Advanced lung segmentation on chest HRCT: comprehensive pipeline for quantification of airways, vessels, and injury patterns.
Authors
Affiliations (5)
Affiliations (5)
- Department of Biomedical Engineering, Istituto Di Ricerche Farmacologiche Mario Negri IRCCS, Villa Camozzi Via G.B. Camozzi 3, 24020, Ranica, BG, Italy.
- Dipartimento Di Elettronica, Informazione E Bioingegneria, Politecnico Di Milano, Via Ponzio 34/5, 20133, Milan, MI, Italy.
- Department of Radiology, ASST Papa Giovanni XXIII, Piazza OMS 1, 24127, Bergamo, BG, Italy.
- Respiratory Unit, ASST Papa Giovanni XXIII, Piazza OMS 1, 24127, Bergamo, BG, Italy.
- Department of Biomedical Engineering, Istituto Di Ricerche Farmacologiche Mario Negri IRCCS, Villa Camozzi Via G.B. Camozzi 3, 24020, Ranica, BG, Italy. [email protected].
Abstract
Chest high-resolution computed tomography (HRCT) is crucial for diagnosing and monitoring pulmonary diseases involving parenchymal, vascular, and airway alterations. However, segmentation faces challenges in distinguishing pulmonary structures due to heterogeneity in image acquisition and pathological manifestations. Unlike existing tools, which usually target a single anatomical structure and rely predominantly on either deep learning or rule-based approaches, our hybrid pipeline pairs U-Net-based AI segmentation with tailored image processing refinements to produce a reliable and simultaneous segmentation of lungs, airways, pulmonary vessels, and parenchymal injury patterns, while enabling quantitative characterization across a spectrum of disease severities and types (inflammatory and infectious). This retrospective observational study employed 19 chest CT scans from COVID-19 public datasets for deep learning, 8 annotated scans from the EXACT'09 challenge to validate airway segmentation, and 20 retrospective HRCT scans from COVID-19 and idiopathic pulmonary fibrosis patients for pipeline validation. The pipeline performs preliminary segmentation of lungs, airways, and pathological regions using U-Nets, followed by image processing to refine results, include vasculature, and classify injury patterns in ground-glass opacities, reticulations/consolidations, and air-filled pathological spaces. Three radiologists validated segmentations on a 1-5 scale, and the Kruskal-Wallis test was conducted to assess differences across raters, pathologies, and severities. The proposed pipeline visually outperformed established tools (LungCTAnalyzer, PTK, TotalSegmentator). Airway's segmentation achieved a Dice coefficient of 0.91 [0.89-0.92] on the EXACT'09 dataset. Radiologists assigned scores of 4 and 5 to segmentation completeness and accuracy, respectively, for both airways and vessels. Parenchymal injury patterns scored 4 for completeness, accuracy, and classification. Ratings were consistently high with no significant differences among raters, diseases, and severity levels. The proposed pipeline introduces a novel, comprehensive, and hybrid approach for simultaneous, multi-structure lung segmentation, demonstrating reliable and potentially generalizable performance across inflammatory and infectious pulmonary diseases.