An Optimized U-Net Encoder with VGG-16 and Additional Feature Block for Brain Tumor Segmentation on MRI ADC Images

Determining the brain tumor area on ADC MRI images is challenging due to the low contrast and variability of water molecule diffusion. This variability affects the differentiation between tumor tissue and healthy tissue, resulting in unclear tumor boundaries. Automatic segmentation of the tumor area using other MRI sequences such as T1C/FLAIR MRI is an effective way to determine the brain tumor area on ADC MRI images by merging the predicted mask from the segmentation results with the ADC MRI image where both have the same dimensions. This research proposes optimizing U-Net with VGG-16 and additional feature blocks to improve brain tumor segmentation on ADC MRI images. The improved U-Net model achieves a dice coefficient of 0.9153, a substantial improvement over the original U-Net model's score of 0.8612, representing a 6.3% enhancement in segmentation accuracy. This demonstrates the effectiveness of incorporating VGG-16 and additional feature blocks in improving segmentation performance. The integration of the segmentation mask with ADC MRI scans allows for accurate identification of the tumor area, showcasing the efficacy of the proposed methodology.