1 Problem: Anamoly detection in patients is challenging because reference values are vague

Identifying patients with enlarged livers is important for identifying hepatitis, systemic disease, and even chronic conditions like alcohol abuse. Oversized is, however, a very poorly defined term. What is normal? Currently medicine operates on the basis of painstaking, time-consuming studies made from relatively homogenous populations. Using this ruler means many sick patients are not identified.

2 Solution: Precision Medicine with IQAE

By combining large sets of information about patients together, finding outliers and treating patients can be much better matched to the individual background of that patients. Rather than comparing the patient to an old publication, they can be compared to a group matching their age, gender, and lifestyle.

The graph below shows the same patient with (left) and without (right) precision medicine.

Despite recent boosts in popularity, both personalized and precision medicine are still not part of the standard diagnosis process for many diseases.

2.1 Image Query and Analysis Engine

Liver Image

Standard imaging modalities like thorax CT and MRI are commonly used to diagnose a number of different diseases. These images are normally examined once for a single specific diagnosis and then archived. Hospitals are required to archive a majority of patient imaging studies and scans for a minimum of 10 years. These essential data are just sitting unused on PACS, tapes, and other storage volumes.

Unlock your data’s potential

These archives of images contain exactly the sort of information required to make precision diagnoses, they are simply difficult to access.

IQAE is the key to unlocking the potential of this data and turning it from a burden into an asset. The data can be analyzed in a completely anonymized manner to obtain general, precise statistics as references against which to compare new patients. Additionally careful analysis of all data can reveal mistakes, missed diagnoses, and new potential high risk patients.

2.2 Real-time Querying of Images

A hospital database with thousands of patient images can be quickly screening for the right type of images

SELECT Image as ChestCT FROM PatientImages WHERE Modality="CT" AND Region="Chest"
Tracks

\[\rightarrow \textrm{Perform Segmentation}\]

SELECT CHEST_SEGMENTATION(Image) as ChestSeg FROM ChestCT
Segment \(\underbrace{\rightarrow}_{\textrm{To individual organs}}\) Final Forms

\[ \downarrow \textrm{ Extract the bones} \]

SELECT BoneImage FROM ChestSeq
Final Forms

\[ \downarrow \textrm{ Extract Meaningful Information} \]

SELECT pt.Age,EstimateBoneMineralDensity(cs.BoneImage) as BMD
  FROM PatientDatabase AS pt JOIN ChestSeq AS cs ON pt.id == cs.PatientId

2.3 What else?

Our Image Query and Analysis Engine runs in the cloud or locally and is used as Software as a Service to ensure you are always working with the latest generation of algorithms and tools. Once the image data is inside the possibilities are endless.

SELECT * FROM ChestSeq GROUP BY Tissue.Type

Organ Subdivisions Size Density Granularity Contrast
Organ-1 330 19716 167.005424 5.236076 79.237187
Organ-2 304 24070 114.700884 15.393268 114.506345
Organ-3 1527 34052 108.518688 2.199077 38.206649
Organ-4 923 45009 167.644877 5.812359 93.827160
Organ-5 2241 420616 4.335617 3.901147 1.638404
Organ-6 1249 111752 87.660901 5.682593 36.652440
Organ-7 1810 42470 92.295132 2.201355 31.504958
Organ-8 981 46241 98.798077 3.735183 36.448815
Organ-9 1218 132143 87.141019 6.333512 36.699420
Organ-10 1212 116776 102.505143 6.505061 46.829904
Organ-11 1179 303374 167.225698 6.278122 41.896361
Organ-12 995 264328 74.416866 6.585659 20.422562

From all of the images general trends can be identified by examining all of the phenotypes and trying to identify the important ones for differentiating disease. The following figure shows the relationship between shape (of the super-pixels) and the healthy segments as pink dots and the abnormal as blue dots. The shape and position provide some differentation but are not definitive enough to clearly distinguish the two groups.

Here we show the relative color components for each channel (red, green, and blue) and how they are related to the tissue labeled as healthy and abnormal.

Here rather than showing the relative intensities we show the median absolute deviation which is better at characterizing the variation inside each structure.

The results can also be summarized as statistical outputs for comparing the phenotypes values for the groups of normal and abnormal tissue.

2.4 Machine Learning

The quantitatively meaningful data can then be used to train machine learning algorithms (decision trees to SVM) in order to automatically high-light many of the interesting regions and label them as such, dratistically reducing the required time to analyze a single patient.

Here we show a simple decision tree trained to identify lesions using color, position, texture and shape.

Classification Tree (Whole)

Furthermore the ability to parallelize and scale means thousands to millions of videos can be analyzed at the same time to learn even more about the structures of the digestive track and identify new possibilities for diagnosis.

2.5 How?

The first question is how the data can be processed. The basic work is done by a simple workflow on top of our Spark Image Layer. This abstracts away the complexities of cloud computing and distributed analysis. You focus only on the core task of image processing.

The true value of such a scalable system is not in the single analysis, but in the ability to analyze hundreds, thousands, and even millions of samples at the same time.

With cloud-integration and Big Data-based frameworks, even handling an entire city network with 100s of drones and cameras running continuously is an easy task without worrying about networks, topology, or fault-tolerance.

3 Technical Aspects

3.1 Processing the Data

Once the cluster has been comissioned and you have the SparkContext called sc (automatically provided in Databricks Cloud or Zeppelin), the data can be loaded using the Spark Image Layer. Since we are using real-time analysis, we acquire the images from an archive of images and create a database out of the results.

val iqaeDB = sc.createImageDatabase("s3n://chest-ct/scans/*/*.avi",
  patientInfo="jdbc://oracle-db/PATIENTS")
iqaeDB.registerImageTable("PatientImages")

Although we execute the command on one machine, the analysis will be distributed over the entire set of cluster resources available to sc. To further process the images, we can take advantage of the rich set of functionality built into Spark Image Layer.

The entire pipeline can then be started to run in real-time on all the new images as they stream in. If the tasks become more computationally intensive, then the computing power can be scaled up and down elastically.

4 Learn More

4Quant is active in a number of different areas from medicine to remote sensing. Our image processing framework (Spark Image Layer) and our query engine (Image Query and Analysis Engine) are widely adaptable to a number of different specific applications.

4.2 Technical Presentations

To find out more about the technical aspects of our solution, check out our presentation at:

5 Acknowledgements

Analysis powered by Spark Image Layer from 4Quant, Visualizations, Document Generation, and Maps provided by:

To cite ggplot2 in publications, please use:

H. Wickham. ggplot2: elegant graphics for data analysis. Springer New York, 2009.

A BibTeX entry for LaTeX users is

@Book{, author = {Hadley Wickham}, title = {ggplot2: elegant graphics for data analysis}, publisher = {Springer New York}, year = {2009}, isbn = {978-0-387-98140-6}, url = {http://had.co.nz/ggplot2/book}, }

To cite package ‘leaflet’ in publications use:

Joe Cheng and Yihui Xie (2014). leaflet: Create Interactive Web Maps with the JavaScript LeafLet Library. R package version 0.0.11. https://github.com/rstudio/leaflet

A BibTeX entry for LaTeX users is

@Manual{, title = {leaflet: Create Interactive Web Maps with the JavaScript LeafLet Library}, author = {Joe Cheng and Yihui Xie}, year = {2014}, note = {R package version 0.0.11}, url = {https://github.com/rstudio/leaflet}, }

To cite plyr in publications use:

Hadley Wickham (2011). The Split-Apply-Combine Strategy for Data Analysis. Journal of Statistical Software, 40(1), 1-29. URL http://www.jstatsoft.org/v40/i01/.

A BibTeX entry for LaTeX users is

@Article{, title = {The Split-Apply-Combine Strategy for Data Analysis}, author = {Hadley Wickham}, journal = {Journal of Statistical Software}, year = {2011}, volume = {40}, number = {1}, pages = {1–29}, url = {http://www.jstatsoft.org/v40/i01/}, }

To cite the ‘knitr’ package in publications use:

Yihui Xie (2015). knitr: A General-Purpose Package for Dynamic Report Generation in R. R package version 1.10.

Yihui Xie (2013) Dynamic Documents with R and knitr. Chapman and Hall/CRC. ISBN 978-1482203530

Yihui Xie (2014) knitr: A Comprehensive Tool for Reproducible Research in R. In Victoria Stodden, Friedrich Leisch and Roger D. Peng, editors, Implementing Reproducible Computational Research. Chapman and Hall/CRC. ISBN 978-1466561595

To cite package ‘rmarkdown’ in publications use:

JJ Allaire, Joe Cheng, Yihui Xie, Jonathan McPherson, Winston Chang, Jeff Allen, Hadley Wickham and Rob Hyndman (2015). rmarkdown: Dynamic Documents for R. R package version 0.7. http://CRAN.R-project.org/package=rmarkdown

A BibTeX entry for LaTeX users is

@Manual{, title = {rmarkdown: Dynamic Documents for R}, author = {JJ Allaire and Joe Cheng and Yihui Xie and Jonathan McPherson and Winston Chang and Jeff Allen and Hadley Wickham and Rob Hyndman}, year = {2015}, note = {R package version 0.7}, url = {http://CRAN.R-project.org/package=rmarkdown}, }

To cite package ‘DiagrammeR’ in publications use:

Knut Sveidqvist, Mike Bostock, Chris Pettitt, Mike Daines, Andrei Kashcha and Richard Iannone (2015). DiagrammeR: Create Graph Diagrams and Flowcharts Using R. R package version 0.7.

A BibTeX entry for LaTeX users is

@Manual{, title = {DiagrammeR: Create Graph Diagrams and Flowcharts Using R}, author = {Knut Sveidqvist and Mike Bostock and Chris Pettitt and Mike Daines and Andrei Kashcha and Richard Iannone}, year = {2015}, note = {R package version 0.7}, }