Computational Model Based on CNN to Identify Masses from Liver Images

Therefore, Convolutional Neural Networks (CNNs) are capable of predicting or extrapolating from identified patterns to unknown data, enabling the analysis and extraction of knowledge from large images, in order to produce consistent outputs for inputs not previously presented. On the downside, the training process takes a long time as it is computationally expensive. Aiming to reduce the computational time to perform the training of the neural network (NN), and the analysis of images during the testing process, as well as taking advantage of the natural distributed characteristic of ANN, it was decided to transfer the processing-intensive parts of the tool from CPU (Central Processing Unit) to a GPU (Graphics Processing Unit). Finally, simulator tests were carried out, with consequent verification of the effectiveness of diagnoses on ultrasound images of the liver. The accuracy and total processing time for carrying out the analyses proposed during this research were then evaluated, making a comparison between the network training and testing times obtained on the CPU and GPU. It can also be seen - in both Tr&Tt cases - that GPU had a higher gain than CPU (about 10.6 times faster), even when dealing with a small database, with 120 images of 202500 pixels each.