Enhanced Cross-Modality MRI Segmentation Using Dilated Convolutions and Multi-Scale Gradient Map

In recent years, convolutional neural networks have significantly advanced image segmentation, particularly for brain images, where important edge features are automatically found. However, accurate segmentation of tumors in a brain remains a challenge across different magnetic resonance modalities, like T1, T2, T1ce, and FLAIR. Using a simple gradient map as an input to the neural networks is not effective due to variations in cross-modality image characteristics. To address this issue, we introduced multi-scale gradient maps that incorporate Holistically Nested Edge Detection (HED) and dilated convolutions into the UNet model. The HED model captures detailed gradient information, enhancing structural feature identification across modalities, while dilated convolutions expand the UNet receptive field for better contextual understanding without increasing parameters. Our method was trained and evaluated on the BraTS2018 dataset. The experimental results demonstrate significant improvements in segmentation accuracy and robustness. Specifically, our method achieved a Dice Similarity Coefficient (DSC) of 0.6902 for T2 to T1ce, 0.6858 for T2 to T1, 0.4329 for FLAIR to T1, and 0.6004 for FLAIR to T1ce, outperforming previous state-of-the-art methods. This demonstrates the effectiveness of our approach in enhancing segmentation performance across different MR image modalities.