Low Processing Power Algorithm to Segment Tumors in Mammograms

The diagnosis based on mammography depends on relevantly upon the radiologist’s experience and accuracy in identifying shapes and tenuous contrast in the images. A possible improvement in the mammography consists of image analysis algorithms suitable for detecting and segmenting tumors. However, such an approach usually demands high computational processing. In this context, the present work objective was to develop an algorithm to accurately search and segment tumors in mammography images without requiring high processing power. Thus, efficient, low computational demanding, and automatic image segmentation occurred by performing traditional image processing, including uniform equalization, adaptive enhancement (CLAHE), simple thresholding, Otsu multiple thresholds, and morphology operators. The development of the algorithm divides into two steps. The first step was the isolation of the breast and removing the pectoral muscle. The breast isolation occurred by removing artifacts outside of the breast image. The pectoral muscle removes as this region presents gray levels similar to the tumor. Finally, the second step segmented the tumor mass. The algorithm validation occurred by determining the accuracy and the Dice similarity coefficient, which values (mean ± standard deviation) were, respectively, 0.75 ± 0.09 and 0.71 ± 0.11. These average values resulted after processing 150 images of the CBIS-DDSM database, signifying a good segmentation outcome. As for the processing speed, the algorithm spent 14,5 s to segment a tumor in an image sizing 5383 × 3190 pixels (16 bits), using a conventional computer. The tool was able to segment both tiny and large tumors, and it may represent a convenient approach to assist in the analysis of mammograms without requiring high computational resources.