Emphysema classification based on embedded probabilistic PCA.

In this article we investigate the suitability of a manifold learning technique to classify different types of emphysema based on embedded Probabilistic PCA (PPCA). Our approach finds the most discriminant linear space for each emphysema pattern against the remaining patterns where lung CT image patches can be embedded. In this embedded space, we train a PPCA model for each pattern. The main novelty of our technique is that it is possible to compute the class membership posterior probability for each emphysema pattern rather than a hard assignment as it is typically done by other approaches. We tested our algorithm with six emphysema patterns using a data set of 1337 CT training patches. Using a 10-fold cross validation experiment, an average recall rate of 69% is achieved when the posterior probability is greater than 75%. A quantitative comparison with a texture-based approach based on Local Binary Patterns and with an approach based on local intensity distributions shows that our method is competitive. The analysis of full lungs using our approach shows a good visual agreement with the underlying emphysema types and a smooth spatial relation.