Markov random field modeling in posteroanterior chest radiograph segmentation

Area of Science:

• Medical Imaging
• Radiology
• Computer Vision

Background:

• Previous work developed a lung region identification algorithm for digitized chest radiographs (DCRs).
• The need for automated segmentation of multiple anatomical structures in DCRs is increasing.

Purpose of the Study:

• To generalize a prior algorithm for identifying multiple anatomical regions in DCRs.
• To classify each pixel into one of six categories: lung, subdiaphragm, heart, mediastinum, body, or background.

Main Methods:

• A probabilistic approach was used, defining optimal pixel classifications (xOPT) by maximizing the conditional distribution P(x/y).
• A spatially varying Markov random field (MRF) model incorporated spatial and textural information for each region type.
• Iterated Conditional Modes (ICM) algorithm was employed to find the maximum of P(x/y) for optimal DCR segmentation.

Main Results:

• The generalized algorithm successfully classified pixels into six distinct anatomical regions.
• The algorithm achieved a high accuracy of 90.0% +/- 3.4% in pixel classification on DCRs.
• Demonstrated the flexibility of the original algorithm by extending its application to multiple structures.

Conclusions:

• The enhanced algorithm provides accurate and flexible multi-region segmentation for DCRs.
• This method has potential applications in automated radiological analysis and computer-aided diagnosis.
• The MRF model effectively integrates spatial and textural features for robust anatomical segmentation.