JoHof/lungmask

Automated lung segmentation in CT

What it solves

Reading CT scans of the lungs is harder when the patient has severe conditions like tumors, fluid buildup (effusions), or heavy scarring (fibrosis). Standard tools often fail to outline the lungs cleanly in these cases. This project provides trained U-net models that automatically segment (outline) the lungs in CT images, even when dense pathologies are present. It also offers models that separate the lungs into left and right, or into individual lobes, and can fuse results to combine strengths.

How it works

The package includes four pre-trained U-net models:

  • U-net(R231) – The default model. It works on individual CT slices, extracts the right and left lungs separately, and includes air pockets, tumors, and effusions. It does not include the trachea. Trained on a diverse dataset.
  • U-net(LTRCLobes) – Trained on a subset of the LTRC dataset. It separates the lungs into five lobes, but performs poorly when dense pathologies are present or when the fissures between lobes aren’t visible on every slice.
  • U-net(LTRCLobes_R231) – Runs both the R231 and LTRCLobes models and fuses the results. False negatives from LTRCLobes are filled by R231 predictions and mapped to a neighboring lobe label; false positives are removed. This fusion is computationally heavy and can take several minutes per volume.
  • U-net(R231CovidWeb) – A variant of R231 that was additionally trained on COVID-19 CT slices that were mapped back from regular image formats to Hounsfield units (HU). Useful for cropped, annotated, or compressed web images.

All models are trained U-nets, a type of neural network for image segmentation. The package can run as a command-line tool or as a Python module. It uses SimpleITK for image I/O (or numpy arrays directly, as of version 0.2.9). GPU support speeds up processing significantly (seconds vs. minutes).

Who it’s for

  • Researchers and clinicians working with lung CT data, especially those dealing with severe pathologies or COVID-19 cases.
  • Developers who need automated lung segmentation in their Python pipelines or as a CLI tool.
  • Anyone who wants to distinguish left/right lungs or individual lobes in CT volumes.

Highlights

  • Four ready-to-use models tailored to different needs (default, lobes, fusion, COVID-adjusted).
  • Handles dense pathologies that confuse other methods.
  • Simple install via pip or conda, and easy Python API (LMInferer).
  • Can process DICOM directories automatically, finding the largest volume.
  • Supports both GPU (fast) and CPU (slow) inference.
  • Includes a model-fusion option to combine lobe and whole-lung segmentations.

Related

  • Project
    Xiaoqi-Zhao-DLUT/MSNet-M2SNet

    A medical image segmentation framework featuring MSNet and M2SNet architectures that use multi-scale subtraction to accurately isolate pathologies like polyps and lung infections.

  • Project
    wasserth/TotalSegmentator

    TotalSegmentator is an open-source deep-learning tool that automatically segments dozens of anatomical structures (organs, bones, vessels, muscles, etc.) in CT and MR images, outputting 3D masks and statistics. It saves radiologists and researchers hours of manual annotation and supports many task-specific models for different body regions.

  • Project
    zsylvester/segmenteverygrain

    segmenteverygrain is a Python package that automatically detects grains in images using a U-Net to create first-pass masks and SAM 2.1 to refine high-quality outlines. It includes interactive tools for cleaning up results and fine-tuning the model for new image types, aimed at geologists and researchers measuring grain size and shape.

  • Project
    MrGiovanni/UNetPlusPlus

    A nested U-Net architecture for medical image segmentation that improves upon the original U-Net by redesigning skip connections and allowing for flexible network depth.

  • Project
    HUANGLIZI/LViT

    LViT is a multimodal framework that combines language and vision transformers to improve the accuracy of medical image segmentation in CT scans and other medical imagery.