Automated classification of cancer tissues using multispectral imagery

Automated classification of medical images for colorectal and prostate cancer diagnosis is a crucial tool for improving routine diagnosis decisions. Therefore, in the last few decades, there has been an increasing interest in refining and adapting machine learning algorithms to classify microscopic images of tumour biopsies. Recently, multispectral imagery has received a significant interest from the research community due to the fast-growing development of high-performance computers. This thesis investigates novel algorithms for automatic classification of colorectal and prostate cancer using multispectral imagery in order to propose a system outperforming the state-of-the-art techniques in the field. To achieve this objective, several feature extraction methods based on image texture have been investigated, analysed and evaluated. A novel texture feature for multispectral images is also constructed as an adaptation of the local binary pattern texture feature to multispectral images by expanding the pixels neighbourhood to the spectral dimension. It has the advantage of capturing the multispectral information with a limited feature vector size. This feature has demonstrated improved classification results when compared against traditional texture features. In order to further enhance the systems performance, advanced classification schemes such as bag-of-features - to better capture local information - and stacked generalisation - to select the most discriminative texture features - are explored and evaluated. Finally, the recent years have seen an accelerated and exponential rise of deep learning, boosted by the advances in hardware, and more specifically graphics processing units. Such models have demonstrated excellent results for supervised learning in multiple applications. This observation has motivated the employment in this thesis of deep neural network architectures, namely convolutional neural networks. Experiments were also carried out to evaluate and compare the performance obtained with the features extracted using convolutional neural networks with random initialisation against features extracted with pre-trained models on ImageNet dataset. The analysis of the classication accuracy achieved with deep learning models reveals that the latter outperforms the previously proposed texture extraction methods. In this thesis, the algorithms are assessed using two separate multiclass datasets: the first one consists of prostate tumour multispectral images, and the second contains multispectral images of colorectal tumours. The colorectal dataset was acquired on a wide domain of the light spectrum ranging from the visible to the infrared wavelengths. This dataset was used to demonstrate the improved results produced using infrared light as well as visible light