[300]
Team: Independent Design: Deepan Islam
Program: [308]
We present a fast deep learningbased framework for deformable, pairwise image registration (DIRGAN). Whereas conventional registration methods require an iterative optimization algorithm which leads to a high runtime, we treat registration as a functional problem and develop a Generative Adversarial Network (GAN) to estimate the deformation vector field needed to align a pair of input scans. We parametrize the function through two adversarially trained convolutional neural networks and refine the model weights on a dataset of medical images. Given a new pair of scans, DIRGAN can quickly and accurately, as measured by lower computational runtime and higher Dice scores respectively, estimate the appropriate deformation vector field needed to align that set of scans. Experiments on FIRE: Fundus Image Registration Dataset confirm that DIRGAN solves the twodimensional registration task and provides proof of concept for the use of generative models in registration. We will modify DIRGAN to solve the threedimensional registration problem on 3D computed tomography (CT) scans from OASIS3 and if time permits, lung CT scans from COPD patients. The result from the threedimensional registration would be integrated with a pipeline to allow for quantification of bronchial wall thickening and other downstream, clinically relevant subtasks.
Sponsorship
[/if 302]
[if 303 not_equal=””]
Mentors
[/if 303]
[if 303 not_equal=””][/if 303]Andrew Gearhart, PhD
Team Members

[if 306 not_equal=””][/if 306]Deepan Islam
[307 size=large show_image=1]