Development of an Algorithm(S) for Human Tissues Characterization

Abstract

Various imaging technologies have been employed to investigate skin tissues over the years, but their inadequate sensitivity, specificity, and accuracy limit their usage. Optical coherence tomography (OCT) is a promising imaging technique in comparison to other imaging modalities since it is a non-invasive imaging modality with a high resolution that can do cellular level imaging as well as provide depth information. This imaging technique has been widely utilised to examine tissues in the human body, demonstrating its clinical promise. Furthermore, OCT can be regarded a possible tool for identification, however modern high-speed OCT systems capture a large amount of data, making human interpretation a time-consuming and tiresome operation. Computer-aided diagnostic (CAD) systems can support clinicians in diagnosing by rapidly assessing large amounts of data. The goal of this thesis is to create a CAD system that uses OCT for human tissue measurement. The feasibility of fully automated quantitative assessment based on morphological aspects of human tissue, which will become a biomarker for the removal of non-viable skin, is described in this thesis research work. We developed an automated algorithm for the classification of malignant and benign human skin tissue, using the dermoscopic images. The resulting algorithm gives a prospective approach for skin tissue characterization, which presents tangible findings in normal and melanoma infected skin tissue by statistical means. Our proposed automated procedure entails building a machine learning based classifier by extracting the features of normal and infected skin images, augmented with various classical transformations and Generative Adversarial Network. The resultant model obtained good accuracy by adding the synthetic data. Further, a robust machine learning approach was utilised to correctly and automatically identify breast cancer tissue. We presented a novel approach combining pre-trained deep convolutional neural network (CNN) inception-V3 with DCGAN and optical coherence tomography (OCT) imaging modality for classification of human breast tissue. The preliminary results demonstrate the feasibility of using deep learning algorithms with OCT images to perform the automated assessment of breast cancer margin. The results obtained from the classifier has better performance metrics for the diagnosis of breast cancer in terms of accuracy, specificity and sensitivity. In the next part of this thesis, an automated full-field polarization sensitive optical coherence tomography diagnostic system (FF-PS-OCT) was developed which would be more accessible to laboratories as a research tool for the investigation of biological applications. Spatial phase features were extracted for the investigation of breast cancer tissue. A 2D camera was used instead of the photodetector that records the entire en-face image (orthogonal to the optical axis) in the single shot. A number of optical parameters of the tissue obtained from their phase images is used to differentiate between healthy and malignant tissues with SVM classifier. Results suggest that FF-PS-OCT can be considered as a strong aspirant for robust and automated diagnosis of breast cancer tissue. In the last part, a new method to detect skin cancer has been developed, which is more accurate than previous methods. The edges are recognised using a canny edge detector after the input PH2 dataset image is transformed to gray images using gray scale conversion. For edge identification, the canny edge detector employs a multi-stage approach that is smoothed and run through a non-linear kernel-based ICA. SVM and the Naive Bayes classifier were used, which produced accurate results than previous approaches.