# Longitudinal assessment of radiosurgery response in small brain metastases: AI‐driven precision tumor segmentation and monitoring on serial MRI

## TL;DR

This paper introduces an AI framework to accurately segment and monitor small brain metastases on MRI scans, improving precision and reducing manual workload.

## Contribution

A novel transformer-based framework with 3D neighborhood attention for precise segmentation of small brain metastases on serial MRI.

## Key findings

- The framework achieved dice scores of 89.8%, 92.0%, and 93.1% for tumors <1 cm, 1-2 cm, and >2 cm respectively.
- It detected tumor size changes and treatment outcomes with over 96% accuracy across different tumor sizes.

## Abstract

The conventional method of assessing radiotherapy outcome in brain metastases (BM) is based on monitoring tumor size alterations on serial magnetic resonance imaging (MRI). To accurately determine changes in tumor dimensions, targets require delineations on several volumetric images acquired before treatment and at multiple follow‐up scans after radiotherapy. However, manual tumor delineation on serial MRI is labor‐intensive, imposes a significant burden on the clinical workflow, and is prone to variability especially for smaller lesions.

This study proposes a novel multi‐step transformer‐based automated framework with a 3D neighborhood attention mechanism, specifically designed to enhance the segmentation precision for BM of various sizes on standard longitudinal MRI. This framework leverages the hierarchical encoding capabilities of transformer architecture to capture intricate tumor characteristics, with a particular focus on improving the delineation of small metastases (<1 cm), which are often overlooked by existing models.

The proposed framework was trained on the BraTS and BraTS‐METS datasets and evaluated on independent external data acquired from 212 patients (508 BM lesions) treated with stereotactic radiosurgery. The framework's performance was evaluated in segmenting tumors across various size categories, monitoring post‐treatment changes in tumor size on serial MRI, and automatically detecting local control/failure (LC/LF) and adverse radiation effect (ARE) following radiosurgery.

The framework achieved a dice score of 89.8 ± 3.4%, 92.0 ± 3.0%, and 93.1 ± 2.3% for tumors with a size of less than 1 cm, between 1 and 2 cm, and larger than 2 cm, respectively. It also demonstrated high performance in longitudinal monitoring of tumor size changes and in detecting LC/LF and ARE, achieving accuracies greater than 96% across different tumor size categories compared to the clinical outcome assessment. The results exhibited a substantial improvement over state‐of‐the‐art segmentation models, particularly for smaller lesions.

This study represents a step forward toward deploying AI‐driven decision support tools to the neuro‐oncology workflow, reducing the assessment burden on oncologists, and improving consistency in routine radiotherapy outcome assessments.