Computer-Aided Diagnosis Systems in the Classification of Neuroblastoma Histological Images

Publication Type:
Issue Date:
Full metadata record
Neuroblastoma is the most common extracranial solid malignancy in early childhood. Optimal management of neuroblastoma depends on many factors, including histopathological classification. Although histopathological classification by a human histopathologist is considered the gold standard, computers can help to extract many more features, some of which may not be recognisable by the human eye. Neuroblastoma histological images have a complex texture with complicated features which are different from appearance-based features. Computer-aided diagnosis (CAD) systems facilitate the analysis and classification of neuroblastoma histological images which are non-trivial tasks due to the differences in staining, intensity, and instrumentation. This motivates the thesis to work on the classification of neuroblastoma histological images. In the past, a small number of methods were proposed by previous studies for the classification of neuroblastoma histological images. These methods are based on the geometry and appearance of the different cells. However, there is a high intra-class variation of intensity and size of the neuroblast cells within the same classification group. Therefore, these methods are not applicable to neuroblastoma histological images. This research proposes a solution based on traditional machine learning approaches and deep learning approaches to extract non-appearance-based features in small regions. This thesis will investigate two research areas of feature extraction: low-level feature extraction and high-level feature extraction. Low-level features are minor details of the image such as lines, curves and edges. However, high-level features are on top of the low-level features to detect object and larger shape in the image. Feature extraction is aggregated with the classifier in this research to classify neuroblastoma histological images into five categories. This thesis makes four contributions. Contribution 1 is the construction of a dataset comprising neuroblastoma histological images which are labeled by an expert histopathologist. Contribution 2 is the proposal of a local feature extraction method which can extract local features which are robust to high intra-class variations of intensity. Contribution 3 is the extraction of discriminative features which are robust to high intra-class variation of scale of the neuroblast cells within the same class. Contribution 4 is the proposal of deep networks to extract high-level features which are difficult for the human eye to recognise. The performance of all the proposed methods in this research is evaluated on a dataset collected from The Children's Hospital at Westmead, Sydney, Australia. As there was no publicly available dataset in this field, the proposed algorithms were evaluated on the second dataset of neuroblastoma provided by the University of Bristol and the public breast cancer dataset. All the results are compared with state-of-the-art methods. The results indicate the effectiveness of the proposed algorithms. This is the first time that neuroblastoma histological images have been classified into five subtypes using low-level and high-level features. However, there are limitations in this research. The specificity is not 100% compared with the gold standard. Moreover, the proposed algorithms are confused in the distinction between poorly-differentiated and differentiating neuroblastoma, a distinction that human pathologists also find difficult in limited fields of view.
Please use this identifier to cite or link to this item: