Unsupervised X-ray image segmentation with task driven generative adversarial networks.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Semantic parsing of anatomical structures in X-ray images is crucial for clinical applications.
• Deep learning methods require extensive labeled data, which is difficult to obtain for X-rays due to complex anatomy and texture.
• Labeled 3D CT data is more accessible for organ delineation.

Purpose of the Study:

• To develop a model for automatic X-ray image parsing using labeled 3D CT scans.
• To overcome the challenge of limited annotated X-ray data.
• To achieve accurate multi-organ segmentation in X-ray images without X-ray domain annotations.

Main Methods:

• A Deep Image-to-Image network (DI2I) was trained on Digitally Reconstructed Radiographs (DRRs) from 3D CT volumes.
• A Task Driven Generative Adversarial Network (TD-GAN) was developed for simultaneous synthesis and parsing of real X-ray images.
• The entire pipeline was designed to avoid the need for X-ray image annotations.

Main Results:

• The proposed model achieved an average Dice score of 86% on topograms, comparable to supervised methods (89%).
• The model successfully segmented anatomical structures without requiring any topogram labels.
• The TD-GAN demonstrated generality on public datasets.

Conclusions:

• The proposed framework effectively learns X-ray image parsing from labeled 3D CT data.
• This approach addresses the challenge of limited annotated X-ray data for semantic parsing.
• The method shows promise for accurate and efficient anatomical segmentation in clinical settings.