Deep anatomy learning for lung airway and artery-vein modeling with contrast-enhanced CT synthesis

Endobronchial intervention requires detailed modeling of pulmonary anatomical substructure, such as lung airway and artery-vein maps, which are commonly extracted from non-contrast computed tomography (NCCT) independently using automatic segmentation approaches. We aim to make the first attempt to jointly train a CNN-based model for airway and artery-vein segmentation along with synthetic contrast-enhanced CT (CECT) generation.

A multi-task framework is proposed to simultaneously generate three segmentation maps and synthesize CECTs. We first design a collaborative learning model with tissue knowledge interaction for lung airway and artery-vein segmentation. Meanwhile, a conditional adversarial training strategy is applied to generate CECTs from NCCTs guided by artery maps. Additionally, CECT identity and reconstruction help to regularize the model for plausible NCCT to CECT translation.

Extensive experiments were conducted to evaluate the performance of the proposed framework based on three datasets (90 NCCTs for the airway task, 55 NCCTs for the artery-vein task and 100 CECTs for the artery task). The results demonstrate the effective improvement of our proposed method compared to other methods and configurations that can achieve more accurate segmentation maps (Dice score coefficients for these three tasks are: 93.6%, 80.7% and 82.4%, respectively) and realistic CECTs simultaneously. The ablation study further verifies the effectiveness of the components of the designed model.

This study demonstrates the model potential in multi-task learning that integrates anatomically relevant segmentation and performs NCCT to CECT translation. Such an interaction approach promotes mutually for both promising segmentation results and plausible synthesis.