Introduction and Aims: Over the past decade, we have performed over 1000 cranial stereotactic radiosurgical treatments using a Novalis TX linear accelerator, resulting in a significant collection of MRI data from patients with brain metastases. With the use of this data, we aimed to facilitate radiosurgical planning by developing a system capable of automatically contouring metastatic lesions on contrast-enhanced T1 brain MRI scans.

Material and methods: We gathered contrast-enhanced T1 MR images of brain metastatic cases from our Brainlab Iplannet database. These images were manually contoured for previous radiosurgical treatment. A portion of this data was input into nnU-Net, a deep learning-based segmentation method that adapts automatically to new tasks. The evaluation was based on Dice’s coefficient, ranging from 0.0000 to 1.0000, representing the degree of overlap between manually and automatically contoured lesions.

Results: We utilized 72 cases with suitable technical data, each containing 1-4 manually contoured metastatic lesions. Seventy-five percent of the data were used to train nnU-Net, while the remaining 25% assessed the accuracy of automated contouring. The Dice’s coefficient averaged 0.812. We excluded 7 cases due to undetected lesions in 5 instances (3-4mm), while in 2 cases, manually non-contoured extracranial lesions were found. After correction, the Dice’s coefficient rose to 0.890.

Conclusion: Our findings demonstrate successful automatic contouring of predefined lesions, yielding promising results compared to the interobserver Dice’s coefficient reported as 0.91 in the literature. Deep learning-based segmentation can effectively develop a tool capable of contouring targeted metastatic lesions for radiosurgical treatment. The results may be further improved with additional MRI sequences, such as post-contrast FLAIR. Our study suggests future potential for the automated identification of metastatic lesions.