Fig. 1: Data-centric, active learning, artificial intelligence-driven precise segmentation of complete murine lung airways.

a Schematic illustration of the key features of LungVis 1.0 including 3D LSFM lung imaging and the development of artificial intelligence-driven airway segmentation to resolve the spatial and temporal nanoparticle (NP) deposition profiles in the mouse lung. Illustrations created partially with Biorender.com. b Manual extraction of ground truth (MS_GT, lung annotations) from the raw, non-stained LSFM images. 2D and 3D images show the original LSFM lung structure (autofluorescence, AF in green), manually segmented GT (in blue), and merged images. c LungVis 1.0 AI pipelines overcome multiple imaging challenges for high-performance airway segmentation. Occasionally, poor image quality arises from imaging shadows, blurring effects, out-of-focus central region (poor illumination in lung center), inconsistent slice illumination, poor and variable signal-to-noise ratio, false gray structures, etc. can be observed in some of LSFM lung images. With the data-centric active learning approach and method improvement, LungVis 1.0 achieved high quality and robust segmentations even in the most challenging cases, as demonstrated for label-free AF lung images in the visible (high AF1 - default AF channel) and near-infrared channel (low AF2). d Two exemplary AI segmentations of complete bronchial trees from either a whole lung or a single lung lobe with different imaging errors (i.e., imaging shadow and blurring) are displayed. Representative data from n = 78 biological samples. e Average time investment for complete airway labeling in lungs via manual versus AI segmentation. Data are presented as mean ± SD, n = 4 biological replicates. f–g The Dice Score and centerline Dice Score were evaluated across three GT lungs in three AI iterations from the test datasets. Data are presented as mean ± SD, n = 3 biological replicates. Scale bars:1000 µm. Source data are provided as a Source Data file.