A new ensemble method for brain tumor segmentation

Brain tumor localization and segmentation from magnetic resonance imaging (MRI) are crucial and challenging tasks for several applications in the field of medical analysis. Tumor segmentation can help in diagnosis and prognosis, overall growth predictions, Tumor density measures, and care treatment plans needed for patients. However, this task is extremely demanding due to low contrast, noise in medical images, and the voluminous size of the data. This work focuses on improving the semantic segmentation of medical images using a time-efficient ensemble learning approach. We propose a novel ensemble approach using a Convolutional Neural Network and three Autoencoders to extract relevant features from brain MRIs, reduce dimensionality, and then apply supervised learning for pixel-by-pixel binary classification to achieve tumor segmentation. Experiments show very promising results of the proposed ensemble segmentation model when hyper-tuned with Bayesian Optimization, hence, demonstrating the model’s ability to identify the tumor’s pixels with precision while consuming fewer resources. Experiments show that the segmentation performance is sufficiently addressed in terms of accuracy, reliability, and computational speed.