Abstract

This thesis highlights the development and evaluation of a hybrid neuroevolutionary approach for designing Convolutional Neural Networks (CNNs) for image classification and segmentation tasks. We integrate Cartesian Genetic Programming (CGP) with Novelty Search and Simulated Annealing algorithms to optimize the CNN architectures efficiently.

The challenge lies in reducing the computational demands and inefficiencies of traditional Neural Architecture Search (NAS) techniques. To address this, a flexible framework based on CGP is utilized for evolving network architectures. Novelty Search facilitates the exploration of varied architectural landscapes, promoting diversity of solutions. Simulated
Annealing further refines these solutions, optimizing the balance between exploring new possibilities and exploiting known good solutions within the search space.

Our experiments, conducted on benchmark datasets such as DRIVE and MSD, demonstrate the method’s effectiveness in generating competitive segmentation models. On the DRIVE dataset, our models achieved Dice Similarity Coefficients (DSC) of 0.828 and 0.814, and Intersection over Union (IoU) scores of 0.716 and 0.736, respectively. For the MSD dataset, our models exhibited DSC scores up to 0.924 for the Heart task, showcasing the potential of our method in handling complex segmentation challenges across different medical imaging modalities.

The significance of this research lies in its hybrid approach that efficiently navigates the search space for CNN architectures, thus reducing number of fitness evaluations while achieving near state of art performance. Future work will explore enhancing the algorithm’s effectiveness through advanced data preprocessing techniques, and the exploration of more complex network layers. Our findings highlight the potential of evolutionary algorithms and local search in advancing automated CNN design for medical image segmentation, offering a promising direction for future research in the field.