Original Article
Robust principal component analysis in optical micro-angiography
Abstract
Background: Recent development of optical micro-angiography (OMAG) utilizes principal component analysis (PCA), where linear-regression filter is employed to separate static and blood flow signals within optical coherence tomography (OCT). While PCA is relatively simple and computationally efficient, the technique is sensitive to and easily skewed by outliers. In this paper, robust PCA (RPCA) is thus introduced to tackle this issue in traditional PCA.
Methods: We first provide brief theoretical background of PCA and RPCA in the context of OMAG where coherent (complex) OCT signals are utilized to contrast blood flow. We then compare PCA and RPCA on sets of 4D-OCT complex data (3 dimensions in space and 1 dimension in time), which are collected from microfluidic phantoms and in vivo nail-fold tissue.
Results: In phantom experiments, both analyses perform relatively well since there are little motion within our observation time window, albeit small tail-noise artifacts from PCA. In nail-fold experiment, PCA suffers from tissue motion, from which RPCA does not seem to be affected. Results from RPCA also show enhancements of other dynamic signals, which are likely from the intercellular fluid. This unwanted result is yet to be proven useful for clinical applications.
Conclusions: Traditional PCA method employs linear-regression filter and is sensitive to outliers (tail-noise and motion artifacts). RPCA method is robust against outliers, but is currently computationally expensive.
Methods: We first provide brief theoretical background of PCA and RPCA in the context of OMAG where coherent (complex) OCT signals are utilized to contrast blood flow. We then compare PCA and RPCA on sets of 4D-OCT complex data (3 dimensions in space and 1 dimension in time), which are collected from microfluidic phantoms and in vivo nail-fold tissue.
Results: In phantom experiments, both analyses perform relatively well since there are little motion within our observation time window, albeit small tail-noise artifacts from PCA. In nail-fold experiment, PCA suffers from tissue motion, from which RPCA does not seem to be affected. Results from RPCA also show enhancements of other dynamic signals, which are likely from the intercellular fluid. This unwanted result is yet to be proven useful for clinical applications.
Conclusions: Traditional PCA method employs linear-regression filter and is sensitive to outliers (tail-noise and motion artifacts). RPCA method is robust against outliers, but is currently computationally expensive.
