tl;dr RKHS make subsequent optimization learning easier to implement and more likely to generalize better [primer].
tl;dr RKHS make subsequent optimization learning easier to implement and more likely to generalize better [primer].
\subsection{Optical Coherence Tomography of the Retina}
\subsection{Optical Coherence Tomography of the Retina}
[Ronald?] -- ~ 4 paragraphs summarize impact of retinal OCT (Glaucoma, occulomics, etc.), background on VFMD, background on metadata, imaging physics.
[Ronald?] -- ~ 4 paragraphs summarize impact of retinal OCT (Glaucoma, occulomics, etc.), background on VFMD, background on metadata, imaging physics.
Key recent deep learning papers [Schuman et al.]. Other learning approaches [NNMF et al.].
Key recent deep learning papers [Schuman et al.]. Other learning approaches [NNMF et al.].
...
@@ -101,16 +102,20 @@ Optical coherence tomography (OCT) is a non-invasive diagnostic imaging tool whi
...
@@ -101,16 +102,20 @@ Optical coherence tomography (OCT) is a non-invasive diagnostic imaging tool whi
Key observation -- inherent anisotropy of imaging implements Frangi-like plate filter along the high resolution axis.
Key observation -- inherent anisotropy of imaging implements Frangi-like plate filter along the high resolution axis.
\subsection{Kolmogorov complexity and the normalized information distance (NID)}
\subsection{Kolmogorov complexity and the normalized information distance (NID)}
FLIF \cite{FLIF}\dots
FLIF \cite{FLIF}\dots
\subsection{Semi-supervised spectral learning}
\subsection{multi-scale image structure enhancement and positive semi-definite functions}
Frangi's 3-D vesselness filter was a big advance in our ability to extract biological structure from low-SNR images.
\subsection{Kolmogorov structure function vs. }
$h_{Kolmogorov}$ [Vereschegin and Vitanyi] \emph{vs.}$h_{Frangi}$
\subsection{Quantization}
\subsection{Semi-supervised spectral learning}
\subsection{Kolmogorov structure function vs. multi-scale vessel ehancing image filters}
$H_{Kolmogorov}$ [Vereschegin and Vitanyi] \emph{vs.}$H_{Frangi}$
\section{The retinal structure function}
\input{algorithmicStructure.tex}
\section{Validating the NCD vs. prediction against functional changes of the visual field }
\section{Validating the NCD vs. prediction against functional changes of the visual field }
...
@@ -132,14 +137,18 @@ RKHS combinations: macula + ONH simultaneously, multi-resolution by channel vs.
...
@@ -132,14 +137,18 @@ RKHS combinations: macula + ONH simultaneously, multi-resolution by channel vs.
\section{Acknowledgements}
\section{Acknowledgements}
\section{Hardware, software and data availability}
A sample dataset of labeled OCT imaging data is being made available concurrent with this manuscript. The images and metadata can be explored using the LEVERSC [ref] web visualization toolkit: https://todo, or direct download at ???
\section{Software and data availability}
All of the software tools used are available free and open source, see \url{https://git-bioimage.coe.drexel.edu/opensource/rsf}.
All of the software tools used are available free and open source, see \url{https://git-bioimage.coe.drexel.edu/opensource/rsf}.
The image data together with segmentation and tracking results can be viewed interactively at \url{https://leverjs.net/ssfCluster}. The LEVERSC 4-D WEBGL viewer \cite[]{Cohen2022} renders 3-D kymographs and images, and the web API also supports downloading metadata and results directly.
The image data together with segmentation and tracking results can be viewed interactively at \url{https://leverjs.net/ssfCluster}. The LEVERSC 4-D WEBGL viewer \cite[]{Cohen2022} renders 3-D kymographs and images, and the web API also supports downloading metadata and results directly.
The image processing pipeline described here requires dedicated hardware. Structure filters are GPU intensive, FLIF compression is CPU intensive. The reference hardware platform is 128 AMD X-???, 3 NVIDIA A6000 GPU, 2TB RAM, 1PB HDD.