitivity, even with increased Nassar grade complexity, but for the cystic duct in Nassar grade IV patients the detection rate decreased to only 55.88% and 60.61% respectively.

Discussion

This study validated an object detection model to detect critical surgical landmarks during fluorescence-guided LC. By integrating fluorescence imaging and deep learning, the model achieved an AP of 0.864 for CBD detection, demonstrating feasible performance across patients with varying degrees of surgical complexity, including grade IV cases.

AI-deep learning technology has been assessed for objective evaluation of intraoperative images during LC. Madani et al. and Laplante et al. developed segmentation models to differentiate safe (Go) and dangerous (No-Go) dissection zones. Mascagni et al. demonstrated the potential of deep neural networks for automated segmentation of hepatocystic anatomy in laparoscopic images with early-stage clinical evaluation¹⁶. Tatsushi et al. used the YOLOv3 algorithm to analyse landmark indications in 22 of 23 LC videos, achieving relatively low AP scores. Notably, their study excluded videos with significant fibrosis, scarring, bleeding, or less-visible landmarks.

Compared to other object detection models used in conventional LC, the AI model in this study introduced key advancements. The integration of ICG fluorescence enhanced both intraoperative visualization and functions as data preprocessing, improving model training by facilitating recognition of fluorescent structures like the CBD and cystic duct. YOLOv7 offers significant improvements over previous YOLO models, with re-parameterized modules and dynamic label assignment strategies reducing parameters by 40% and computational demands by 50%, while delivering faster and more accurate detection. The study hypothesized that including fluorescent anatomical structures and instruments in training enhanced recognition across different surgical scenes.

Tashiro et al. used an AI model with ICG fluorescence to identify loose connective tissue during LC in real time, achieving a Dice coefficient of 0.60¹⁷. Other models combined AI and ICG fluorescence, mainly in colorectal cancer detection, classifying cancer progression with high accuracy and minimal background noise due to delayed ICG retention¹⁸. Another model applied AI and ICG fluorescence angiography in flap reconstruction, accurately identifying excisable tissue based on fluorescence intensity¹⁹. Compared to traditional methods, AI with ICG fluorescence requires fewer annotated data sets while maintaining or surpassing accuracy, enhancing visualization, and reducing manual annotation effort. With increasing use of ICG-guided surgeries, this combination offers a cost-effective approach to improve surgical precision and real-time decision-making.

Before AI advancements, procedures like CVS and fluorescence-guided surgery were used to reduce BDI. Recent data from Japan’s National Clinical Database indicate that BDI rates have remained steady over the last decade²⁰. Both methods have limitations. A photo evaluation study revealed that only 26.8% of patients without cholecystitis had satisfactory CVS. In contrast, only 15.7% of patients with acute cholecystitis achieved a satisfactory CVS²¹. Stefanidis et al. also reported inconsistent adherence to CVS criteria, with only one-quarter of surgeons properly following them²². Fluorescence-guided LC presents challenges as well, with variations in bile duct fluorescence intensity due to tissue depth, diminishing effects over time, and background fluorescence interference. In this study, the AI-deep learning system integrated anatomical and fluorescence features to identify the CBD in real-time, aiding in achieving CVS. By continuously tracking the CBD throughout the dissection and before cystic duct transection, the system reduced the risk of BDI during LC.

In the subgroup analysis, the model’s recognition of the CBD and cystic duct showed little variation across different clinical indications. Higher Nassar grading scales were associated with more difficult dissections and reduced fluorescence visibility, significantly altering the appearance of anatomical structures. Although poor cystic duct recognition in Nassar grade IV cases was found, CBD identification remained feasible.

This study had several limitations. Although it included patients with varying degrees of surgical difficulty for model training, the overall sample size was small, and the distribution of Nassar grading scale cases was uneven. This imbalance may lead the model to perform better in simpler cases, limiting its effectiveness in more complex scenarios. Given the low clinical incidence of BDI at approximately 1% in LC patients, a large-scale, multicentre trial would be required to assess whether the AI tool can directly reduce BDI rates. The data set originated from a single centre with a consistent near-infrared camera system and ICG injection protocol, suggesting that further cross-validation and transfer learning may be needed to adapt the model to other systems. Variability in fluorescence visibility, influenced by factors such as visceral fat, inflammation, background fluorescence, and the duration of the ICG effect, presents a challenge for model training. Insufficient fluorescence intensity can lead to confusion between the cystic duct, cystic artery, and CBD. The manual selection of frames is time-consuming and may introduce human bias, potentially leading to false positives when visibility is poor during adhesiolysis or gallbladder take-down, underscoring the need for a comprehensive training data set.

In conclusion, this study demonstrated the feasibility of integrating an AI-deep learning system with fluorescence-guided surgery to identify extrahepatic bile ducts and anatomical structures during laparoscopic cholecystectomy. Future research should focus on improving cystic duct detection, expanding the data set for broader validation, and further exploring AI’s role in enhancing patient safety and surgical outcomes.

Funding

Chang Gung Memorial Hospital, grant number CMRPG8M1501 to Shih-Min Yin, funded this research.

Acknowledgements

The authors thank the colleagues of the general surgery department for the surgical video recording. J.-J.J.L. and I.-M.C. share the same contribution as the corresponding author.

Disclosure

The authors declare no conflict of interest.

Supplementary material

Supplementary material is available at BJS Open online.

Data availability

The data supporting this study’s findings are available from the corresponding author upon reasonable request.

Author contributions

Shih Min Yin (Conceptualization, Data curation, Funding acquisition, Project administration, Resources, Visualization, Writing—original draft, Writing—review & editing), I-Min Chiu (Formal analysis, Investigation, Methodology, Software, Supervision, Writing—review & editing), and Jenn-Jier Lien (Methodology, Resources, Supervision, Writing—review & editing).

References