Friday, 6 October 2017
SPEAKER: Russel Mesbah, Department of Computer Science
TITLE: Convolutional Neural Networks for Medical Image Semantic Segmentation
WHERE: Owheo G34
WHEN: 1:00 pm, Friday 6 October
CNNs outperform many other techniques in medical semantic segmentation using learnt hierarchical representations of the patterns of interest. These universal function approximators are capable of inferring low-resolution high-level concepts from the observed data. The deeper the network, the lower the resolution of the deducted abstractions are. The resolution loss through the pyramid of abstractions brings about the trade-off between detection and localisation accuracy. However, enhancing semantic segmentation performance requires the improvement of both the aspects simultaneously.
Many approaches addressed the issue by integrating conditional random fields with CNN, employing fully convolutional networks, reconstructing the low-resolution learnt CNN features at the output using encoder-decoder architectures, or designing different architectures based on the combination of the above techniques. Despite the improvements in performance, these approaches can't fully address the inaccuracy driven from the resolution loss through the layers which is mainly due to the presence of downsampling functions in CNN architectures. Additionally, current CNN-based approaches don't fully exploit information about the local neighbourhood of the pixels being classified.
In this talk, I will address three main issues of using CNNs for medical image semantic segmentation. Moreover, since it will be a summary of my PhD work done over the past three years, I look forward to your comments/advice to help my research.