Thesis about resource efficient automatic segmentation of medical images
How to improve resource efficiency in work with medical images by automatic segmentation is in focus for Minh H. Vu’s doctoral thesis at the Department of Radiation Sciences.
Text: Ola Nilsson
The core benefits of convolutional neural networks (CNNs) are weight sharing and that they can automatically detect important visual features. Minh H. Vu and his group found that CNNs are very efficient in automatically segmenting tumors, organs, and structures, which means that CNNs can save radiation oncologists much time when delineating.
First, an end-to-end cascaded deep learning network is effective and promising for quantifying uncertainty in the segmentation of medical images. Second, the proposed novel loss function, the so-called “data-adaptive loss function,” demonstrated that it can address diverse issues in deep learning, including imbalanced datasets, partially labeled data, and incremental learning. Third, one of the works, designed for compressing high-dimensional activation maps, showed that it induces a regularization effect that acts on the layer weight gradients. By employing the proposed technique, we reduced activation map memory usage by up to 95 per cent
Overall, this thesis aims at the classification and segmentation of medical images. Both public and in-house datasets were used. The deep learning architectures used in this thesis were generative adversarial networks (GANs) and convolutional neural networks (CNNs). We also used numerous methods throughout the thesis: statistical tests (Friedman test followed by a Nemenyi post-hoc test) to find the methods that are significantly different from the others, hyper-parameter search, cross-validation, and ensemble, to name a few methods.
Minh Hoang Vu is currently a doctoral candidate at the Department of Radiation Sciences, Umea University. He received his M.Sc. degree in Erasmus+ Joint Master Program in Medical Imaging and Applications (MAIA) in 2018 and an M.E. degree from Nanyang Technological University, Singapore, in 2015. His research interests include semantic segmentation, network compression, generative adversarial network, and visual question answering.