Lumen Segmentation in Endobronchial Optical Coherence Tomography with Deep Learning

Chronic lung allograft dysfunction (CLAD) is a condition that kills approximately 50% of lung transplant (LTx) patients who survive their first year post-surgery. A correlation between CLAD status and dilation of the small airways has been shown, and assessment of these airways by endobronchial optical coherence tomography (EB-OCT) could prove beneficial in identifying CLAD in its early stages.
To perform this task, a deep learning approach is presented to segment cross-sections of EB-OCT, identifying the luminal boundary. With this information, assessment of dilation is likely possible, and further correlations between airway morphology and CLAD could be investigated.
The proposed method utilizes a U-Net architecture implemented in PyTorch, designed for the segmentation of medical images. The model is trained from scratch on 532 cross-sections acquired in vivo from 44 patients (9 CLAD), utilizing cross-validation as a model-building technique and an external validation set to report metrics. Key metrics are the average Hausdorff distance (AVD), precision, and recall.
Results on external data show an AVD of 1.25 pixels, or approximately 12.5 μm. Precision and recall were 0.631 and 0.847, respectively. These values are satisfactory and reflect the challenging nature of the task.