Aplicación móvil para clasificar regiones en imágenes mamográficas dicom

Abstract

Breast cancer causes millions of deaths around the world every year, which makes early detection essential. As technology advances, the need for the daily use of increasingly smaller devices is seen. In the present work, a mobile prototype was developed that allows classifying regions of interest present in DICOM mammographic images. A data set of 211 mammography images was used. The process was divided into four phases: image preparation, preprocessing, segmentation, and classification. In the preparation phase, the DCMTK library was used for reading and managing images. In the preprocessing phase with algorithms such as Clahe for contrast enhancement. In addition, the HMF technique to eliminate salt and pepper type noise. The segmentation phase with Canny was used for edge detection and Watershed for segmentation. What allowed to obtain regions corresponding to Masses, Microcalcifications and tissues present in the breast. The HMF results in a maximum average signal-to-noise ratio of all images used of 77 dB. The Canny algorithm presented a 21% CPU usage and Watershed Memory usage in execution of 3709.97 Megabytes (MB). To then be classified using Support Vector Machines (SVM) with accuracy results: 82.43% in masses, 83.78% in micros and 75.36% in masses with micros. Where the classification tends to go down for images of the masses and micros type, and only masses or micros remain at similar values.