Medical image registration using extremal region based interest points

`Maximally Stable Extremal Regions´ (MSER) based interest points are frequently used for medical image registration on account of their robustness to noise, better localization, and good repeatability. However, if the objects in the image do not have sharp boundaries (as is the case with medical images), the number of MSER´s detected is low. Also, MSER´s are highly sensitive to image blur. This paper proposes new interest points (specifically for medical images) based on `Extremal regions´ (ERs) (as against MSERs). The proposed interest points inherit favorable attributes of MSERs in addition to being detected in good numbers and robustness to blur. Our experiments reveal that ER based interest points perform better, especially for medical images. The computational complexity for proposed points is same as `quasi linear´ as for MSERs using union find algorithm.