U-net Based Deep Architecture For Brain Tumor Segmentation

Résumé: Deep learning has achieved very high and significant results in many interesting fields, the medical imaging field is one of these active areas and it is advancing each day. In this work, we are mainly interested in brain tumor segmentation. Glioma is one of the most common brain tumors, it is divided according to its grade. MRI images are relevant and commonly used on a wide scale by scientists in the diagnostic of LGG images this is what makes it recommended for efficient results. Accurate brain tumor diagnosis and the ability to identify size, location, and shape are very important to save patients’ lives. In view of the impressive performance of U-Net architecture in brain tumor segmentation, as revealed by several literature studies, we propose in this work a U-Net-based architecture for brain tumor segmentation. In particular, we considered an ensemble learning scheme in which three pre-trained networks are incorporated to achieve the final decision. These networks are MobileNet, DeepLabV3+, ResNet, and DenseNet we use them as an encoder part with the U-Net architecture then ensemble learning is applied in many ways to get the best result. However, our methodology could be well generalized as well as could be investigated by so many other architectures and meth ods. Generally, our obtained results were promising for IOU, Dice-coeff, and accuracy we achieved 0.86, 0.92, and 0.99 respectively thus our followed method improves the importance of applying deep learning in the brain tumor segmentation domain

Publié dans la revue: