A NOVEL SYSTEM FOR BRAIN TUMOR SEGMENTATION USING ARTIFICIAL NEURAL NETWORKS AND 3D SIFT

Brain tumor segmentation is an essential step in medical image analysis for surgical planning, cancer treatment, or monitoring of therapy. The task is to classify each voxel in a magnetic resonance image as either normal or tumorous tissue, with four sub-categories for the latter. Clinical guidelines rely on manual segmentation, which is time-consuming and requires expert work. It is of great interest to develop a fully automatic, efficient, and accurate segmentation system.

The novel system relies on a histogram matching normalization procedure, an extension of a technique called SIFT, and an Artificial Neural Network. It was predicted that using the novel system would lead to effective processing for tumors of different shapes, sizes, and locations in the brain.

Test results show that the novel system had an overall average Dice accuracy score of 90.47% for whole tumor segmentation. This is competitive with current leading algorithms, including those using different methods besides neural networks. Other studies have shown that manual segmentations by trained experts have intra- /inter-rater variability up to 20% and 28%, respectively. This also reveals that the novel system is of comparable or greater accuracy than its manual counterparts.

Hence, the goal of the project has been achieved. A fully automated, cost-effective, and efficient brain tumor segmentation system is proposed, and it is very promising for future clinical use. This system is both practical, as it is competitive with manual expert segmentation, and versatile, as it can also be applied to image segmentation and analysis for other medical fields.