Automatic Detection of Injuries in Mammograms Using Image Analysis Techniques Carlos B. Fiallos 1 , Maria G. Pérez 2 , Aura Conci 3 , Víctor H. Andaluz 4 1 Universidad Técnica de Ambato, Ambato-Ecuador, 2 Escuela Politécnica Nacional, Quito-Ecuador, 3 Dep. Ciência Computação, Instituto de Computação, Universidade Federal Fluminense, Niterói, Rio de Janeiro-Brazil, 4 Universidad de las Fuerzas Armadas ESPE, Sangolqui-Ecuador, spacarlos@hotmail.com, maria.perez@epn.edu.ec, aconci@ic.uff.br, vhandaluz1@espe.edu.ec Abstract - Breast cancer is the most common cancer and the second cause of cancer death among women. Early detection is the key to reducing the associated mortality rate, for this identify the presence of microcalcifications is very important. This paper presents an approach for micro calcification detection in mammography based on the following steps: noise reduction, image segmentation, extraction of the region of interest (ROI) and features that describe the possible asymmetries between the ROI of both breasts. The new aspect of our work is how we detect the microcalcifications by using wavelet decomposition. All decompositions were conducted using orthogonal wavelet filter set to computes the four filters associated with the scaling filter corresponding to a wavelet: low-pass filter and high-pass filter. Several mother families have been tested and we are confident to recommend the coiflets as the best one. Keywords - ROI; Microcalcification; Mammographic images; Image segmentation; Texture descriptor. I. INTRODUCTION Mammography is the most efficient, effective and currently the most reliable technique by detect breast cancer at different stages [1]. Early detection increases the survival rate [2] and computer systems to aid in the detection and diagnostics is very important [3][4]. Several studies have been developed and involve lesion detection, classification of regions (tumors, calcifications, etc.), and search for similar cases in databases [5]. An important topic for these systems is the identifications of micro calcification, especially in postmenopausal women. Many micro calcifications are related to benign tumors, but some patterns are related to malignant cases. Micro calcifications, in some cases, are difficult to be detected because it has small size and low contrast, especially if it is superimposed on a dense glandular tissue. Different methods and computer algorithms have been proposed to detect micro calcifications. Table 1 summarizes the most relevant of these from 2006, others previous works are easily found in surveys on this subject. TABLE I. SUMMARY OF SOME RECENT WORKS ON TECHNIQUES FOR MICROCALCIFICATIONS DETECTION Author Year Method and work description Images details Evaluation Area Juarez et al. [6] 2006 Creation of negative image, decomposition by WT, binary image, pre-detecting micro calcifications, identification of pixels by threshold. Applied Daubechies wavelets: db2, db4, db8 and db16. Include characteristics of background tissue (fatty, fatty-glandular, or dense glandular). Consider class of abnormality (calcification, masses and speculated masses). MIAS database: 30 mammograms: 15 with calcifications and 15 with glandular tissue. Accuracy Full image Quinta- nilla et al. [7] 2011 Use top-hat transform to enhancement microcalcifications. Sub-segmentation based on fuzzy c-means algorithm. Window-based features (mean and standard deviation) are extracted from ROI. Neural network used to identify the microcalcifications or healthy tissue. Mini-MIAS. Accuracy, sensitivity, specificity. ROC ROI Bose et al. [8] 2012 Pre-processing for noise removal by adaptive median filtering, change the range of pixel intensity values. Normalization of the image, fuzzy segmentation, 2D DWT, with Daubechies, db1. Neural network for classification into normal or abnormal images. MIAS /322 mammograms Accuracy. Full image Hamad et al. [9] 2013 1-D discrete WT, choice optimal level of WT- 2D approximation coefficient set to zero, detail coefficients are thresholded, image reconstruction. Mini-MIAS: 40 images, 317 with microcalcificatios. TP, TN and FP ROI Grigor- yev et al [10] 2014 Compare ultrasound and mammography in microcalcification detection. The breast was examined by ultrasound (9 MHz, Aplio XG/500) with additional use of 420 images (4 per patient: B-mode and level 1MicroPure images, in sagittal and axial planes) and 105 video of the ultrasound examination. MAMMOMAT Inspiration, Siemens AG, sensitivity, specificity, ROC and AUC Full image Krishna- veni et al. [11] 2014 Chain code, Enhancement, Feature extraction (HOG), Naive Bayes. These images are normalized to 256x256 ROI. Histogram of Oriented Gradients is applied, Gray Level Co-occurrence Matrix (GLCM) and Intensity based features (mean and standard deviations). Fatty, dense and glandular tissue. MIAS Normal:40 and 40 abnormal 1024X1024 Accuracy, sensitivity, specificity, precision. ROI Dheeba et al. [12] 2014 Laws Texture Energy Measures are extracted, Classification by pattern classifier using Particle Swarm Optimized Wavelet Neural Network (PSOWNN). Multi centric clinical database, 216 mam. 54 patients ROC Full image 978-1-4673-8353-0/15/$31.00 ©2015 IEEE 245