Detección de Cáncer de Mama a partir de Imágenes Histopatológicas y de mamografías mediante redes neuronales y segmentación de imágenes.

Abstract

The detection of breast cancer through artificial intelligence is a great advance in medical support for the early diagnosis of breast cancer, and to help many women who present with its symptoms. Because it is the most frequent cancer and the one with the highest mortality rate, the main objective of this project is to develop an algorithm based on artificial intelligence for the early detection of breast cancer from histopathological images and mammograms of scanned film using deep neural networks (DNN) to provide the doctor with support in the diagnosis of the patient efficiently, thus preventing metastasis to parts of the body near or far from the breasts. Two types of images are used to classify breast cancer: mammography images (1124 from Cancer Imaging Archive from the CBIS-DDSM dataset), and histopathology images (1204 from Kaggle from the Breast Histopathology dataset). Preprocessing was performed on the primary data: color field change from RGB to grayscale, followed by a binarization stage and filtering by object removal according to their areas (regionprops). Applying Otsu's method, we obtain the mask of the objects of interest on each image in RGB, followed by an adaptive equalization by histogram to increase the contrast of these found areas. On each set of images, data set augmentation techniques are applied using ImagingDataGenerator from the Keras package in the Python programming language, the creation of deep neural networks such as the Residual Neural Network and a ResNet50 was developed for each set of images, in the training the data augmentation was applied by means of iterations, the histopathological image data set performed 33 iterations throughout the training, the mammography image data set performed 28 iterations in total. It is validated which type of image presents better results for the classification of breast cancer. After this, image segmentation was performed to locate benign and malignant tumors applied to the mammogram image dataset; finally, an Api was developed in the flask micro framework, importing the Tensorflow and Keras libraries, loading the best model for breast cancer classification. The EC2 instance of AWS was used to create a virtual environment for hosting the API, this service is performed via SSH protocol.