Performance-aware programming for intraoperative intensity-based image registration on graphics processing units

Abstract

Recent advancement of intraoperative imaging technologies allows real-time view of the tissue morphologies to ensure safer and efficient interventions. Particularly, intraoperative imaging is extensively used in surgical scenarios that require accurate target localization and precise movement of surgical tools. Stereotactic neurosurgery and cardiac catheterization are typical examples. However, the intraoperatively acquired image may not be aligned with the preoperative image used for planning, due to motion, gravity, or interventions. Intensity-based non-rigid image registration is able to resolve such misalignment, but it suffers from prolonged registration time due to its high computation requirement. This extended registration time makes intensity-based registration inadmissible to the highly dynamic surgical scenarios. To allow seamless application of intra-operative application without disrupting the surgical workflow, there is a constant demand for having a fast intensity-based registration. Graphics processing units (GPU) have attracted the most attention in the recent years due to its unmatched parallel computing power. However, many works of GPU-based image registration have overlooked the underlying memory transaction patterns, which can hamper computation efficacy if not appropriately managed. In view of achieving fast computation, performance-aware programming is a specialized practice that involves repeated profiling, micro-benchmarking, and code optimization to ensure full device utilization. In this thesis, performance-aware programming techniques were employed on GPU to resolve for the high computation requirement in the diffeomorphic log-demons algorithm, which is one of the most popular intensity-based image registration algorithms. The GPU implementation of the algorithm was tested and analyzed extensively. By successfully pinpointing and optimizing for the blocking operations, significant (>200×) performance speed-up has been achieved as a promising result.