Challenges and Difficulties of Multi-Spectral MRI Based Brain Tumor Detection and Segmentation

Introduction: The automatic detection and segmentation of brain tumors based on multi-spectral MRI data represents a widely investigated topic, which captured special attention with the Brain Tumor Segmentation challenges organized yearly since 2012 jointly with MICCAI conferences. Up to one thousand different solutions may have been presented by the research community, involving the whole arsenal of artificial intelligence starting with classical machine learning methods and developing together with convolutional neural networks and deep learning. Methodology: This paper proposes to discuss the general challenges and difficulties of automatic brain tumor segmentation methods, comparing the needs and capabilities of a classical machine learning based method using ensemble learning, and a U-net solution involving deep learning, both very popular at their respective time. The main difference of the two approaches is that the first relies on handcrafted features extracted for each pixel separately and applies a classification to estimate the tissue to which each pixel belongs, while the other works with whole slices or volumes at once and extracts its own features needed for decision making. Our team investigated the necessary preprocessing steps of MRI data, and their optimization to suit the needs of both approaches, and trialed various post-processing techniques that can improve the final segmentation outcome after the decision made by the classification algorithm. Results and Discussion: The U-net based solution can achieve finer segmentation quality, reaching Dice scores approximately 2 % higher than the ensemble learning, but its stability is weaker meaning that two different training sessions of the same network may lead to significantly different results. Another main difference consists in the most suitable histogram uniformization technique required, as the ensemble learning works better when the method of Nyul et al (IEEE T. Med. Imag. 19:143-150, 2000) is employed with only a couple of intermediary milestones, while U-net is best supported by the spatial CLAHE technique. Conclusions: Both approaches had their merits at their time but CNN and deep learning based methods are more likely to evolve and be employed in the future.