Investigation of Human Skin Tissues Using Optical Coherence Tomography

Abstract

Over the decades various imaging technologies have been used for the investigation of skin tissues, but their poor sensitivity, specificity and accuracy limits their applications. In comparison to other imaging modalities, optical coherence tomography (OCT) is a preferred technique as it is a non-invasive imaging modality with a high resolution and is able to perform cellular level imaging as well as providing depth information. This imaging modality has been widely used to image tissues in the human body and thus manifests its potential for clinical applications. Further, OCT can be considered as the potential tool for the identification but the modern high-speed OCT system acquires huge amount of data, which will be very time-consuming and tedious process for human interpretation. However, OCT was used for a qualitative investigation of the human skin tissue but does not employ the automatic classification of the tissues (i.e. healthy and unhealthy tissue). This thesis research work describes the possibility of fully automated quantitative assessment based on morphological features of human skin tissue, which will become biomarker for the removal of non-viable skin. We developed an automated algorithm for the classification of infected and normal human scalp in-vivo, using OCT images. The resulting algorithm gives a prospective approach for scalp characterization, which presents tangible findings in normal and fungal-infected scalps by statistical means. Our proposed automated procedure entails building a machine learning based classifier by extracting quantitative features of normal and infected scalp images recorded by OCT and obtained good sensitivity and specificity. Furthermore, the study was performed for the classification of thermally damaged tissue using polarisation sensitive (PS-OCT) images. It is ascertained that the birefringence of the damaged tissue changes due to the change in the alignment of epidermis and dermal layer and can be detected and quantified using PS-OCT. The output is also correlated with the corresponding histopathology images. Results suggest that PS-OCT can be considered as a strong aspirant for robust and automated diagnosis of thermal damaged tissue. Next, for accurately and automated identification of breast cancer tissue a robust machine learning approach was used. Textural features were extracted for the investigation of breast cancer tissue. The best-first search algorithm was used to select the best subset of features for each base classifier. The result obtained from multi-level ensemble classifier has better performance metrics for the diagnosis of breast cancer and will help to perform a more precise biopsy or intraoperative margin assessment. In the last part of this thesis, the hybrid deep convolutional neural network (CNN) architectures was employed for identifying the burn skin margin using OCT images. We evaluate the performance of commonly used CNNs (VGG19, ResNet50, and Inception-V3) on the information extracted from patch and whole slide images for the classification the burn and normal tissue. It overcomes the major challenge of the traditional patch-based approach, which shows good performance classifying isolated patches, but unable to maintain the same performance of the whole slide image.