Over the years, I've accumulated a wealth of medical knowledge across a very large number of specialties (the boundaries of which I largely ignore). More recently, I've started synthesizing new knowledge in this domain and contributing it back to the literature. But this still falls short of where I would like to be in medicine: that is, devising new and innovative treatments for cancer and autoimmune diseases. I have a large number of ideas to do exactly this when I have the chance. My lack of a formal medical degree has hindered me somewhat from doing this, and it is somewhat grating, as lost time equates to lost lives here. But one way or another, through increasingly ambitious research, commercial medical developments which I bring to market (first stop: radiology), and additional informal medical training, I will find an "in" into this field or I will make one.
I defended my dissertation, entitled "Mining Complex High-Order Datasets", on April 23, 2010, and graduated with my Ph. D. in Computer and Information Sciences (GPA: 3.92) from Temple University on May 13, 2010.
My dissertation work is based primarily on higher-order analogues of singular value decomposition (high-order SVD, PARAFAC, and Tucker) coupled with a variety of wavelet-based techniques and involved creation of a comprehensive data mining framework for classification, co-clustering, concept discovery (e.g. detecting handedness in subjects based on fMRI scans), summarization, and compression of high-order spatiotemporal datasets, with principal applications in fMRI. In the course of this work I also devised a tensor-theoretic multidimensional discrete wavelet transform, extended the WaveCluster algorithm to accept real-valued high-order image data, created a novel clustering algorithm based on WaveCluster called TWaveCluster, mitigated the massive partial volume effect inherent in grid-based segmentation (by deforming the grid in a context-sensitive manner prior to clustering), released the first public implementation of WaveCluster (which I have opensourced under the GPLv3), and discovered quite a few interesting new results about motor and cognitive task activation patterns in the human brain.
For instance, I discovered that performing difficult cognitive tasks lights up the same areas of the brain that process physical pain, added evidence supporting a hypothesis that ipsilateral activation is important in motor task based learning (most activation is contralateral), and localized the spatiotemporal patterns of activation found in working memory tasks.
My next task is to fuse functional and structural data in a tensor model and to begin reading people's minds using the high-tech equivalent of phrenology :)
Though not in my dissertation, I also created texture-based classifiers for diagnosing breast and brain cancer and created a galactographic ductal topology classifier for many types of breast pathology.
Despite this, my goal is to move away from diagnosis and into developing treatments. I have far more medical knowledge than my formal background suggests, but this is a difficult industry to convince of any new innovation.
Date: May 13, 2010.
I received a Master's Degree in Computer and Information Sciences from Temple University (as a continuing Ph. D. student) on August 31, 2007, one year and three days after enrolling. My GPA was 3.89. My master's project is the Medical Image Data Mining System, a content-based image retrieval system for MRIs and other medical scans.
MIDMS was published in ISBI 2008, in the article "A Web-Accessible Framework for the Automated Storage and Texture Analysis of Biomedical Images".
Date: August 31, 2007.
"Purpose. Humanity. Success." It's what I live by.
By sheer data mining, determination, and force of will, created a product which can automatically diagnose breast cancer on a mammogram with above 90% accuracy (on a dataset of 2,000 film mammograms, no less - FFDM should be even more accurate), coded it up, wrote all of the copy, designed and coded the website, incorporated a business around it, and brought it to beta by October 2011. Wrote up and filed a provisional patent, published a journal paper, and started on a mobile application. Responsible for all facets of system administration and maintenance, and just about all things technical.
Published "Mammographic Segmentation Using WaveCluster" in the 2012 medical informatics special edition of "Algorithms".
Date: October 4, 2011.
This is a paper that I began writing on August 6, 2007, and completed on August 7, 2007. It deals with classification of tissue type in individually extracted regions of interest from MRI image datasets. We achieve 89% classification accuracy between 11 tissue types using our texture-based approach in a spontaneous murine model.
Date: August 7, 2007.
I designed a "smart" machine learning-enabled knee brace with an embedded accelerometer (pictured) that can automatically distinguish between three different types of Parkinson's Disease symptoms (tremor, freezing, dyskinesia, and normal) with 80% sensitivity and near-100% specificity. This level of diagnostic accuracy allows the device to be worn continually with minimal disruption to the patient's lifestyle.
I partnered with a company called BioMotion Suite to bring this device to market. Just as I ran into medical regulations and an impossible burden of proof (multi-institution studies all over the globe) when trying to directly market my mammographic CAD system for Living Discoveries, so too did BioMotion Suite hit a circular dependency between commercializing this technology and gaining funding to commercialize this technology.
As a result, this is likely my last foray into the medical device space, either independently or through an intermediary. It's too painful creating lifesaving devices and watching their impact zeroed by medical gatekeepers.
My next entrepreneurial activity was Clipless, a deals app. Without the gatekeepers of medicine, it became an overnight success and was acquired within 8 months of its founding.
I intended it at least partly as commentary on the misplaced priorities of our society.
Date: May 27, 2012.
The Medical Image Data Mining System is a framework for automated medical (primarily brain) image dataset submission and analysis which I wrote as a Master's Project (with the help of Jingjing Zhang, another graduate student) over the course of just one week. The system is written in Perl and Matlab and employs a methodology that achieved 89% classification accuracy on MRI images of the brain in another study (which we submitted in a separate paper to the same conference). Thus the system has diagnostic accuracy sufficient for clinical use in applications such as cancer diagnosis.
The associated paper was presented at ISBI 2008 and published in proceedings.
Date: August 1, 2007.
This is a journal paper published in IEEE Transactions on Medical Imaging, Vol. 28, Issue 4, pp. 487-493. It significantly extends our branching pattern analysis approach and is conducted on a much larger dataset than our previous studies in this area.
Date: November 4, 2007.
During my first semester at Temple, I submitted research work by myself and the research group I work with (DEnLab) to ISBI 2007. I am the first author on this paper because I performed much of the research, ran the experiments, and wrote the paper (the first paper I wrote from scratch).
The paper was rejected by ISBI, likely because our results were initially poor.
However, I revisited the paper over the summer of 2007 and obtained much better results. We submitted the revised paper to IWDM 2008, where it was accepted for publication.
Date: December 15, 2007.
I spoke at a regional New Jersey Coast ACM meeting, giving viewers an overview of what the AI landscape looked like (vs. popular concepts of AI), then discussed medical applications of AI and gave a rare sneak peek into a few of the methods that would form the basis of the classifier used by Living Discoveries - knowledge that doesn't exist anywhere else.
Date: November 18, 2011.
This paper, accepted to PAKDD 2010, was the first semi-complete record of the techniques which would eventually form my "TWave" framework. The key result in this paper was a classification technique which preserved the accuracy of wavelet-based classification by subject and task in an fMRI dataset, while executing two orders of magnitude faster (2 hours vs. 8 days!)
http://books.google.com/books?id=aEmO_PH5_UsC&lpg=PA246&ots=ZdQO-NPEMP&dq=high-order%20analysis%20of%20spatiotemporal%20data&pg=PA246#v=onepage&q&f=false
Date: June 21, 2010.
This paper introduces a high-order data summarization technique I developed using tensor decompositions which is capable of identifying the handedness of subjects based solely on their fMRI scans. I developed the method as part of my dissertation work, and it eventually formed a major component of the dissertation and defense.
http://ieeexplore.ieee.org/xpl/freeabs_all.jsp?arnumber=5490087
Date: April 10, 2010.
This is a conference paper published at ISBI 2009 which augments our tree-like structure analysis by automatically discovering branching points. The first author on this paper is Haibin Ling; I am the second. This is fitting; my role in this paper was mostly supportive, although I did provide a slight amount of aid with the experimental design.
Date: January 15, 2009.
This paper was accepted to CARS 2008 in Barcelona, Spain, and will be presented in a poster session on June 25, 2008.
It is on using wavelets to exploit the spatiotemporal locality of 4D fMRI motor task data. The absolute accuracy of the classifier is reasonable at 81%, but it is a 30% improvement over voxelwise approaches. This was also my first exposure to wavelets, and as such, I chose the Haar wavelet for the sake of simplicity. I could likely improve the accuracy even more with a different wavelet function.
Because this was my first exposure to wavelets, I really hope I don't have any mistakes in the background section of this paper. The reviewers did not identify any, but I've found peer review to be next to useless in performing its ostensible function, instead choosing to enforce petty guidelines of presentation.
Date: February 14, 2008.
My group presented its research at the 2006 Society for Neuroscience Meeting.
Our presentation encompassed six different research topics: application of our novel classifier, known as Dynamic recursive partitioning (DRP) on an Alzheimer's disease fMRI dataset and a study of sexual dimorphism in the human corpus callosum (two projects), an image linearization procedure using space filling curves, a feature extraction technique using concentric spheres, use of this feature extraction technique to segment and classify brain tumors (with a very high degree of accuracy; this really should be employed clinically to assist diagnosis), and an image similarity search utility using wavelets and vector quantization.
I was responsible for writing the actual demonstration and GUI code. Additionally, I was responsible for the two DRP projects and, to a lesser extent, the similarity search project.
Working on these projects is how I learned Matlab. My knowledge probably still has some holes, but I learned the language incredibly rapidly - I was able to program proficiently in it after about one week. The only languages I learned faster than this were HTML (but that hardly counts) and Perl.
Date: October 16, 2006.
A followup to the prior paper on branching node detection, with an improved technique.
Probabilistic Branching Node Detection Using AdaBoost and Hybrid Local Features: By Tatyana Nuzhnaya, et al. (2nd author). Published in Proceedings of ISBI 2010.
Date: April 10, 2010.
Spatial Feature Extraction Techniques for the Analysis of Ductal Tree Structures: By Aggeliki Skoura, Michael Barnathan, and Vasileios Megalooikonomou. Published in Proceedings of EMBC 2009, Minneapolis, Minnesota, September 2 - 6, 2009.
Date: September 2, 2009.
Another paper on galactogram classification using topology, accepted to ISBI 2009. I provided "the competing method" in this paper. Angeliki Skoura is the first author; I am second.
Date: January 15, 2009.