JoHof/lungmask

What it solves

It provides automated segmentation of lungs in CT scans, specifically designed to handle severe pathologies (such as tumors, effusions, and severe fibrosis) that often cause standard segmentation tools to fail.

How it works

The project uses trained U-net models to process CT images. It offers four distinct models depending on the goal:

• U-net(R231): A general-purpose model that separates left and right lungs and includes dense pathological areas.
• U-net(LTRCLobes): A specialized model that segments individual lung lobes.
• U-net(LTRCLobes_R231): A hybrid approach that fuses the lobe-specific model with the general model to fill in missing areas (false negatives).
• U-net(R231CovidWeb): A version augmented with COVID-19 data to better handle cropped or non-standard images from the web.

Who it’s for

Medical researchers and clinicians working with lung CT data who need reliable, automated masks for lung volumes or specific lobes.

Highlights

• Pathology Robustness: Capable of including tumors and effusions in the segmentation.
• Flexible Input: Supports DICOM series, ITK formats, and Numpy arrays.
• GPU Acceleration: Supports CUDA for significantly faster inference (seconds vs. minutes per volume).
• Multiple Modalities: Provides both simple left/right lung separation and detailed five-lobe segmentation.