Enhancing CNN-based Brain Metastasis Detection in MRI by Integrating Locoregional 3D Deformation Technique

Abstract

AbstractIntroduction: The detection of brain metastases in magnetic resonance imaging (MRI) remains a significant challenge due to the small size and subtle nature of metastatic lesions. This study introduces a novel deep learning-based approach that integrates a locoregional 3D image deformation technique with a 3D U-Net model to enhance segmentation accuracy for brain metastases in T1-weighted contrast-enhanced (T1c) MRI scans. Methods: To improve the visibility of small and subtle metastases, a local 3D magnification technique was applied to MRI images, producing a "fish-eye" effect that emphasizes lesion structures while preserving global anatomical context. This deformation was systematically applied throughout the brain to enhance the detection of metastases. A 3D U-Net model was then trained using a dataset of 236 T1c MRI scans from the BraTS-METS 2023 challenge. The dataset was divided into training (70%), validation (10%), and test (20%) subsets. The proposed model was evaluated against a baseline 3D U-Net model trained on non-deformed images, using Dice coefficient, sensitivity, and precision as performance metrics. Results: The integration of the 3D deformation technique improved segmentation performance, achieving an average Dice coefficient of 0.7023, sensitivity of 0.3110, and precision of 0.2021. In comparison, the baseline model without deformation achieved a Dice coefficient of 0.6472, sensitivity of 0.2139, and precision of 0.1533. The results demonstrate that the 3D deformation method enhances the model’s ability to detect small metastases, thereby improving segmentation accuracy. Conclusion: This study highlights the effectiveness of a locoregional 3D deformation approach in improving the detection and segmentation of brain metastases in T1c MRI scans. By selectively magnifying key regions of interest, the proposed method enhances deep learning model sensitivity and precision. Future work will focus on validating the approach with larger and more diverse datasets, as well as exploring its potential applications in other medical imaging tasks.