Artificial Intelligence-Enabled Skin Perforator Segmentation

The aim of this single-arm clinical trial is to validate the performance of the artificial intelligence-enabled skin perforator segmentation tool (artificial intelligence tool) in computer-assisted fibula free flap harvest.

To test the predictive accuracy of the artificial intelligence tool in identifying the targeted skin perforators (primary endpoint)

• The predictive accuracy is calculated from the number of true/false positive, true/false negative targeted perforators.
• The secondary outcomes include the sensitivity, specificity, positive, and negative predictive values (positive predictive value and negative predictive value respectively).

Our hypothesis is that the artificial intelligence-enabled vessel segmentation tool will demonstrate a high level of accuracy in identifying the fibula skin perforators.

Plan of Investigation:

(i) Study design The current study is a single-arm single-centre clinical trial to explore the performance of an artificial intelligence-enabled skin perforator segmentation tool in computer-assisted surgery fibula free flap harvest. The study protocol shall conform to Standard Protocol Items: Recommendations for Interventional Trials and Practical Robust Implementation and Sustainability Model.

Data collection procedure A case report form has been designed to collect pre- and intra-operative research data.

Patient demographics will be collected, including age, gender, body height, weight, body mass index, diagnosis, medical history, and smoking status.

Preoperative measurements will also be obtained, including but not limited to:

• Total fibula length
• Number and lengths of fibula segments
• Number of immediate dental implants
• Number of perforators
• Type of perforators:

1. Septocutaneous vs musculocutaneous
2. Targeted (to be incorporated into the virtual surgical planning and harvesting) vs accessory (other perforators)

• Relative location of the perforators, measured as the distance from the most inferior point of the lateral malleolus to:

1. Level of the perforator, as it crosses the posterior aspect of the fibula
2. The point where it bifurcates from the peroneal artery

• Time spent incorporating the skin perforator model to the original virtual surgical planning workflow.

Intraoperatively, the following information will be recorded:

• Deviation between the locations of the skin perforators identified by the artificial intelligence tool and during the surgery
• Need for intraoperative modifications of the plan due to the perforator identification discrepancy
• Operator satisfaction

The primary endpoint of the predictive accuracy will be calculated according to the definition by Šimundić et al. When the skin perforator is identified by both the artificial intelligence-segmentation tool and during the surgery, it will be counted as a true positive. When the perforator is identified by the artificial intelligence-segmentation tool, but not found during the surgery it is counted as a false positive. When the perforator is seen during the surgery, but not shown by the artificial intelligence tool, it is a false negative. Finally, a true negative perforator count will be derived from those subjects, who do not exhibit a false negative perforator. The predictive accuracy is identified as the percentage of true perforators among all the perforators, and is calculated as (true positive + true negative)/(true positive+false positive+true negative+false negative)*100%.

The secondary outcomes of sensitivity, specificity, positive and negative predictive values will be calculated.

(ii) Methods

The workflow has already been set up for the previous projects. The efficiency, predictability, and accuracy of the protocol have been demonstrated in previously published articles.

For the currently proposed project, the investigators aim to incorporate the artificial intelligence tool into the current virtual surgical planning workflow and assess its performance. The complete process will be divided into three parts: 1) preoperative preparation; 2) surgery and intra-operative recording; 3) postoperative revaluation. In the preliminary study, the investigators have established the feasibility of the methodology via several pilot cases.

1. Preoperative preparations: Preoperative computed tomography imaging: Patients enrolled in the study will receive a computed tomography scan on a General Electric revolution system (General Electric Healthcare, Chicago, IL) with slice thicknesses of 0.625 mm for both the head and neck, and lower limbs. Iodinated contrast medium (Iopamiro 370, Bracco, Israel) at a concentration of 370 mg/ml will be administered intravenously at a flow rate of 4 to 5 ml/sec, computed tomography angiogram of the lower extremity will be taken at the arterial phase. The imaging data will be exported in the Digital Imaging and Communications in Medicine format. Artificial intelligence-enabled segmentation of skin perforators from computed tomography angiogram: The Digital Imaging and Communications in Medicine data of the computed tomography angiogram will be imported to the artificial intelligence workstation (United Imaging Intelligence, Shanghai, China). The tibia and fibula bone, skin, blood vessels of the lower extremities including the anterior tibial artery, posterior tibial artery, and peroneal artery will be segmented with pre-set threshold range to generate the stereolithography (.stl) models. Based on the fully convolutional neural network of V-net, the blood vessel branches from the peroneal artery will be automatically segmented to build the .stl model. The small perforating vessels from the peroneal artery, traveling laterally, crossing the posterior edge of the fibula bone to the skin surface, will be marked as skin perforators. The .stl models of the fibula bone, skin, anterior tibial artery, posterior tibial artery, peroneal artery, and skin perforators will be exported from the workstation for incorporation into the virtual surgical planning workflow. Virtual surgery planning: Preoperative imaging and building of models Preoperative computed tomography scan will be imported to ProPlan CMF 2.1 software (Materialise, Leuven, Belgium). Volume segmentation will be performed to build the 3D virtual models of the jaws. Incorporating the skin perforators The .stl models exported from the artificial intelligence workstation will be imported to the ProPlan file. The fibula length and skin perforator location will be measured and recorded. Virtual fibula flap harvest, segmentation, and dental implant insertion Virtual surgical planning will be performed using the methodology described in the previous publications. The extent of resection will be pre-determined, based on the type and extension of the pathology. The osteotomies for fibula harvest and segmentation will be planned to suit the defect after guided tumour resection. The location of the immediate dental implants will be marked in the fibula segments. Taking the location of the skin perforators into consideration, the osteotomies can be adjusted to include the take-off point of the skin perforators, while avoiding placing the perforator directly in planned osteotomies or implant insertion points. Design and fabrication of surgical guides and/or plates After the virtual surgical planning is confirmed, the surgical guides and/or plates will be designed using 3-matic 13.0 software (Materialise, Leuven, Belgium). The location of the perforators as they cross the posterior edge of the fibula will be marked with a small sphere on the fibula harvest guide. The guides will be printed with biocompatible and autoclavable resin - either MED610 (Stratasys Ltd, Eden Prairie, MN, USA) or NextDent SG (Vertex Dental, The Netherlands).
2. Surgical procedures and intraoperative recording All the surgical procedures will be conducted in the routine manner of computer-assisted surgery jaw reconstruction using fibula free flap by the principal applicant. Briefly, osteotomies, direct dental implant insertion, bone movements, and bone inset will be guided by the prepared surgical templates. Intraoperative measurements will be conducted. Any false positive or false negative perforators will be recorded. Any unanticipated event that requires the modification of the fibula harvest plan will be recorded. Standard perioperative management will be conducted in a routine manner.
3. Postoperative evaluation Postoperative evaluation will be conducted between the surgical team led by the principal applicant and the computer science team led by the co-applicant. Any intraoperative events will be discussed in a timely manner. Data entry will be performed for each case right after the surgery. (iii) Sample size The calculation of the sample size is by the primary endpoint. As the success rate of the conventional method in identifying a reliable correlation was 67%, the aim of the artificial intelligence-segmentation tool is to achieve a minimum accuracy of 80%. A one-sided test for non-inferiority/superiority is employed. The sample size is calculated using the exact binomial test for a one-sample non-inferiority/superiority clinical study. The hypothesis can be unified as: H0: p = p0 + δ versus Ha: p = p1. Here, δ represents the intervention effect size; p0 equals 80% and p is the true accuracy which is hypothesised to exceed 80% with a statistical power of at least 90% with a 5% level of significance.

According to a previous report on the artificial intelligence-enabled skin perforator segmentation tool, the diagnostic accuracy was reported as 97%. The investigators can conservatively estimate the true success rate as 95% (representing an intervention effect size of 15%). Based on the calculation, a sample size of 44 is expected to have a 93.3% statistical power. Assuming at least one targeted perforator for each case and a dropout rate after signing the informed consent of 10%, a total of 49 patients will be recruited to the study.

(iv) Data processing and analysis

All the data in this protocol will be recorded using a specially designed case report form (Appendix). Discrepancies in the database will be traced to the original documents for resolution. All records will be audited 100% for accuracy. Data entry will be performed by two independent assistants using the IBM SPSS Statistics Version 25. Consistency of input data between assistants will be compared using an in-built function.

For the demographic information, data will be presented as mean values with standard deviations for continuous data, and as counts with proportions for categorical data. The predictive accuracy will be presented as a frequency. The one-sided 95% binomial proportion confidence interval will be computed, and the lower confidence limit will be compared with 80%. The sensitivity, specificity, and positive and negative predictive values will be calculated. Continuous data will be expressed as mean values with standard deviation for normally distributed data and expressed as the median with interquartile range or range for skewed data. To evaluate the effect of various confounding variables on the precision of perforators identified by artificial intelligence, suitable tests such as Pearson or Spearman's correlation test, t-Test, or Wilcoxon Signed-rank test will be selected depending on the normality of the data. Statistical significance will be set at p< 0.05. All statistical analyses will be performed using IBM SPSS Statistics Version 25.