Proceedings Article | 12 March 2018
KEYWORDS: Image segmentation, Computed tomography, Brain, Neuroimaging, Convolutional neural networks, Head, Injuries, Magnetic resonance imaging, Tomography, Tissues
While Computerised Tomography (CT) may have been the first clinical tool to study human brains when any suspected abnormality related to the brain occurs, the volumes of CT lesions usually are usually disregarded due to variations among inter-subject measurements. This research responds to this challenge by applying the state of the art deep learning techniques to automatically delineate the boundaries of abnormal features, including tumour, associated edema, head injury, leading to benefiting both patients and clinicians in making timely accurate clinical decisions. The challenge with the application of deep leaning based techniques in medical domain remains that it requires datasets in great abundance, whilst medical data tend to be in small numbers. This work, built on the large field of view of DeepLab convolutional neural network for semantic segmentation, highlights the approaches of both semantics-based and patch-based segmentation to differentiate tumour, lesion and background of the brain. In addition, fusions with a number of other methods to fine tune regional borders are also explored, including conditional random fields (CRF) and multiple scales (MS). With regard to pixel level accuracy, the averaged accuracy rates for segmentation of tumour, lesion and background amount to 82.9%, 85.7%, 85.3% and 81.3% while applying the approaches of DeepLab, DeepLab with MS, DeepLab with MS and CRF, and patch-based pixel-wise classification respectively. In terms of the measurement of intersection over union of two regions, the accuracy rates are of 70.3%, 75.1%, 77.2%, and 63.6% respectively, implying overall DeepLab fused with MS and CRF performs the best.
