Personalized computational mechanobiology allows predicting the progression of emphysema from CT images

抄録

Introduction: Emphysema is a progressive disease that gradually destroys alveolar tissue leading to airspace enlargement. CT imaging have demonstrated that tissue density decreases in a heterogeneous manner producing a distribution of clusters of low attenuation areas (LAA). Although computational models have been proposed to understand the process via mechanical failure, they lack patient specificity. The aim of this work was to develop a computational model that allows spatiotemporal simulations of patient specific tissue loss on CT images.

Materials and Methods: We created a large 2-dimensional hexagonal elastic network to recapitulate the alveolar geometry. Initially, the network had a rectangular boundary. The CT image of a patient was thresholded and the boundary of the lung field was approximated with a closed polygon. The coordinates of the vertices were mapped onto the network and nodes outside the polygon were set to be fixed whereas nodes inside the polygon were allowed move in subsequent optimization procedures. The line elements were linear springs with their initial length smaller than the distance between their nodes creating a prestress. The equilibrium configuration was solved and an apparent CT image was created by placing a square grid on the network and within each square, the number of springs were counted which was set to be proportional to CT density. From the network, a stress map can be computed by averaging the network forces within a pixel. To mimic emphysema progression, first a set of initial holes was created in the network by cutting springs in a random fashion. Next, the internal node positions were found that minimized the total elastic energy of the network. The forces on all springs were computed and 10 springs carrying the largest force were cut followed by another optimization. These steps were then repeated 5 times where each iteration represents the progression of emphysema.

Results and Discussion: The original CT image and the LAA cluster structure created by the model were similar and the corresponding network also predicted the LAA clusters and a stress map. The model captured the large and medium sized LAA clusters but not the smallest clusters. The model was then applied to predict a patient's CT image 1 or 2 year after the first image was taken by advancing the progression of emphysema on the images obtained at time 0. Despite some differences, the model was able to capture major trends in structural alterations. The stress map predicted by the model can then be used as a patient and location specific risk predictor.

Conclusions: We have introduced a personalized network model approach to convert a CT image to a stress map which allows predicting the spatial location of tissue deterioration. Our approach may find implications for predicting the personalized rate of decline of lung structure and function in response to interventions such as drug treatment or lung volume reduction.

Acknowledgements: This study was funded by NIH grant U01 HL-139466.