Data Preprocessing Via Compositions Multi-Channel MRI Images to Improve Brain Tumor Segmentation

Nikolay Tolstokulakov, Evgeniy N. Pavlovskiy, Bair Tuchinov, Evgeniya Amelina, Mihail Amelin, Andrey Letyagin, Sergey Golushko, Vladimir Groza

Abstract: Magnetic resonance imaging (MRI) stays one of the most essential noninvasive methods for brain diagnostics. It allows obtaining the detailed 3D image of the brain, including various types of soft tissues. In this paper, we compare the influence of the multichannel data composition approach on the model’s performance. We consider the binary brain tumor segmentation problem evaluating the Dice, Recall and Precision metrics. One common way to process the medical images with the use of neural networks is to use 2D slices as the input. In contrast to the RGB images, there are plenty of methods of how to combine the multi-channel MRI data structure into the common format for ML-based algorithms. After evaluating several possible combinations we demonstrate the most performance improvement by 6–7% in Dice & Recall metrics using the pseudo-RGB approach.