Purpose: To develop a tool for generating high resolution, realistic and random human airway trees that abide by physiological and anatomical constraints. This tool is meant to facilitate training and validation of deep learning airway segmentation algorithms for surgical planning and pulmonary disease diagnosis.
Methods: Airway tree generation begins with selecting terminal airway segments (endpoints) within the lung by randomly sampling a given oxygen demand map. The tree is progressively generated by calculating a connection cost between every existing branch within the tree and an endpoint. A new branch is then created from the endpoint to the lowest cost branch by forming a new bifurcation. The cost function is designed to optimize the physiological principle of minimum work (Murray’s law and volume-minimization) under anatomical constraints (main airway branch to each lobe). The algorithm was adjusted to assure no airway intersections. Airways were simulated down to .4mm, equivalent to the average minimum human airway diameter. Realism was evaluated by comparing length to diameter ratios (LDR), homothety ratios (HR), and branching angles with literature data for human subjects. Simulation randomness was evaluated through a dissimilarity metric (1-Dice Coefficient) measured between airway trees.
Results: For 40 synthetic airway trees, the LDR, HR, and branching angles [mean(std)] were found to be 2.9(1.3), 0.72(0.2) and 76.3(13) degrees, respectively. Corresponding literature values for human subjects were 2.9(2.0), 0.71(0.1) and 76.1(46) degrees. The airway tree randomness as characterized with the dissimilarity metric was 0.73(.01).
Conclusion: A method for the simulation of realistic airway trees with LDR, HR, and branching angles close to human subjects has been developed. This algorithm can be paired with any existing digital phantoms for realistic background and could allow the training of airway segmentation algorithms with large datasets and known ground truth.