Avaliação do contorno ativo e da influência de pré-processamentos na segmentação de nódulos mamários

Abstract

Breast cancer is the most commonly diagnosed neoplasm worldwide, representing one in every eight cancer diagnoses recorded. Early detection and routine mammographic screening, performed through mammography, play a fundamental role in reducing the morbimortality rate associated with the disease and improve the prognosis and the quality of life of affected patients. Radiologistconducted diagnoses become more accurate with the aid of digital processing of mammographic images, enhancing the visualization of breast nodules and the interpretability of mammographic examinations. Thus, this study assesses nodule segmentation in mammographic images based on active contour and the influence of preprocessing techniques on segmentation quality. For the conduction of the tests, 47 images containing nodules from the INbreast image database were employed. Processing was divided into four preprocessing and segmentation techniques: technique 1 involved direct segmentation without any preprocessing of the original image; technique 2 applied a Wiener filter before active contouring; technique 3 used adaptive histogram equalization with contrast limited adaptive histogram equalization (CLAHE) only on the approximation coefficients of the discrete wavelet transform, followed by active contour; and, finally, technique 4 involved using the Wiener filter and a combination of wavelet with CLAHE before active contour application. Active contour segmentation was based on the level-set sparse field method. Results were assessed using eccentricity, Hausdorff distance, quantity of white pixels and visual analysis. For both quantitative and qualitative analyses, the gold standard was established by the ground-truth provided by the image database. The metrics provided an analysis of the format, precision, and extent of the segmented nodule area, allowing a more detailed evaluation of the effects of the techniques used and the validation of segmentation efficacy alongside visual analysis. Thus, it was observed that, for breasts categorized under a and b density patterns, technique 1 yielded the best results. Conversely, for denser breasts classified as c and d patterns, the best results were achieved with technique 3. Therefore, the Wiener filter was found to negatively impact segmentation. Consequently, it is concluded that breast density and applied preprocessing significantly influence segmentation quality. Overall, active contour segmentation proved to be effective for all types of breast densities. However, some images experienced under or over-segmentation. To enhance the findings in this study, there is a need to expand the analysis by utilizing other types of preprocessing, segmentation, and metrics, as well as a larger quantity of images.