Detection and Segmentation of Arteriovenous Malformation Lesions Using a Two-Stage Deep Learning Strategy

Abstract

<h3>Purpose/Objective(s)</h3> Identifying and accurately delineating malformed vascular lesions is critical in radiosurgery for arteriovenous malformations (AVMs). Applying deep learning to automatic detection and segmentation of AVM lesions may help the treatment planning be more efficient. This study aims to develop an artificial intelligence model to recognize and contour the AVM nidus using deep learning methods. <h3>Materials/Methods</h3> We retrospectively collected two hundred twenty-three AVM patients treated by radiosurgery from 2003 to 2020. Among the 233 cases, 179 were in the training set, and 44 were in the testing set. The contrasted magnetic resonance image of time-of-flight was the sequence for image acquisition. The targeted area depicted in the previous treatment planning was the ground truth. We utilized a two-stage deep learning strategy. The first stage is the yolov5 model to detect the bounding box of the AVM lesions. Then we used the U-Net++ for the auto-segmentation based on the bounding box generated from the first stage in the second stage. We trained and evaluated the first and second stage models separately for 100 epochs, and the model with the lowest error was chosen as the optimal model. <h3>Results</h3> In the first stage, the recall and precision of the yolov5 model were 0.93 and 0.91 for the training data set and 0.88 and 0.93 for the testing data set, respectively. The second U-Net++ segmentation model achieved a Dice score of 0.98 for both the training and testing data sets. The segmentation results demonstrated the statistically significant difference between the Dice score calculated from the predicted segmentation and reference bounding box (p<0.0001). Besides, we observed that segmentation accuracy gradually decreased as the bounding box expanded. <h3>Conclusion</h3> Our two-stage model exhibits satisfactory accuracy in the images' auto-detection and auto-delineation of AVM lesions. The model needs to be further verified by more data before clinical use. It is expected that such an AI model will improve the clinical process and reduce interobserver variation.