# Development of an automatic segmentation system for anterolateral thigh flap perforators in maxillofacial reconstruction

## TL;DR

This paper presents an AI-based system to automatically identify blood vessel perforators in thigh flaps used for facial reconstruction, improving accuracy and efficiency over manual methods.

## Contribution

A novel CNN-based automatic segmentation system for detecting ALT flap perforators in CTA images with a two-stage cascaded approach.

## Key findings

- The system achieved a DSC of 69.67 ± 1.48 and JSC of 67.81 ± 1.70 for perforator segmentation.
- Distance differences between manual and automatic detection were 38.28 ± 15.52 mm on the left and 31.96 ± 18.11 mm on the right.
- The system shows potential for clinical use in improving surgical planning for maxillofacial reconstruction.

## Abstract

The anterior thigh (ALT) flap is commonly used in reconstructive surgery, especially in maxillary reconstruction. Accurately identifying the perforator that supplies blood to the flap is critical for surgical success but is time-consuming and prone to variability since it is traditionally performed manually. Advances in artificial intelligence have shown convolutional neural networks (CNN) the potential to automate medical image segmentation. However, ALT flap perforator segmentation poses a unique challenge due to the small size of the perforator and its high anatomical variability. To address this challenge, we developed and validated a CNN-based automatic segmentation model for detecting ALT flap perforators on computed tomography angiography (CTA). Manual annotations of bilateral lateral femoral circumflex artery perforators were obtained from 80 patients using an image tracing program for comparison. The training for the development of an automatic segmentation system was then conducted based on these manual segmentation. The automatic segmentation system employed a two-stage cascaded approach: 2D detection with DeepLabv2 and 3D segmentation with ResNet152. Data augmentation techniques were applied to improve model generalization. Performance metrics included the dice similarity coefficient (DSC) and jaccard similarity coefficient (JSC). The automatic segmentation system achieved DSC and JSC values of 69.67 ± 1.48 and 67.81 ± 1.70, respectively. The distance differences between manual and automatic detection were 38.28 ± 15.52 mm on the left side and 31.96 ± 18.11 mm on the right side. The automatic segmentation system for ALT flap perforators demonstrates promising accuracy, highlighting its potential for clinical application. By reliably identifying perforator locations in CTA, the system can enhance the efficiency and precision of surgical planning, particularly for maxillofacial reconstruction.