Improved Brain Tumor Segmentation in MRI Images Based on Thermal Analysis Model Using U-Net and GPUs

Abstract

This paper presents a new approach to improve brain tumor segmentation in Magnetic Resonance Imaging (MRI) using brain temperature distribution and U-Net convolutional neural network. As we presented in our recent works, brain tumors generate more heat than healthy tissues and based on tumor temperature profile we can determine tumor borders and reinforce segmentation in conventional MRI sequences such as T1-weighted, T1-weighed contrast-enhanced, T2-weighted, and Flair images. In the present work, U-Net architecture was used for brain tumor segmentation from thermal images. The results were compared to segmentation with U-Net in Flair images. Pennes bioheat transfer equation discretized with Finite Difference Method (FDM), was used to calculate the temperature distribution of the brain with tumors, 2% of additional Gaussian noise was added to the calculated temperature and transformed to grayscale thermal images. Then, U-Net was used for brain tumor segmentation from thermal images. We have generated 276 images to train the U-Net model, and 25 images were used to test the model. The dataset is containing thermal and the corresponding Flair images with the ground truth of tumors of the same patient at the same level. The training computation time of U-Net in 276 thermal images was about 15 h using NVIDIA GTX 1060 GPU with 6 GB memory. The obtained segmentation from thermal images was compared to segmentation of tumors from Flair images using U-Net in 10 images. An average of 0.33% of tumors was detected only in thermal images, and an average of 2.05% of healthy tissues was detected only in thermal images. The obtained results demonstrate the importance of thermal information of brain tumors to improve segmentation in MRI towards an effective diagnosis.