CMIC Seminar: Tim Boers and Adeyemi Akintonde

About

Tim Boers

Title: Imaging in pancreatic cancer and interactive segmentation with convolutional neural networks

Pancreatic cancer has rising incidences yet with an extremely low 5 year overall survival that has not been improved for the last decades. Even today, clinicians understand very little about the mechanisms underlying the development and characteristics of the disease. CT remains one of the most widely-used imaging technique to diagnose, monitor and evaluate the malignance with rich functional and anatomical information. First, I will give a brief overview on the role of CT imaging in pancreatic cancer. I will then discuss computational methods to aid the analysis of the pancreatic CT, in particular, the work on interactive segmentation of pancreas that has traditionally been posed as either a labour-intensive manual task or an automated computer segmentation algorithm.

Adeyemi Akintonde

Title: ROI Data Driven Respiratory Motion Extraction from Cone Beam CT for Radiotherapy Application

The ability to identify respiratory motion prior to radiation therapy treatment and planning verification would be beneficial to patient. Currently most clinical radiotherapy treatment machine have gantry-mounted cone beam CT devices. However, CBCT acquisition takes approximately 1 minute, and such a relatively long acquisition time causes respiratory induced motion artefact in the reconstructed images. Currently there are two general approaches to correct for this: Respiratory binning to form 4D-CBCT data, and the use of a respiratory motion model to compensate for motion during the image reconstruction. Both methods require a surrogate breathing signal that can be easily related to the actual patients motion. In this study we introduced a novel data driven method based on principal component analysis (PCA) to extract a signal related to respiratory motion from cone beam CT projection data. Our method takes into account that projection data acquired on cone beam CT devices typically have two motion components: (1) respiratory induced motion and (2) detector rotational induced motion. Our approach is based on selecting region of interests within the projection data, removing background information, computing PCA for different sections of the data set independently, and introducing a technique of combining the extracted signals from each section in a manner to represent the respiratory signal from the entire data set. We validated the proposed method on simulated data with real ground truth respiratory motion under different clinical conditions. In addition, we provide proof of concept on real CBCT projection data of one patient with encouraging results.

Visitors from outside UCL please email in advance.