Deep Cascade Learning for Optimal Medical Image Feature Representation

Abstract

Cascade Learning (CL) is a new and alternative form of training a deep neural network in a layer-wise fashion. This varied training strategy results in different feature representations, advantageous due to the incremental complexity induced across layers of the network. We hypothesize that CL is inducing coarse-to-fine feature representations across layers of the network, differing from traditional end-to-end learning, advantageous for medical imaging applications. We use five different medical image classification tasks and a feature localisation task to show that CL is a superior learning strategy. We show that transferring cascade learned features from cascade trained models from a subset of ImageNet systematically outperforms transfer from traditional end-to-end training, often with statistical significance, but never worse. We demonstrate visually (using Grad-CAM saliency maps), numerically (using granulometry measures), and with error analysis that the features and also errors across the learning paradigms are different, motivating a combined approach, which we validate further improves performance. We find the features learned using CL are more closely aligned with medical expert labelled regions of interest on a large chest X-ray dataset. We further demonstrate other advantages of CL, such as robustness to noise and improved model calibration, which we suggest future work seriously consider as metrics to optimise, in addition to performance, prior to deployment in clinical settings.