Desarrollo de una herramienta para detección de tejido anómalo en mamografías digitales usando redes neuronales convolucionales

Abstract

Nowadays, convolutional neural networks have gained popularity and are increasingly used in the field of medicine, thanks to its ability to detect patterns in image classification by discerning certain specific characteristics for detecting lesions or tumors. For this work it will be used in the detection of calcifications in mammography images. This technical project uses the Mini-MIAS database that contains digital mammograms in which a preprocessing process is performed, allowing to modify its contrast and increase the number of images, and thereby providing a sufficient number of calcification mammograms. Additionally, a convolutional neural network is used together with the transfer learning technique, for which the vgg16 neural network model is used having the advantage of its previous training and the versatility of modifying and adapting it to meet the objective of performing the classification of digital mammograms containing calcifications between benign or malignant. The neural network training process was carried out with 70% of the images with calcifications and the validation of the algorithm with 30% of them. Finally, based on the confusion matrix obtained from the final model, metrics such as precision, recall, accuracy and specificity were calculated.