Room: AAPM ePoster Library
Purpose: Lung lobe segmentation is currently done in a semi-automatic manner, which is infeasible for processes that require segmentation of multiple CT scans. In this abstract, we aim to automatically segment the lung lobes in Fast Helical Free Breathing CT (FHFBCT) scans.
Methods: We developed a machine learning based framework for automated segmentation of lung lobes from FHFBCT scans. A set of 10 patient datasets (25 3D FHFBCT scans each) were retrospectively employed for this study. 150 FHFBCT scans were employed for training, in which the lobes were first automatically segmented using Hessian matrix and fissure identification based algorithms (conventional method). Motion blurring caused the segmentations to exhibit errors, which were manually corrected using the Pulmonary Toolkit interface. The final segmentations were used for training an adversarial neural network. The neural network consisted of generator and discriminator deep neural networks. The generator network generated the lung lobe segmentations while the discriminator ensured lobe segmentation accuracy. At the end of the training, the generator network segmentation was tested using the remaining 100 FHFBCT scans, and the segmentations were evaluated using image a normalized cross correlation (NCC) metric.
Results: The generator networks accurately segmented the lobes. The generator segmentation NCC was >0.9, while the conventional automated segmentation NCC was <0.7. The improved NCC indicated that the proposed segmentation method was superior to the conventional approach.
Conclusion: Using an automated machine learning network, we demonstrated the feasibility to segment lung lobes from FHFBCT scans in near real-time. This provides a tool for automating a key component of pulmonary research and technique development for applications using FHFBCT.