Logo image
Back
MEDPSeg: an offline pulmonary segmentation and quantification tool for ground‑glass opacities, consolidation, and other pulmonary structures on computed tomography
Conference proceeding

MEDPSeg: an offline pulmonary segmentation and quantification tool for ground‑glass opacities, consolidation, and other pulmonary structures on computed tomography

Diedre S. Carmo, Jean A. Ribeiro, Alejandro P. Comellas, Joseph M. Reinhardt, Sarah E. Gerard, Letícia Rittner and Roberto A. Lotufo
Vol.13926, pp.139260A-139260A-11
04/02/2026
DOI: 10.1117/12.3087511

View Online

Abstract

A tool for automated pulmonary characterization using chest computed tomography (CT) would need to understand both healthy structures and abnormal findings, requiring the joint segmentation of healthy parenchyma, lesion involvement, and neighboring anatomical structures. Yet existing open-source tools focus on only part of this problem, or require setting up complex interactions between multiple heavyweight models. We introduce MEDPSeg, a lightweight method that leverages novel hierarchical polymorphic multitask learning strategies to learn from heterogeneous labels and deliver six segmentation masks: lung; lung lesion; ground-glass opacity (GGO), consolidation, airway tree, and pulmonary artery. All in a single forward pass. Its development took into consideration future distribution in lightweight and offline environments, allowing for execution in low resource infrastructure without needing internet access or the sharing of sensitive data. MEDPSeg’s Python package (https://github.com/MICLab-Unicamp/medpseg) offers both a graphical user interface (GUI) and a command-line interface (CLI). We also host an online demo (https://medpseg.neuralmind.ai). Trained on 6000 CT volumes aggregated from COVID-19, Cancer, COPD, airway tree, pulmonary vessel and lung datasets, MEDPSeg attains better or comparable results to the state-of-the-art in multiple tasks. It outperforms or matches recent specialized methods in the segmentation of lung, lung opacity, airway, pulmonary vessel, GGO and consolidation. Prediction for a high resolution scan takes less than 1 min on a budget GPU, using less than 8 GB of VRAM to generate comprehensive quantification reports and multiple segmentation masks. The open-source tool has been validated by other internal and external independent studies. The MEDPSeg tool delivers comprehensive pulmonary assessment with commodity hardware, bridging the gap between research prototypes and practical understanding of thoracic CT imaging.

Details

Metrics

Logo image