Purpose: Deformable image registration (DIR) is a necessary step in radiotherapy treatment planning and daily patient alignment during fractionated treatment. Current registration methods do not enforce boundary conditions on the lobes of the lung, allowing for the unrealistic deformation of voxels originating in one lobe to another. The task of lobe segmentation is typically a tedious manual process for medical professionals. The implementation of a fully automated artificial intelligence approach drastically expedites this process and decreases both inter- and intra-observer variability.
Methods: In this study, we propose an accurate and efficient delineation of the lung lobes using a machine learning method. A GPU-based lobe segmentation model was constructed consisting of five generative adversarial networks (GANs), one for each lobe. 10 lung datasets were manually segmented using the program Pulmonary Toolkit and each GAN was trained on a dataset of 5 patients, totaling 2,560 images. Provided a thoracic CT as input, the model returns a binary mask volume for each lobe. The GAN-generated lobe segmentation was then compared with the ground truth segmentations. Using the masks, we performed a deformable image registration of the 4DCT lung volumes to quantify the DIR improvements enabled by the automated lobe segmentation.
Results: Our results showed that using the adversarial learning approach, an accuracy of >93% was observed in segmenting the lung lobes. The GAN was able to generate masks associated with each of the lobes with which the registration process was automated. The voxels along the lobe fissure were captured with > 97% accuracy, which resulted in an improved DIR results.
Conclusion: These results suggest the feasibility of implementing a fully automated lobe segmentation workflow. This machine learning segmentation method could be a useful tool in improving the physiological accuracy of image registration, while eliminating the time spent manually segmenting the lobes.
Funding Support, Disclosures, and Conflict of Interest: This research was funded by National Institute of Health grant R56-HL139767.