Brain tumor detection poses serious challenges not only because the brain is a complex structure, but also because tumors can be serious and often fatal. Of course methods already exist to scan the brain, even in a state of consciousness, and classify the results; luckily, science is an ever-evolving endeavor, so new methods continuously appear that may eventually replace older ones.

This paper presents precisely one such method that registers data from brain scans, extracts from them features of interest, and orders these features according to some classification scheme so that appropriate actions can be taken. The paper starts from existing methods, presented first with a synoptic table and then described one by one with their respective strengths and weaknesses; from this analysis, the authors decide to concentrate on wavelet texture, a promising field of research not yet fully explored. This approach gives a sense of perspective to the paper as a step within a scientific path. This is unusual in a scientific work, and makes it all the more enjoyable reading even for someone not directly involved in this particular field.

After this introduction, the authors devote the central part of the paper to an in-depth description of their new method, which they propose as an extension of existing ones. Again, they present it in general with a flow diagram and then describe each block extensively. The method is composed of three steps: preprocessing, eliminating insignificant data and normalizing the rest; feature extraction, where the focus is on wavelet texture, which, thanks to a more efficient use of available bandwidth, assigns brain images to one of five classes--background, necrosis, edema, enhancing tumor, and non-enhancing tumor; and classification, where the method feeds these results to a random forest classifier, which defines the images as complete tumor, active tumor, or enhancing tumor. In this section, the authors also compare several algorithms for classification and explain why they select the random forest classifier among them; they also develop a computer algorithm to test this method in practice. The next section presents experimental results, performed on a standard laptop computer, from live runs of the algorithm on an existing medical image computing and computer assisted intervention brain tumor segmentation (MICCAI BraTS) dataset containing high- and low-grade data from real patients. Reasons for measuring certain variables (Dice coefficient, Jaccard coefficient, sensitivity, and specificity) over others are explained, and then results are presented in table format. These tables show improvements for all of the previous variables over almost any other method. Of course, in some specific cases, existing methods may still perform better, but on average this new method is far superior. Finally, a concluding part summarizes the place this work occupies within the current state of research and gives an extensive bibliography.

As noted before, the strength of this scientific paper lies in its sense of perspective, which makes for very interesting reading, even for people not directly involved in this particular field of research.