Reconstrução de imagem mamográfica com alta qualidade usando deep learning com imagens de baixa dose

This paper presents the preliminary findings of a research whose main objective is to develop methods to reduce the radiation dose required for mammograms without compromising the exams result. The approach used is to improve the quality of images obtained with low radiation doses by detecting characteristic patterns of images resulting from high radiation doses and joining they with the original low radiation images. The quality of images obtained with low radiation doses is lower than those obtained with high doses because there is an increase in quantum noise in the first ones. Our method uses wavelet and deep learning algorithms and was tested on a set of mammogram images obtained with lower radiation doses. Regarding the peak-signal-to-noise ratio (PSNR) metric, there was a media increase from 37 in the control group to 58 in the test group, and the highest PSNR value obtained by an image was 60.7485683800. Regarding the structural similarity index measure (SSIM) metric, the average 80% of the control group increased to approximately 90% in the test group, and the highest similarity value found in an image was 0.912399596. Preliminary results indicate real benefits for patients by showing that it is possible to obtain quality images close to those obtained at high doses by exposing patients to reduced doses