Advanced image understanding with deep learning in real-world applications

Abstract

Image augmentation and segmentation are crucial tasks in biomedical imaging applications. Deep learning has attained promising accuracy in many real-life applications. However, most of the cases rely on supervised learning requiring a large amount of labelled data. Obtaining medical images is difficult due to privacy issues, and labelling medical images needs significant time and expertise. Recent works have proposed an unsupervised approach image registration methods to synthesise new medical labelled examples. However, there is still room for segmentation performance improvement. In this project, we aim to improve the augmentation and segmentation performance of the state- of-art unsupervised learning-based registration model. To achieve that, we analysed the behaviours of the state-of-art unsupervised learning-based registration model, and designed image filtering methods for the current model to synthesize more reliable brain magnetic resonance images (MRI) and labels. The image filtering methods are based on Siamese network and classification model. Siamese network selects the unlabelled images that are more similar to labelled reference volume x. The classification model filters out the unlabelled images with irregular shapes. The experimental results show that implementing image filtering and adjusting loss functions provide significant improvements over the state-of-the-art model for segmentation performance. Asides from improving segmentation performance, we also did the experiments on synthesizing more labelled examples through sampling the transformation from a continuous set of spatial transformation.