کنفرانس مهندسی برق ایران

صفحه اصلی / سی امین کنفرانس بین المللی مهندسی برق

DRAU-Net: Double Residual Attention Mechanism for automatic MRI brain tumor segmentation

نویسندگان :

Mohammad Soltani gol¹ Morteza Fattahi² Hamid Soltanian zadeh³ Samd Sheikhaei⁴

1- دانشگاه تهران 2- دانشگاه تهران 3- دانشگاه تهران 4- دانشگاه تهران

کلمات کلیدی :

Attention mechanism، brain tumor segmentation، deep convolutional neural networks، magnetic resonance imaging، residual blocks، U-net

چکیده :

Abstract— Accurate tumor segmentation is necessity for reliable diagnosis and treatment of brain cancer. Glioma is a very common and life-threatening type of brain tumor. Various Magnetic Resonance Imaging (MRI) modalities contribute to segmentation accuracy since they provide complementary information. Deep Convolutional Neural Networks (DCNNs) have provided remarkably good performance in the field of image segmentation. However, because of difficulties in detecting gliomas due to their intensity and shape variations, development of an efficient network with an appropriate loss function is needed. DCNNs developed for segmentation include 2 main parts. The first part performs as an encoder and extracts spatial information, while the second part generates a full resolution probability map. Our proposed network is based on the U-net structure. We use double residual blocks to generate a unique mapping to activations earlier in the network and neutralize the decay problem. Besides, we utilize an attention mechanism after the low and high-level features for adaptively weighing the channels. Instead of feeding the 3D data to the network, we use 2D plains in the axial view. We tested the network successfully using the Brain Tumor Segmentation (BraTS) 2018 dataset. We used the Dice score for network evaluation and obtained scores of 0.891, 0.849, and 0.834 for WT, TC, and ET, respectively.