Segmentation of Brain Metastases in MRI: A Two-Stage Deep Learning Approach with Modality Impact Study

TL;DR

This paper introduces a two-stage deep learning approach for brain metastasis segmentation in MRI, emphasizing the importance of modality selection and demonstrating improved accuracy over existing methods.

Contribution

The study presents a novel two-stage detection and segmentation model that leverages specific MRI modalities for enhanced brain metastasis segmentation.

Findings

Combining T1c, T1, and FLAIR modalities improves segmentation accuracy.

The proposed model outperforms single-pass models in detecting small metastases.

The approach sets a new benchmark in brain metastasis segmentation.

Abstract

Brain metastasis segmentation poses a significant challenge in medical imaging due to the complex presentation and variability in size and location of metastases. In this study, we first investigate the impact of different imaging modalities on segmentation performance using a 3D U-Net. Through a comprehensive analysis, we determine that combining all available modalities does not necessarily enhance performance. Instead, the combination of T1-weighted with contrast enhancement (T1c), T1-weighted (T1), and FLAIR modalities yields superior results. Building on these findings, we propose a two-stage detection and segmentation model specifically designed to accurately segment brain metastases. Our approach demonstrates that leveraging three key modalities (T1c, T1, and FLAIR) achieves significantly higher accuracy compared to single-pass deep learning models. This targeted combination allows for precise segmentation, capturing even small metastases that other models often miss. Our model sets a new benchmark in brain metastasis segmentation, highlighting the importance of strategic modality selection and multi-stage processing in medical imaging. Our implementation is freely accessible to the research community on \href{https://github.com/xmindflow/Met-Seg}{GitHub}.