Brain tumor MRI images classification based on machine learning

Abstract

Recent research has shown machine learnings outstanding performance on image classifying tasks, including applications on Magnetic Resonance Images. While the former models are overly complicated, this paper proposes a simplified model, which is proven to be both accurate and much less time-consuming. Our proposed method is learned from former research and combines Bias Field Correction, DenseNet, and SE-Net to form a concise structure. With small datasets of T1-weighted and T2-weighted labeled MR brain tumor images, our model spent a short training time of 2 hours and showed excellent performance on classifying pituitary, meningioma, glioma or no tumor with an accuracy of 91.32%. After evaluation, our model is proven to be accurate in distinguishing between 3 of the tumor types with an f1-score of 0.96.