Regresar

Mammogram classification schemes by using convolutional neural networks

Abstract:

This work presents the comparison of two schemes of mammogram classification based on convolutional neural networks (CNN). The main difference between these two classification schemes relies on the number of bits per image pixel, the feature extraction techniques and the number of neurons in the fully connected layer. We use 1070 mammograms from the Digital Database for Screening Mammography (DDSM), which are divided into two categories: benign and malignant mammograms. We use CNN for classification by applying the open source library TensorFlow which is configured on the high level library Keras. In order to tune our classification model parameters, we apply random and grid search algorithms, by combining the batch size, the number of layers, the learning rate and three optimizers: Adadelta, RMSProp and SGD. We evaluate the classification algorithm performances through the accuracy and two loss functions: Categorical Cross-Entropy and Mean Squared Error. The model with the best accuracy has 85.00%, and a mean squared error of 15.00%.