[Establishment and preliminary clinical validation of an artificial intelligence-assisted perforator identification model for anterolateral thigh perforator flaps based on ICGA videos].

Abstract

Objective: To establish and preliminarily clinical validate an artificial intelligence-assisted perforator identification model for anterolateral thigh perforator flaps based on indocyanine green angiography (ICGA) videos. Methods: This study was a cross-sectional study. From January to July 2024, 70 patients who met the inclusion criteria and underwent anterolateral thigh perforator flap repair surgery were admitted to Suzhou Ruihua Orthopedic Hospital of Soochow University. There were 54 males and 16 females, aged 10 to 70 years. Fifty patients were selected as the training set for model parameter training using the random function in Python 3.13, and the remaining 20 patients were assigned as the validation set for independent validation. All patients received preoperative color Doppler ultrasound (CDU) and ICGA examinations. Postoperatively, artificial intelligence-assisted identification based on ICGA videos was adopted for perforator localization. Four perforator sites were designated in the anterolateral thigh perforator flap area of each patient. In the training set and the validation set of patients, the positive in perforator localization of CDU, ICGA-manual interpretation, and artificial intelligence-assisted ICGA (ICGA-AI) was collected, and the positive rate was calculated. The sensitivity, specificity, positive predictive value, negative predictive value, accuracy, and balanced accuracy of the three methods for perforator localization were calculated; the consistency of the perforator localization results between the three methods and the intraoperative exploration was recorded. The distance between the perforator localization points determined by the three methods and the corresponding actual intraoperative cutaneous perforator entry point (error distance of perforator localization) was measured. The consistency between the dominant perforator localized by ICGA-AI and that determined intraoperatively was recorded, and the localization accuracy of the dominant perforator was calculated. Results: There were no statistically significant differences in the positive rates of perforator localization among the three methods in both the training set and the validation set of patients (P>0.05). In the training set of patients, the sensitivity, specificity, positive predictive value, negative predictive value, accuracy, and balanced accuracy for perforator localization were 88.70%, 85.88%, 89.47%, 84.88%, 87.50%, and 87.29% for CDU, 93.91%, 89.41%, 92.31%, 91.57%, 92.00%, and 91.66% for ICGA-manual interpretation, and 82.61%, 83.53%, 87.16%, 78.02%, 83.00%, and 83.07% for ICGA-AI, respectively; the consistency between the perforator localization results by CDU, ICGA-manual interpretation, and ICGA-AI and the findings from intraoperative exploration was strong or almost perfect (with κ of 0.74, 0.84, and 0.66, respectively, P<0.05). In the validation set of patients, the sensitivity, specificity, positive predictive value, negative predictive value, accuracy, and balanced accuracy for perforator localization were 95.24%, 78.95%, 83.33%, 93.75%, 87.50%, and 87.09% for CDU, 95.24%, 92.11%, 93.02%, 94.59%, 93.75%, and 93.67% for ICGA-manual interpretation, and 95.24%, 94.74%, 95.24%, 94.74%, 95.00%, and 94.99% for ICGA-AI, respectively; the consistency between the perforator localization results by CDU, ICGA-manual interpretation, and ICGA-AI and the findings from intraoperative exploration was strong or almost perfect (with κ of 0.75, 0.88, and 0.90, respectively, P<0.05). In the training set of patients, the error distance of perforator localization of ICGA-AI was 0.84 (0.70, 0.98) cm, which was significantly shorter than 1.41 (0.80, 2.00) cm of CDU (Z=-2.56, P<0.05) and 1.50 (1.03, 1.80) cm of ICGA-manual interpretation (Z=-5.88, P<0.05). In the validation set of patients, the error distance of perforator localization of ICGA-AI was 0.52 (0.35, 0.71) cm, which was significantly shorter than 1.50 (1.00, 2.25) cm of CDU (Z=-3.23, P<0.05) and 1.56 (1.41, 2.00) cm of ICGA-manual interpretation (Z=-3.52, P<0.05). In the training set and the validation set of patients, the localization accuracy of ICGA-AI for the dominant perforator was 82.00% (41/50) and 90.00% (18/20), respectively. Conclusions: An artificial intelligence-assisted perforator identification model for anterolateral thigh perforator flaps based on ICGA videos is established. In patients of the validation set, this model shows smaller localization errors and higher accuracy in locating dominant perforators, which can assist with preoperative perforator localization.