BMVA 2022 meeting - Unsupervised airway measurement to predict survival in bronchiectasis

BMVA April 2022 Symposium - Abstract

Download Poster

Abnormal airway dilation, termed bronchiectasis, is a feature of many chronic airway diseases. It is identified qualitatively on computed tomography (CT) in the clinic; however, quantitative analysis would make measurements of disease severity more precise and objective. The Full Width Half Maximum (FWHM) method typically used to measure airways often over estimates airway lumen diameter in small airways, making it difficult to measure bronchiectasis on CT. Airway diameters are often close to the limit of resolution and significantly vary in appearance between datasets. We consider an unsupervised deep learning method that couples a Generative Adversarial Network (GAN) with an ellipse fitting Convolutional Neural Network (CNN), similar to [1]. The GAN generates airway labels, mimicking real data to train the CNN to measure real airways. In this pilot study, the two methods are compared on N=108 bronchiectasis patients using survival models that consider comorbidities to predict survival using bronchiectasis metrics. It was found that the GAN-CNN method better predicts survival (p<0.02) compared to the FWHM method (p<0.36) in airways within the peripheral lung, which are typically smaller. This provides evidence for GAN-CNN as a viable method for generating bronchiectasis imaging biomarkers for survival prediction in small airways.

[1] Nardelli, Pietro et al. “Generative-based airway and vessel morphology quantification on chest CT images”. Medical Image Analysis 63 (2020)