Deep Learning-Based Automatic Segmentation Model for Radiotherapy Target Delineation of Brain Metastases and Its Clinical Applications

Objective  To develop an automatic segmentation model based on a U-shaped Convolutional Neural Network (U-Net) for delineating the Gross Tumor Volume (GTV) in Gamma Knife radiotherapy for brain metastases, and to evaluate its accuracy, efficiency and clinical applicability compared with manual delineation by junior clinicians.

Methods  A total of 100 patients with brain metastases who underwent Gamma Knife treatment were included, and the model was trained using data augmentation techniques. Using contours delineated by senior clinicians as the gold standard, the model and junior clinicians were compared in terms of the Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), total time consumption, single-sample processing time, Average Symmetric Surface Distance (ASSD), and Maximum Symmetric Surface Distance (MSSD).

Results  The mean DSC for the model and junior clinicians were 0.84 and 0.74, respectively, while the average single-case processing times were 1.8 minutes and 14.2 minutes, respectively. The differences in segmentation accuracy between the two groups were statistically significant (p < 0.05).

Conclusion  The proposed U-Net+GTV model outperforms junior clinicians in accuracy, efficiency, and stability, demonstrating considerable potential for clinical application in GTV delineation for Gamma Knife radiotherapy.