Extraction of Pulmonary Airway in CT Scans Using Deep Fully Convolutional Networks

TL;DR

This paper presents a two-stage 3D fully convolutional network approach for automatic pulmonary airway segmentation in CT scans, improving accuracy and efficiency for medical diagnosis and surgical planning.

Contribution

It introduces a novel two-stage 3D FCN method that segments pulmonary airways at coarse and fine resolutions, enhancing segmentation performance in multi-site CT data.

Findings

Achieved a Dice Similarity Coefficient of 0.914 on validation data.

Reduced false negative and false positive errors in airway segmentation.

Validated effectiveness on multi-site, multi-domain datasets.

Abstract

Accurate, automatic and complete extraction of pulmonary airway in medical images plays an important role in analyzing thoracic CT volumes such as lung cancer detection, chronic obstructive pulmonary disease (COPD), and bronchoscopic-assisted surgery navigation. However, this task remains challenges, due to the complex tree-like structure of the airways. In this technical report, we use two-stage fully convolutional networks (FCNs) to automatically segment pulmonary airway in thoracic CT scans from multi-sites. Specifically, we firstly adopt a 3D FCN with U-shape network architecture to segment pulmonary airway in a coarse resolution in order to accelerate medical image analysis pipeline. And then another one 3D FCN is trained to segment pulmonary airway in a fine resolution. In the 2022 MICCAI Multi-site Multi-domain Airway Tree Modeling (ATM) Challenge, the reported method was evaluated on the public training set of 300 cases and independent private validation set of 50 cases. The resulting Dice Similarity Coefficient (DSC) is 0.914 $\pm$ 0.040, False Negative Error (FNE) is 0.079 $\pm$ 0.042, and False Positive Error (FPE) is 0.090 $\pm$ 0.066 on independent private validation set.