CNN optimization using surrogate evolutionary algorithm for breast cancer detection using infrared images

Convolutional neural networks (CNNs) have shown great potential in different real word application. Defining a suitable CNN architecture is vital for obtaining good performance. In this work we propose a random forest surrogate combined with two bio-inspired optimization algorithm, genetic algorithms (GA) and particle swarm optimization (PSO) used to find good CNN fully connected layer architecture and hyperparameters for three state of the art CNNs: VGG-16, Resnet-50 and Densenet-201. The proposed model is used to classify breast thermography images from the DMR-IR database in order to find whether or not the patient has cancer. The proposed model improved F1-score from 0.92 to 1 for the Densenet using the GA and also Resnet from 0.85 of F1-score to 0.92 using the PSO. Moreover, the surrogate model also helped reducing training time.