2D Brain Tumor Segmentation Based on Thermal Analysis Model Using U-Net on GPUs

Brain tumor segmentation allows separating normal and abnormal pixels. In clinical practice, stills a challenging task, due to the complicated structure of the tumors. This paper aims to improve the process of segmentation based on brain tumor thermal profile. Brain tumors are a fast proliferation of abnormal cells, which thermally represent a heat source. In this work, we segment brain tumors using U-Net fully convolutional neural network based on the change on the temperature in the tumor zone. The temperature distributions of the brain including the tumor were generated using the Pennes bioheat transfer equation and converted to grayscale thermal images. Next, U-Net was applied to segment tumors from thermal images. A dataset containing 276 thermal images was created to train the model. As the process of training the model is time-consuming, we used massively parallel architecture based on graphical processing unit (GPU). We tested the model in 25 thermal images, and we obtained a precise segmentation with Accuracy = 0.9965, Precision = 0.9817, Recall = 0.9513, and F1 score = 0.9338. The training time was 20 h in NVIDIA GTX 1060 GPU. The obtained results prove the effectiveness of deep learning and thermal analysis of brain tumors to reinforce segmentation using magnetic resonance imaging (MRI) to increase the accuracy of diagnosis.