Deformable image registration applied to lung SBRT: Usefulness and limitations

Highlights

• Review the status of DIR for lung RT, its main applications, its associated uncertainties and its limitations.

• Large bibliography, classified according to the applications.

• Focus on the application of DIR in clinic.

Abstract

Radiation therapy (RT) of the lung requires deformation analysis. Deformable image registration (DIR) is the fundamental method to quantify deformations for various applications: motion compensation, contour propagation, dose accumulation, etc. DIR is therefore unavoidable in lung RT. DIR algorithms have been studied for decades and are now available both within commercial and academic packages. However, they are complex and have limitations that every user must be aware of before clinical implementation. In this paper, the main applications of DIR for lung RT with their associated uncertainties and their limitations are reviewed.

Introduction

Deformable image registration (DIR) has been studied for more than 20 years and it has a long history with research in radiation therapy (RT). Indeed, the clinical interest is large with numerous applications: motion compensation, auto-contouring, dose accumulation etc. Since the early years of DIR, important progresses have been made: algorithms are faster, more precise and more accessible than ever. However, several challenges and limitations remain such as validation, tissue appearance/disappearance and robustness. This review focuses on applications of DIR in lung cancer and CT images, but DIR can be used in many other sites (head and neck, prostate etc) and with other image modalities (MRI, PET, SPECT, US), every situation having specific challenges. General practical recommendations in RT may be found in the recent AAPM TG report [1].

DIR in a nutshell. First, the main concepts at the core of most DIR algorithms are briefly summarized below. For more details, several excellent reviews are available which cover in depth biomedical image registration methods [2], [3]. DIR is an ill-posed problem formalized as the optimization of a function balancing the similarity between images and the plausibility of the deformation. This tradeoff is at the heart of all DIR algorithms. The three main components are 1) the measurement of image similarity, 2) the parameterization of the deformation and 3) the optimization method. Image similarity can be estimated via numerous approaches, e.g. the popular Mutual Information metric, or metric mixing voxel-based and geometrical extracted features. Deformation vector fields (DVF) may be directly estimated or they may be parameterized with fewer unknowns, e.g. using the popular B-spline basis functions. The cost function is composed of an image similarity term and a transformation plausibility term. It may be optimized via gradient-based continuous methods or discrete approaches (graph-based). This is a very active field of research – around 150 publications per year in PubMed in the last few years – applied to a wide range of applications. In RT, usage of DIR has significantly progressed [3], particularly for thorax images. However, ten years after our review optimistically presenting the potential of DIR in IGRT [4], it can be observed that clinical use of DIR is “like sex for teenagers: everyone talks about it, nobody really knows how to do it, everyone thinks everyone else is doing it, so everyone claims they are doing it too¹”.

Evaluation Like other key components, such as the dose computation engine in a TPS, it is necessary to evaluate DIR. However, a ground truth is generally not available. Indeed, the accuracy is often measured via anatomical landmarks, e.g. bifurcation of airways, using Target Registration Error (TRE) criteria averaging the distances between landmarks. Additionally, other anatomical structures, e.g. lines corresponding to vessels or organ contours can be used. Several open databases of thoracic images with their corresponding evaluation data are available (see Table 1) and they have proved to be very useful as demonstrated by their high number of citations. For example, the EMPIRE10² challenge [5] compared more than 40 algorithms with a database of 30 pairs of thoracic CT images: the first 10 methods depicted TRE lower than 0.9 mm. Instead of relying on manually defined landmarks or segmented delineated structures, other authors have proposed to automatically estimate local DIR uncertainty [6], [7], [8]. More detailed analysis on DIR evaluation may be found in [9], [10].