Development of a Biomechanical Basis for Lung Image Registration
Hamed Minaeizaeim · ResearchSpace (University of Auckland) · 2019
Respiratory disease places an exceptionally high economic and social burden on society. Due to limitations of pulmonary function tests, most lung abnormalities are diagnosed using imaging techniques such as computed tomography and chest X-ray. In the clinical setting, in order to diagnose, prognose and evaluate lung diseases and overcome limitations of each imaging technique, multiple images may be taken at different volumes or postures. To improve interpretation, these images need to be accurately mapped together to relate information in one to another. Although a number of image registration techniques have been developed, either based on image intensity or landmarks, these techniques are not robust when significant changes in lung volume or postural changes occur, as the lung is highly deformable. Furthermore, they are not typically constrained to physical tissue deformation, so their results can be non-physical. In this thesis, a physics-based lung image registration using finite element method was developed and incorporated into an existing intensity based free form registration. When we breathe, the shape of the lung changes non-uniformly, as most of the lung is constrained by the chest wall but the diaphragm moves more freely. This deformation plays an important role in the physiology and mechanics of breathing. However, no biophysical accurate model has been published on the effect of the pleural cavity shape changes during breathing or posture changes. In this study, quantitative measurement of how the shape of pleural cavity for the left lung changes between two different postures and volumes were made. Then, cavity shape changes were incorporated in a biophysically based model of lung tissue deformation. Both left and right lungs deform significantly during the breathing cycle and both lungs have similar physical structure, and so likely material properties. In this study, the left lung was selected as it has more complex deformation than the right lung due to the location of the heart, which has been proposed to interact with the lung and influence its deformation. The model was assessed in healthy subjects imaged at functional residual capacity and total lung capacity in supine posture and ten healthy subjects imaged at total lung capacity in supine and prone postures. The biophysical model of the lung was used to develop a physics-based lung registration that can map the material points between a source and target image. This method can register lung images despite different postures and volumes. Furthermore, a hybrid method combining the biophysical model and free form deformation were developed to create a robust registration methodology that can rely on both the physics of the lung and image intensity. This novel registration technique was examined in two case studies of clinical interest. The first is in an adult population where multiple high-resolution computed tomography (CT) images are available, a cohort with idiopathic pulmonary fibrosis to register multiple CT images in the same subjects at different time points. Idiopathic pulmonary fibrosis can be difficult to assess as patients may find it difficult to breath to reproducible volumes at repeat visits, in this study we show how registration can help to quantitatively evaluate progression of disease features in imaging by mapping data to a consistent lung volume. The second case focuses on a more challenging population, a cohort of children with cystic fibrosis, for whom both high-resolution computed tomography and X-ray images are acquired to monitor disease status. These images are typically analysed qualitatively or quantitatively without applying registration. In this study, a biophysically based model of the left lung was created using a CT image acquired in the supine position. Then, deformation of lung tissue in the upright position was computed and areas with abnormalities mapped to an X-ray image. A machine learning method was then employed to automatically differentiate between normal and abnormal areas in X-ray. The methodologies presented a tool for mapping abnormal regions between images to identify locations where abnormalities potentially change, and for multimodal and multidimensional registration.