1. Visualisation, optimisation and machine learning : application in PET reconstruction and pea segmentation in MRI images
- Author
-
Al-Maliki, Shatha and Vidal, Franck
- Subjects
006.3 - Abstract
This study aims to investigate the behaviour and applications of an Evolutionary Algorithm (EA) based on a particular approach of Cooperative Co-evolution Algorithm (CCEA), the "Parisian approach" where the solution of an optimisation problem is a set of individuals (e.g. the whole population) instead of a single individual (the best one) as in typical EAs. The CCEA we selected is called "Fly algorithm". It is named after flies, because the individuals are extremely primitive and correspond to three-dimensional (3-D) points. The focus of this study relies on visualisation to examine the Fly Algorithm (FA) to solve complex problems - reconstruction and segmentation of two types of medical imaging modalities, Positron emission tomography (PET) and MRI. For image reconstruction, we compare the performance of FA with traditional non-cooperative optimisation schemes, such as Real-Coded Genetic Algorithm (RCGA), Particle Swarm Optimization (PSO) and Covariance Matrix Adaptation Evolution Strategy (CMA-ES) algorithms on two test cases: A toy problem, the Lamps; and a complex inverse problem, PET reconstruction. This choice is based on three facts: i) FA has been built on top of RCGA, the comparison yield an assessment of the cooperative component that has been added, ii) PSO has been sometimes opposed to FA, and iii) CMA-ES is considered as the state-of-the-art for continuous optimisation. Our experiments highlight some structural differences and experimentally compare these algorithms. In both test cases FA exhibits a better scalability. In another work, we propose using information visualisation and user interaction techniques to explore the algorithm's internal data. Our aim is to better understand what happens during the evolutionary loop. Using PET reconstruction, we demonstrate that it is possible to interactively discover when an early termination could be triggered. It is implemented in a new stopping criterion that reduction of the number of iterations without any loss of accuracy. This methodology lead to the segmentation where, we combine optimisation, computer vision and visualisation/data exploration to analyse MRI data and detect peas inside the human stomach. We propose to perform the image analysis task as a multi-objective optimisation. We rely on the FA implemented using NSGA-II. The output of the optimisation is a succession of datasets that progressively approximate the "Pareto front", which needs to be understood and explored by the end-user. Using interactive Information Visualisation (InfoVis) and clustering techniques, peas are then semi-automatically segmented. Once a labelled dataset became available, we performed a binary classification as a food recognition problem, implemented using a deep Convolutional Neural Network (CNN). The results have been analysed using interactive visualisation. We prove in this work that advances in computer vision and machine learning can be deployed to automatically label the content of the stomach even when the amount of training data is low and the data imbalanced. To make the work more robust by taking advantage of the labelled data, we compare the performance of some more traditional machine learning classifiers by using online application. Also, we deploy the multi-objective optimisation NSGA-II to use it as classifier and feature selection to improve classification accuracy and reduce the computational complexity. The result of this classifier is then refined using a residual neural network (ResNet). We have presented a fully-automatic segmentation of the peas using a combination of evolutionary computing, machine learning and computer vision techniques. The final results were confirmed by experts. In conclusion, our investigations confirm that the Fly algorithm works well with a complex search space. We prove the use of a simple but effective visualisation can help to understand the behaviour of the FA and extract early results without need to wait until the optimisation loop finished also, to analyse the output of the FA. We open the door for the researchers to try use the algorithm in segmentation processing images with consideration the description of the object they want to segment it.
- Published
- 2021