Venous leg ulcers assessment using convolutional restricted Boltzmann machines with genetic programming for thermal imaging systems

Abstract

Deep learning techniques have been extensively used throughout the last years due to their promising results in a wide range of applications, such as medical image analysis. A traditional approach used to solve this problem is the well-known Restricted Boltzmann Machine (RBM), capable of extracting features and learning good data representations. Nevertheless, one of its main shortcomings is to ignore the 2-D structure of images. An improved version of the RBM, known as Convolutional Restricted Boltzmann Machine (CRBM), employs a convolutional concept to its architecture, where CRBM's hidden and visible layers weights are shared, mitigating the 2-D structural problem. However, there is still the problem of adequately selecting a set of parameters. Therefore, in addition to studying CRBMs, this proposal aims at using Genetic Programming (GP), a meta-heuristic optimization technique, to select CRBMs' most suitable parameters. Additionally, the proposed approach will be validated in the context of venous leg ulcers thermal images under the supervision of Prof. Dinesh Kumar, Royal Melbourne Institute of Technology. (AU)