Back

2D versus 3D artificial intelligence-driven segmentations of airway alterations in cystic fibrosis: which one is better?

Hadj Bouzid, I.; Benlala, i.; Dournes, G.

2023-11-20 radiology and imaging
10.1101/2023.11.18.23298712 medRxiv
Show abstract

Purpose or Learning ObjectiveArtificial intelligence (AI) with convolutional neural network allows fully automated detection and segmentation of bronchial changes on CT-scans of cystic fibrosis (CF). However, the superiority of two-dimensional (2D) versus three-dimensional (3D) architectures remains to be explored. Method or BackgroundCT-scans from fifty CF patients were retrospectively included at two CF reference centers. The nnUnet model was implemented in both 2D and 3D, and trained to segment five structural alterations: bronchiectasis, wall thickening, mucus plugs, bronchiolar impactions and consolidations. A semantic validation was done by using fifty CTs with a five-fold cross validation strategy, by comparing normalized Dice-Sorensen coefficient (DSC) between 2D and 3D architectures, with manual segmentations as Gold Standard. Results or FindingsThe 3D nnUnet was found able to segment the five CF main hallmarks such as bronchiectasis, wall thickening, mucus plugs, bronchiolar impactions and consolidations. Metrics obtained with the 3D architecture were superior for mucus plugs, bronchiolar impactions and consolidations (p<0.001) but not significantly different for bronchiectasis and wall thickening (p>0.05). ConclusionAI with the 3D-nnUnet model can perform fully automated segmentation of CF-related structural hallmarks on CT scans, and overcome 2D implementation. Non-invasive, holistic 3D quantifications are allowed for promising next clinical applications.

Matching journals

The top 4 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.