Classification of Brain Tumor Leveraging Goal-Driven Visual Attention with the Support of Transfer Learning
Kazihise Ntikurako Guy-Fernand, Juanjuan Zhao, Fredy Malack Sabuni, Jiawen Wang
Abstract
The task of distinguishing types of tumors in brain MRI scans is of paramount importance in the medical field. In the recent years, deep learning methods have produced leaps of advancements, especially in the field of computer vision, which allowed to address important challenges such as image classification in the medical field. The rise of convolutional neural networks brought about the achievement of new heights in many tasks, include medical image analysis. They produced very good results in terms of accuracy and performances, by recognizing repeating patterns and shapes in a image, they still fall short of taking advantage of spatial information. As inspiration was also taken from natural visual system, concepts visual attention have also inspired architectural advancements that yielded higher performance in computer vision tasks. Visual attention allows to differentiate relevant areas of a picture for improved recognition abilities. In this paper, we propose using a pre-trained attention mechanism for the task of brain tumor classification. We counter-intuitively use transfer learning on a specific, restricted dataset for the goal of finding relevant areas of brain tumors. This goal-oriented attention mechanism used in a general convolutional-based architecture allowed us to achieve state of the art classification performance on a brain tumor dataset.