Author: Minbin Chen, Ke Lu, Kaizhong Shi, Chunhao Wang, Chuan Wu, Zhenyu Yang, Fang-Fang Yin, Jingtong Zhao 👨🔬
Affiliation: The First People's Hospital of Kunshan, Duke University, Medical Physics Graduate Program, Duke Kunshan University, Duke Kunshan University, Department of Radiation Oncology, Duke Kunshan University 🌍
Purpose: MRI-based automatic detection of brain metastases is often challenged by the small size and subtle nature of metastases. This study aimed to develop a novel deep learning-based brain metastasis detection model that incorporates a locoregional 3D deformation technique to improve the efficiency and accuracy of detecting brain metastases from T1-contrast enhanced MRI scans.
Methods: We hypothesized that magnifying the structures of small and subtle brain metastases in MRI images could enhance automatic detection performance. In this work, a local 3D magnification technique was applied to brain MRI images to create a “fish-eye” like effect, emphasizing smaller metastases while preserving the global anatomical context. Such deformation was systematically applied throughout the entire brain MRI to amplify the visibility of lesions distributed within the brain. The deformed images were subsequently fed to a 3D U-net model for precise segmentation of metastases. The model was developed using a dataset comprising a total of 236 T1-weighted contrast-enhanced MRI scans collected from the BraTS-METS 2023 challenge. A train-validation-test split of 70/10/20 was implemented. A comparative study was conducted using the same CNN segmentation model without incorporating the 3D image deformation technique. The lesion wise Dice, sensitivity, and precision were employed to evaluate the model performance.
Results: The proposed segmentation model with 3D deformation achieved superior performance with an average lesion-wise Dice=0.7023, sensitivity=0.3011, and precision=0.2021. In contrast, the model without deformation yielded significantly lower performance, with a Dice=0.6472, sensitivity=0.2139, and precision=0.1533.
Conclusion: The locoregional 3D deformation technique presents a promising approach for improving the detection and segmentation of brain metastases in T1c MRI scans. Preliminary results indicate that the model is effective in identifying brain metastases with improved accuracy and sensitivity.