Steerable Convolutional Neural Networks

Convolutional Neural Networks (CNNs) are widely used and have an impressive performance in detecting and classifying objects. However, the CNN's performance is sensitive to variations in rotation, position or scaling of the objects to be detected. Fully Convolutional Neural Networks were trained for guidewire detection and retinal vessel detection in this dissertation. We highlight what challenges are encountered during training for the guidewire detection. We present a novel method for simultaneously detecting the guidewire pixels and predicting the guidewire orientation using trained oriented filters. We also show how to train, in the same framework, these oriented filters as steerable filters in a low rank representation. We introduce the Spherical Quadrature Filters (SQF) for guidewire detection and show how they can be used to improve the training data. We propose a steerable CNN that can detect an object rotated by an arbitrary angle without being rotation invariant. The proposed steerable CNN is discriminative like a regular CNN, but it has a latent parameter representing the object's 2D orientation. For any value of this parameter, the steerable CNN will be sensitive to detect only objects having that orientation. We apply the SQF, CNN and steerable CNN to detect the guidewire in fluoroscopy (real-time X-ray) images and to detect vessels in retina (fundus) images. The guidewire is a thin wire used in coronary angioplasty interventions, which are visualized using fluoroscopic images. The fundus images are noisy because of the similarity between the background and the vessels. Experiments show that the steerable CNN outperforms the regular CNN and other popular approaches such as the Frangi filter, the Steerable Quadrature Filter and a state of the art trained classifier based on hand-crafted feature.