• Light
  • Dark
  • Auto
Select Page

An important challenge in the use of artificial intelligence (AI) for medical image segmentation tasks is the lack of high-quality, scan protocol-specific datasets. AI performs best on narrow tasks with homogenous specifications. Thus, pre-trained models may be inadequate for use in centre-specific studies if the scan protocols do not match.

For the airway segmentation task in the Imaging in Lifelines study (ImaLife), we started with an open-source 3D-Unet deep learning model (Bronchinet) to segment airways. As Bronchinet was trained on scans with higher dose and less noise, the resulting airway segmentations for ImaLife images were incomplete and contained some disconnected components. To improve Bronchinet performance in the ImaLife study, we required a high-quality ground-truth airway segmentation dataset. Due to manual segmentations of airways taking a significant amount of time (up to 15 hours on average), we developed a manual correction approach using open-source software; this reduced the time to create a high-quality ground-truth airway segmentation to an average of 2 hours per scan.

Using the new ground-truth dataset, we re-trained Bronchinet and improved its performance on ImaLife images. The outlined method can be applied to other segmentation tasks where openly available trained models can be used to obtain an initial segmentation and then corrected to a high-quality standard. Overall, our paper shows that it is possible to create ground-truth segmentations without a drawn-out time investment for AI training.

Key points

  • Artificial intelligence (AI) segmentation tools require high-quality training data matching the population and scanning parameters of the use case.
  • Manually correcting initial airway segmentations based on free tools is an efficient way to create an optimal dataset for AI training purposes.
  • Performance of an existing AI model trended towards more complete airways following retraining with corrected data.

Article: Creating a training set for artificial intelligence from initial segmentations of airways

Authors: Ivan Dudurych, Antonio Garcia-Uceda, Zaigham Saghir, Harm A. W. M. Tiddens, Rozemarijn Vliegenthart & Marleen de Bruijne

Latest posts