Towards automated classification of clinical optimal coherence tomography (OCT) data obtained from dense tissues

• Main Author: Bazant-Hegemark, Florian
• Other Authors: Meglinski, Igor ; Stone, Nicholas
• Published: Cranfield University 2008
• Subjects: 611
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.512841

Cervical cancer can be prevented if its precursors are recognised. Those lesions that justify preventive treatment are currently identified using methods that suffer from delayed results, false positives and subjective judgement. Optical coherence tomography (OCT) is a novel imaging modality that provides high-resolution backscattering data similar to ultrasonography. It could potentially provide in vivo and real-time imaging from within the entire cervical epithelium, where cervical cancer predominantly develops. In this study, we used a bench top OCT system with a 1310 nm light source. It employs fibre optics and operates in the time domain. A collection of 1387 images from 212 ex vivo tissue samples from 199 participants requiring a histopathologic examination of the cervix has been created. Images from this collection were assessed in respect to their benefit in providing markets or evidence of early developments representative of cervical cancer. In our images, the contrast in dense tissue is weak and specific markers that could be associated with a higher cancer risk were difficult to establish. For two reasons it was decided to use an algorithm for classifying the images: 1) Modern OCT systems acquire gigabytes of data per second which cannot be assessed in a clinically meaningful time. 2) An unsupervised classification tool can provide an objective assessment. There is no established method for evaluating OCT images of dense tissue. A classification algorithm was designed that uses Principal Components Analysis as means of data reduction and Linear Discriminant Analysis as a classification tool. This approach does not rely on clinical markets to be designated a priori. The algorithm was applied to the clinical data set to separate samples with mild from severe risk of cancer progression. The performance after leave-one-patient-out cross-validation reaches 61.5% (sensitivity = 66.7%, specificity = 47.3%, kappa = 0.52). These results are not convincing enough to let OCT replace current systems as clinical tools in cervical precancer assessment. Routes for improving results are suggested. This thesis provides a novel, generic algorithm for rapidly classifying OCT data obtained from dense tissues.