Automatic segmentation and evaluation techniques for free flap in reconstruction surgery using deep learning

Postoperative free flap monitoring would be crucial to achieve the best outcomes of microsurgical reconstruction, for which meticulous clinical examination has served as the gold standard. Despite its high reliability, requirement of excessive consumption of manpower has been addressed a main drawback. To overcome this issue, this study aimed to develop a novel system for free flap monitoring, which is based on artificial intelligence (AI) using deep learning system. Photographs of postoperative appearance of free flaps were gathered retrospectively from patients who underwent free flap-based reconstruction between 2020 and 2021, and prospectively from those in 2022, and used with dividing into 80%:20% for model training and test. The development process proceeded with categorizing into two parts; flap segmentation model that can identify the flap region in the photographs, and flap status classification model that can evaluate the flap perfusion status based on its color. A total of 2,068 photographs of 433 patients were used. The most reliable model for flap segmentation was developed based on U-Net algorithm, achieving a Dice Coefficient of 0.972. For the flap status classification model, the Support Vector Machine algorithms was adopted, showing an accuracy of 0.926, from which the ratio of red pixels in the flap region was extracted as a quantitative indicator. Our results suggest that this novel AI-based model for flap segmentation and status classification could achieve reliable outcomes. A further upgraded system based on this model may provide optimal flap monitoring, keeping high accuracy and reducing medical staffs' labor burden.