Cancer Detection and Identification on Scarce and Low- Resolution Data

Abstract

Machine learning algorithms have been contributing immensely in the biomedical sector with the innovation of several automatic and semi-automatic diagnostic devices. They are well suited for applications such as the diagnosis of cancer, which is a prevalant and devastating disease nowadays. A computer-aided diagnostic system can detect and classify various tumor tissues and thereby ensures a reliable and rapid screening procedure. They serve as an additional confirmatory tool which is independent of pathologist expertise and experience. In this thesis study, we perform comparative evaluations among several recent approaches for cancer detection and identification on a scarce and low-resolution biopsy image dataset. The biopsy samples comprise of normal as well as cancerous colorectal tissues, collected from the Al-Ahli hospital, Qatar. We have built two separate image datasets, multispectral and optical from the collected samples. Using our multispectral image acquistion system, images are acquired in various wavelength bands spanning from visible to near infrared to build the first dataset. The second dataset is composed of optical images (in RGB raw format) of the same samples. A Multispectral image based tumor identification system was developed using rotation invariant Local Phase Quantization technique and Support vector Machine (SVM) classifier. The comparative evaluations demonstrate that it could outperform Local Binary Pattern for the feature extraction and Random Forests (RF) in classification of the colorectal tumor tissues. We compared the classification accuracies yielded with the two image modalities- multispectral and RGB and the former one exhibited higher accuracy. Furthermore, we have presented a band selection strategy to eliminate the redundant bands from the multispectral imagery. This could reduce the computation time along with improving the classification accuracies. As the main contribution of the thesis, we propose a compact and adaptive CNN approach for the detection and identification of the tumor tissues on the RGB image dataset. This approach is fully automatic with the absence of any manual pre-processing, tuning or prior segmentation phase to aid the classification algorithm. Its performance is compared against the SVM classifier with three different kernel types and five state-of-the-art texture feature extraction methods including rotation invariant Local Binary Pattern, rotation invariant Local Phase Quantization, and Haralick features. The proposed systematic approach with adaptive and compact CNNs and the top performing conventional method with the best texture feature have achieved the highest identification accuracies with respect to the task of discriminating four classes of colorectal tissues. However, the proposed method has achieved the highest cancer detection performance, around 94.5%, as compared to the best detection score of 87% achieved by the best conventional method. This is despite the fact that the proposed method used low-resolution image data (64x64 pixels) in contrast to the original patch resolution (300x300) used by the conventional methods. Finally, the proposed approach can further exhibit a superior computational complexity and minimal false alarms. The promising results throw light on the competence of adaptive CNNs for cancer detection in low-resolution images from a limited dataset.