XSEDE-enabled High-throughput Caries Lesion Activity Assessment

1. XSEDE-enabled High-throughput Caries Lesion Activity Assessment Hui Zhang, Guangchen Ruan, Hongwei Shen, Michael Boyles, Huian Li, Masatoshi Ando Hui Zhang huizhang@iu.edu XSEDE'13 San Diego July 24th , 2013

2. Outline • Background • What is caries lesion activity • Scientific goal and computing objective • Dataset and Methods • Computing task implemented in a serial means • How Map-Reduce framework can be applied • Assessment Examples • Visualization and analysis • Qualitative and quantitative lesion activity assessment • Conclusion and Future Work

3. Introduction • Dental caries management project in IUSD (2010 ~) • Scientific goal: reduce, or reverse the prevalence of dental caries lesion active → inactive → reversed • Activelesion is a caries lesion that exhibits evidence of progression for a specific period of time • losing mineral content (or, demineralization) • Inactive/arrestedlesion is a caries lesion that exhibits no evidence of progression for a specific period of time • Reversed (with treatments) • gaining mineral content (or, remineralization )

4. Introduction • Lesion activity assessement (arrested or active) is important • essential and critical in dental studies • critical impact on dental treatment decision-making • incorrect determination can easily result in wrong treatment

5. Introduction • But ……. Today in dental clinical practice visual and tactile inspections are commonly used : • subjective • dependent on observer's experience to be accurate • results often in-consistent • tracking • temporal comparison Visual Assessment Tactile Sensation

6. Introduction • (Dental) Computing objective • Bring computers and computing technologies to dentistry research • dental imaging technology (µ-CT imaging→ cross-sectional dental scans) • image segmentation (cross-sectional scans→ ROIs) • visualization and analysis (lesion activity assessment → 3D-time series analysis) • Design methods not only for "marking" on dental scans, but also quantifying the volumetric information in the assessment • Use HPC and parallel computing to scale to larger datasets

7. a: Dimension b: Region of interest (ROI) Schematic diagrams showing specimen dimension (a), and region of interest (b). Datasets and Methods • The study reported 195 ground/polished 3x3x2mm blocks prepared from extracted human teeth collected from Indiana dental practitioners (approved by IU IRB#0306-64)

8. Datasets and Methods • Longitudinal dental experiment • uses 5-phase dem./rem. model • healthy1→dem2 →dem3→dem4 →rem5 • temporal evaluation • U-CTs • specimen/phase

9. Datasets and Methods • µ-CT Dental Scans • ~1000 scans per specimen per time point • each u-CT scan • 16-bit gray-scale image • 1548×1120 resolution • ~1.65 MB size • lesion on u-CT scan shows observable gray-scale difference

10. Datasets and Methods • 3D-Time Series Analysis Workflow (to quantify and compare volumetric lesion information over time) • Pre-analysis training • threshold, pivot values (based on histograms) • Region-of-interest (ROI) segmentation • blob detection, morphological operation • 3D construction • stacking ROIs, generating isosurface and geometry • Visual analysis (on volumetric models) • temporal comparison • How lesion evolves on same specimen • cross-conditional comparison • How lesion evolves with different treatments

11. Datasets and Methods • The Serial Implementation Model • A small collection of representative dental scans • threshold, valley grayscales, pivot values

12. Datasets and Methods • The Serial Implementation Model • A small collection of representative dental scans • threshold, pivot values • Segment ROIs on all scans (with established parameters) • binary image conversion • apply morphological operations (erosion and dilation) to remove false ROI candidates • blob detection → ROI boundary • processing images to keep only relevant pixels

13. Datasets and Methods • The Serial Implementation Model • Select representative dental scans • Threshold, pivot values • Segment ROIs on all scans • binary image conversion • apply morphological operations (erosion and dilation) to remove false ROI candidates • blob detection → ROI boundary • processing images to keep only relevant pixels • 3D construction • stack ROIs and visual analysis

14. Datasets and Methods • The Parallel Model • MapReduce - center around 2 func. to represent domain problems • General pattern Map(Di) → list(Ki,Vi); Reduce(Ki, list(Vi)) → list(Vf) • Divide the dataset D into individual data values Di • Map(Di)is applied to each individual value, producing many lists of key value pairs list(Ki,Vi) • Data produced by Map operations will be grouped by key Ki, producing associated values list(Vi) • Reduce(Ki, list(Vi))takes each key Ki and associated list of values list(Vi) to produce a list of final output values

15. Datasets and Methods • Lesion activity assessment using Map-Reduce • Map(Di) → list(Ki,Vi): • performs ROI segmentation; • extract image phaseID (encoded in filename); • produce (phaseID, roiByteArray) as key-value pair • Reduce(Ki, list(Vi)) → list(Vf) : • receives ROI collections keyed to phaseID; • performs 3D construction; • produce (phaseID, 3DModelByteArray) pair

16. Datasets and Methods • Better performance with sequence files and data compression • Hadoop excels in processing small # of large files • Too many I/O operations → extra burden • Implementation • Data packing before 3D-time series workflow • Map task loads images • Reduce task • produce sequence files • apply compression

17. Datasets and Methods • Computing setup and parameters • 64-node cluster on SDSC-Gordon • 8 Map slots 4 Reduce slots • Used DEFLATE codec and block compression for sequence files • 40,000 images in 12.62 minutes • More performance and scalability data reported in “ Exploting MapReduce and Data Compression for Data-intensive Applications“

18. Lesion Activity Assessment • Quantitative Assessment • lesion and its volumetric change measured in pixel^3 • objective and consistent comparisons across specimen and across different experimental conditions • scalable to larger datasets

19. Lesion Activity Assessment • 3D-Time Series Visualization • highlight lesion's volumetric changes B/A treatment

20. Lesion Activity Assessment • 3D-Time Series Visualization • show lesion's volumetric changes B/A treatment • combine dem. and rem. enamel in an integrated view with transparency

21. Lesion Activity Assessment • Shape Generation and Depth Measure • some studies concern finding the association between lesion depth and treatment variables previous effort: approximate lesion depth based grayscale on QLF images

22. Lesion Activity Assessment • Shape Generation and Depth Measure • some studies concern finding the association between lesion depth and treatment variables

23. Lesion Activity Assessment • Shape Generation and Depth Measure • some studies concern finding the association between lesion depth and treatment variables • 3D Poisson surfaces constructed for interactive depth measurement and comparison

24. Conclusion • Dental computing gives rise to a broad range of educational and treatment planning applications for dentistry; • A promising research approach that allows users to use imaging technology, computational algorithm, and visualization methods to make lesion activity assessment faster and more accurate; • The workflow can be supported computationally; implemented using parallel programming model such as MapReduce; further automated using HPC resources.

25. Future Work • Provide templates to other domains with similar computing task • Potential improvement of the workflow • The final result is much lighter compared to raw inputs • Data transfer with ROI boundary vectors instead of heavy image arrays • Compression of intermediate analysis results

26. Thank you! Questions?