Lung emboli detection

A project of the Biomedical Image Analysis group, Department of Biomedical Engineering, Eindhoven, together with Philips Healthcare, Best.
PhD student: Henri Bouma (thesis 02-04-2008)
Supervisors: prof. Frans Gerritsen, prof. Bart ter Haar Romeny, dr. Anna Vilanova

Pulmonary embolism

Blood cloths in the blood stream are dangerous. When they come into narrower vessels, they may block the flow. When this occurs in the in the lungs we speak of a pulmonary embolism (PE). There are more than 50 cases of PE per 100,000 persons every year in the USA. Of these cases, 11% die in the first hour and in total, the untreated mortality rate of PE is estimated to be 30% versus 2.5% for appropriately treated PE. 

A lung illustration depicting a pulmonary embolism as a thrombus (blood clot) that has travelled from another region of the body, causes occlusion of the pulmonary bronchial artery, leading to arterial thrombosis of the superior and inferior lobes in the left lung. From: Wikipedia.

High resolution CT is the preferred diagnostic imaging test. Source: Healthline.

In 2003, prof. Frans Gerritsen of Philips Healthcare (part-time professor in the TU/e-BMIA group) asked for assistance to design an automatic image analysis method for PE from 3D CT. We hired Henri Bouma as a PhD student.

Today we learn features from lots of data with deep learning, in those days we designed proper features.  PE appears as a filling defect inside the pulmonary arteries in CT images (see alto the entry picture on top of this page).

Flow chart of the computer-aided diagnosis (CAD) system for pulmonary embolism consisting of vessel segmentation, candidate detection, feature computation and classification. From Bouma 2008.

Source: URL

One particular feature was the local vessel shape, for PE in particular the deviation of a vessel from a 3D tubular shape. This can be measured from the ‘shape space’, i.e. the principal curvatures of a point on a surface.  If we measure the curvature in all directions from a point, there is a maximum and a minimum curvature, always in perpendicular directions. These are the principal curvatures, mathematically found by the eigenvalues of the Hessian matrix, i.e. the second order derivative matrix. 

You can play with these principal curvatures yourself in the interactive plot below. K1 and K2 are the principal curvatures: change them with the sliders (press the + to see the values. Double click the richt top-arrow bracket to see the code). If both principal curvatures zero: surface is flat. If one is zero: cylinder. If both are positive: ellipsoid or sphere. If different sign: saddle point.

Further reading: Henri Bouma (2008), Vessel-diameter quantification and embolus detection in CTA Images, PhD thesis Eindhoven University of Technology.