Comparison of Key Point Detector Methods for Microcalcification’s ROI Identification on Breast Images: An alternative to SIFT

This document analyses different key point detectors: FAST (Features from accelerated segment test [1]), ORB (Oriented FAST and Rotated BRIEF [2]), MSER (Maximally stable extremal regions [3]), KAZE ( KAZE features [4] ), AKAZE (Accelerated-KAZE [5] ), STAR (CenSurE based STAR feature detector [6]), BRIEF (Binary Robust Independent Elementary Features [7]) and BRISK (Binary Robust Invariant Scalable Keypoints [8]) on mammography images from MIAS dataset [9] seeking for the better choice to extract ROI (regions of interest) with microcalcifications which will allow reducing noise and time of execution in microcalcification detection. An Image enhancement was implemented by filtering with a gaussian mask, thresholding to zero, and applying CLAHE to improve contrast. For ROI detection a K-Mean algorithm was implemented for grouping key points, the ROI’s area must include the microcalcification. An analysis was performed using 3 different parameters for each method: 1. if the microcalcification area is included in the ROI (called fitting), 2. time of execution, and 3. the mean number of points generated. The BRISK method had better performance, resulting in the best choice for mammography analysis, with a 100% of fitting, a time of execution of 0.38 seconds, and 408 points per image. This shows that a good Key point method selection may improve the future performance of an AI model for mammography analysis.