Predicting multiple linear stapler firings in double stapling technique with an MRI-based deep-learning model

Multiple linear stapler firings is a risk factor for anastomotic leakage (AL) in laparoscopic low anterior resection (LAR) using double stapling technique (DST) anastomosis. In this study,

our objective was to establish the risk factors for ≥ 3 linear stapler firings, and to create and validate a predictive model for ≥ 3 linear stapler firings in laparoscopic LAR using DST

anastomosis. We retrospectively enrolled 328 mid–low rectal cancer patients undergoing laparoscopic LAR using DST anastomosis. With a split ratio of 4:1, patients were randomly divided into

2 sets: the training set (n = 260) and the testing set (n = 68). A clinical predictive model of ≥ 3 linear stapler firings was constructed by binary logistic regression. Based on

three-dimensional convolutional networks, we built an image model using only magnetic resonance (MR) images segmented by Mask region-based convolutional neural network, and an integrated

model based on both MR images and clinical variables. Area under the curve (AUC), sensitivity, specificity, accuracy, positive predictive value (PPV), and Youden index were calculated for

each model. And the three models were validated by an independent cohort of 128 patients. There were 17.7% (58/328) patients received ≥ 3 linear stapler firings. Tumor size ≥ 5 cm (odds

ratio (OR) = 2.54, 95% confidence interval (CI) = 1.15–5.60, p = 0.021) and preoperative carcinoma embryonic antigen (CEA) level > 5 ng/mL [OR = 2.20, 95% CI = 1.20–4.04, p = 0.011] were

independent risk factors associated with ≥ 3 linear stapler firings. The integrated model (AUC = 0.88, accuracy = 94.1%) performed better on predicting ≥ 3 linear stapler firings than the

clinical model (AUC = 0.72, accuracy = 86.7%) and the image model (AUC = 0.81, accuracy = 91.2%). Similarly, in the validation set, the integrated model (AUC = 0.84, accuracy = 93.8%)

performed better than the clinical model (AUC = 0.65, accuracy = 65.6%) and the image model (AUC = 0.75, accuracy = 92.1%). Our deep-learning model based on pelvic MR can help predict the

high-risk population with ≥ 3 linear stapler firings in laparoscopic LAR using DST anastomosis. This model might assist in determining preoperatively the anastomotic technique for mid–low

rectal cancer patients.

Colorectal cancer (CRC) is the third most prevalent cancer worldwide with high rate of cancer mortality, while approximately one third of all CRCs occur in rectum1. Total mesorectal excision

(TME) is the primary treatment for localized rectal cancer patients, and has high rate of cancer control2,3. It is a widely applied surgical technique to perform laparoscopic low anterior

resection (LAR) using double-stapling technique (DST) anastomosis4,5.

In mid–low rectal cancer patients, anastomotic leakage (AL) is the most common postoperative complication after LAR6. In addition to increasing the incidence of reoperation and mortality, AL

may even negatively affect long-term survival7,8,9. Despite advances in surgical techniques and anastomotic equipment over the past few decades10,11,12, there has been no significant

decrease in the incidence of AL after LAR13.

DST simplifies colorectal reconstruction in LAR, especially in laparoscopic surgery, but the incidence of postoperative AL was not reduced by the application of this technique14,15. Several

studies have proved that ≥ 3 linear stapler firings in laparoscopic LAR is an independent risk factor for AL16,17,18. Thus, the Chinese expert consensus statement proposes ≤ 2 linear stapler

firings in LAR surgery19. To avoid multiple linear stapler firings, a predictive model should be created to predict ≥ 3 linear stapler firings in anastomosis, and alternative techniques

could be considered.

Several studies have shown that pelvic anatomical features (such as the anteroposterior diameter and the transverse diameter of the pelvic outlet, the anteroposterior diameter of the pelvic

inlet) are related to surgical difficulty and the number of the linear stapler firings20,21,22. Those studies mainly measured pelvic parameters, but did not include the influence of rectal

and mesenteric conditions. How to effectively integrate above parameters to predict ≥ 3 linear stapler firing needs further research.

With the advancement of imaging technology, pelvic magnetic resonance imaging (MRI) is a preferred tool for local staging of mid-low rectal cancer before surgery23,24. MRI can obtain

relevant parameters of tumor, meso-rectum and pelvis comprehensively and accurately. Mask region-based convolutional neural network (Mask R-CNN)25 and three-dimensional convolutional network

(C3D)26 image recognition are current techniques for advanced-recognition artificial intelligence (AI), and have been applied in various medical fields27,28,29. With advanced deep-learning

technology, pelvic MRIs’ complex data can be identified, extracted, analyzed, and integrated efficiently, and we can create a predictive model to screen out high-risk patients with ≥ 3

linear stapler firings based on the database.

In this study, we aimed to establish the risk factors for ≥ 3 linear stapler firings in laparoscopic LAR using DST anastomosis, and to create and validate a predictive model for ≥ 3 linear

stapler firings with deep-learning technology based on MRI.

A total of 328 mid–low rectal cancer patients who underwent laparoscopic LAR at Ruijin Hospital Affiliated to Shanghai Jiaotong University School of Medicine, China, between January 2016 and

June 2021 were retrospectively analyzed as the deep-learning set. Clinical data were obtained from Ruijin hospital database and medical records. All methods were performed in accordance

with the relevant guidelines and regulations, and the study was approved by the Medical Ethics Committee of Ruijin Hospital (No. 2019-82). The need for informed consent was waived by Ethics

Committee of Ruijin Hospital. With a split ratio of 4:1, 260 patients were divided into the training set and 68 patients were divided into the testing set on the basis of an unbiased random

sampling method. The prospective validation set comprised 128 patients from an independent clinical trial in our institution (Transanal versus laparoscopic total mesorectal excision for

rectal cancer, ClinicalTrials.gov Identifier: NCT03359616).

The inclusion criteria were: (1) histopathologically confirmed rectal cancer; (2) pelvic MRI examination within 14 days before surgery; (3) tumor distance from the anal verge ≤ 10 cm; and

(4) laparoscopic LAR with DST anastomosis. The exclusion criteria were: (1) LAR without anastomosis (e.g., Hartmann’s operation); (2) anastomotic techniques rather than DST (e.g., manual

rectal anastomosis through the anus); (3) number of linear stapler firings was not recorded on surgical reports; and (4) robotic surgery.

Laparoscopic LAR was performed by the gastrointestinal surgery team with experience in completing more than 100 rectal cancer operations every year. The laparoscopic LAR surgical procedure

was carried out in strict accordance with the national guidelines for laparoscopic radical resection of CRC (2018 edition). During dissection of the distal rectum, the surgeon manually fired

endoscopic linear staplers (Endo-GIA™ Ultra Universal Stapler Reload with Tri-stapler™ Technology; Covidien LLC), which was loaded with 60- or 45-mm staple cartridges which have three types

of heights: 3.0, 3.5 and 4.0 mm.

We collected clinical data that potentially correlated with the number of linear staplers used in surgery, including baseline characteristics, including age, gender and body mass index

(BMI); biochemical data, including hemoglobin, albumin, and carcinoma embryonic antigen (CEA); and tumor characteristics, including tumor distance from the anal verge, tumor stage, tumor

size and circumferential resection margin (CRM). A predictive model of ≥ 3 linear stapler firings was constructed by binary logistic regression. The variables of the clinical model included:

gender, BMI, serum CEA level (> or ≤ 5 ng/mL), tumor distance from the anal verge, tumor size and CRM.

During MRI, patients were in the supine position and scanned using a Philips INGENIA™ scanner with 3.0 T field strength. The pelvic phased-array surface coil covered from the aortic

bifurcation to the anal verge. The scanning parameters were: layer thickness 5 mm; field of view 250 × 340 × 166 mm; echo time 80 ms; repetition time 3565 ms; and image matrix 312 × 357.

With the Picture Archiving and Communication System, fat-suppressed fast spin-echo (FSE) T2-weighted sequences in the axial plane of the pelvis were used for image segmentation. Then, pelvic

MRI specialists who have > 10 years of experience built an image database by an online annotation tool called Labelme (labelme.csail.mit.edu/)30, and labeled three kinds of target area on

each of the T2-weighted images (tumor body, mesorectum, and pelvis represented by green, yellow, and drab, respectively, Fig. 1A,B). Then, all data were transformed into a COCO dataset for

segmentation experiments31.

Examples of target regions. (A,B) manually labeled; (C,D) segmented by Mask region-based convolutional neural network (Mask R-CNN) based model (the regions of tumor body, mesorectum, and

pelvis were represented by green, yellow, and drab, respectively).

As an effective two-stage detection and segmentation algorithm, Mask R-CNN was adopted to identify and segment the three kinds of target areas in the image (Fig. 1C,D). In the first stage,

the ResNet-101-FPN network served as the backbone to extract multiscale and discriminative feature maps. The Region Proposal Network scanned the feature map in a sliding-window and selected

the rough detection rectangle that contained the object. After the regions of interest alignment process, the candidate regions entered the next stage. This consisted of three functional

branches: classification, detection and segmentation, based on fully connected layers and convolutional layers. We trained the Mask R-CNN network on the training set for 200 epochs, and

evaluated the performance of the testing set with standard COCO metrics. We evaluated the trained Mask R-CNN model, obtained average precision (AP) by calculating the precision–recall curve

under different intersection-over-union thresholds, and then calculated the three types of targets. The respective AP values of the regions were weighted to obtain the class-wide mean

average precision (mAP). mAP > 50 indicated that the model performed well32.

We used C3D networks to address 24 images of a case simultaneously and learn 3D spatial features. The C3D network consisted of eight 3D-convolution layers, a softmax layer, two fully

connected layers and five pooling layers. It took the entire image of the case as input and output the probability of ≥ 3 linear stapler firings, and the sample with probability > 50%

(empirical value) was judged as positive. The C3D network was trained until convergence (~ 1000 epochs) and evaluated the performance of the deep-learning model. In the training set, two

C3D-based models were trained, including an image model using only MR images and an integrated model based on both MR images and the six clinical variables in clinical model. In the testing

set, the clinical model, image model and integrated model were examined, and receiver operating characteristic (ROC) curves were plotted. Area under the curve (AUC), sensitivity,

specificity, accuracy, positive predictive value (PPV), and Youden index were calculated.

We used clinical data and T2-weighted images of patients from the validation set to validate the predictive performance of the above three models. ROC curves were plotted, and sensitivity,

specificity, accuracy, Youden index, PPV, and AUC were calculated. The flow chart of the design is shown in Fig. 2.

Flow chart of the design. BMI: Body mass index; CEA: Carcinoembryonic antigen; CRM: Circumferential resection margin; MR: Magnetic resonance; Mask R-CNN: Mask region-based convolutional

neural network.

Statistical analysis was performed using SPSS (version 25.0). Categorical variables were analyzed by Fisher’s exact test or Pearson’s chi-square. In our binary logistic regression models,

only factors with a P value