On March 25-27, the project V|LF-Spiro3D and particularly Université Paris-Saclay were nicely represented by a group of PhD students who showed off some preliminary results of their research at the Artificial Intelligence in Biomedical Imaging Congress (IABM 2024) .
Spirometry is the reference test for assessing ventilatory function using flow-volume curves measured at the mouth. Regional measurements can be obtained via dynamic pulmonary MRI. Segmentation of lobes in each respiratory phase allows estimation of volumes and flows, characterizing ventilatory pathologies. Currently, no automated algorithm exists for segmenting lung lobes from MRI images.
A semi-automatic method based on vascular and bronchial tree partitioning is time-consuming and contrast-sensitive. To enhance efficiency and robustness, we develop a multiclass automatic segmentation method using the nnUNet framework: 5 encoding and 4 decoding steps, each comprising 2 convolutions. A database was constructed from images of 25 healthy volunteers in various positions and respiratory phases. A reference volume was segmented by partitioning, and other segmentations were inferred via elastic registration, yielding 640 training and 168 test volumes. We achieve mean Dice scores ranging from 0.86 (right middle lobe) to 0.95 (right lower lobe). Global flow-volume curves are consistent with literature, and lobalar tidal volumes are provided for the first time: between 30±6 mL (right middle lobe) and 137±23 mL (right lower lobe).
Read the abstract of the presentation here!
