An investigation of Recent Deep Learning Techniques Applied to Blood Cell Image Analysis

Abstract: This project has investigated the performances of Capsule Networks in comparison to Convolutional Neural Networks (CNNs) on white blood cell image classification at CellaVision. The Capsule Network models that were investigated are EM Routing Capsule Networks (EMCNs) and Dynamic Routing Capsule Networks (DCNs). The models were compared with regards to convergence rate, speed, and accuracy performance on datasets with varying size and complexity. The results show that DCNs outperform the other models on small datasets with regards to accuracy and convergence rate, whereas the CNNs outperform the other models on bigger datasets with higher complexity. With regards to speed, CNNs outperform the other models on both CPU and GPU, with DCNs being very slow. EMCNs, meanwhile, give an indication of learning spatial concepts better, as they are more invariant to spatial- and noise transformations of the underlying test datasets.