Master thesis presentation by Juan Pedro Vigueras Guillén, MPBME

Welcome to the master thesis presentation: “Automated detection of immune cells in IHC images presented by Juan Pedro Vigueras Guillén, MPBME”

Wednesday, February 25th at 13.00
Lunnerummet (room 3311), Hörsalsvägen 11, 3rd floor

Examiner: Olof Enqvist

Abstract
Chronic Obstructive Pulmonary Disease (COPD) is a progressive disease characterized by a persistent blockage of airflow from the lungs. The main cause is cigarette smoking, although long-term exposure to other lung irritants, such as air pollution or chemical fumes, may also influence. The development of more effective treatments is hampered by the lack of knowledge of the disease, especially concerning how the damaged tissue become inflamed. Therefore, further study of those immune responses could lead to a better understanding of the disease.

The standard technique to study immune cells is called immunohistochemistry (IHC). This procedure makes use of the fact that antibodies bind to specific to antigens. Hence one can detects antigens or proteins in tissue sections by attaching fluorescent dye or enzymes to the right antibodies. Manual analysis of the resulting images is extremely time-consuming, and thus, there is a need for an automated system.

The aim of this thesis is to explore and implement an algorithm that detects cells in IHC images with large reliability. Two different methods are suggested: The first one, which can be described as `a rank-SVM algorithm for cell detection’, only provides detection, whereas the second one, `a flooding algorithm for cell detection’, also provides segmentation.

The first approach consists of a difference-of-Gaussians detector, a ranking support-vector-machine (rank-SVM) classifier, and an optimization block that finds the best solution. The goal is to rank a set of points in the image, called hypotheses, so that those at the centre of cells obtain a higher rank and thus a threshold can be employed to perform the classification. The features used for ranking are histogram of gradients, mean colour, colour standard deviation, and colour histograms. The second algorithm, includes a block between the detector and the classifier to perform segmentation using a variant of the watershed algorithm. The two algorithms provide similar results, obtaining high sensitivity and specificity.