Beyond Low Rank + Sparse: Multi-scale Low Rank Reconstruction for Dynamic Contrast Enhanced Imaging Frank Ong 1, Tao Zhang 2, Joseph Cheng 2, Martin Uecker.

PowerPoint Presentation

Beyond Low Rank + Sparse:Multi-scale Low Rank Reconstruction for Dynamic Contrast Enhanced ImagingFrank Ong1, Tao Zhang2, Joseph Cheng2, Martin Uecker1 and Michael Lustig1

Contact: [email protected]

1 University of California, Berkeley2 Stanford UniversitySpeaker Name: Frank Ong

I have the following financial interest or relationship to disclose with regard to the subject matter of this presentation:

Company Name: GE HealthcareType of Relationship: Research SupportDeclaration ofFinancial Interests or Relationships

23D Dynamic Contrast Enhanced (DCE) MRIDetect and characterize lesionsAny imaging plane and any time point

Challenges with 3D DCE-MRIHuge amount of data in a fixed time windowTradeoff between spatial and temporal resolution33D DCE Dataset# 0575 Multi-scale Low Rank MRIDCE imaging has become a popular imaging technique to detect and characterize lesions3D DCE MRI can further allow radiologists to obtain these information at any imaging plane and any given time point----- Meeting Notes (5/22/15 12:06) -----bleeding, labeling3GoalCapture true contrast dynamics of DCE

we dont need to achieve high spatiotemporal resolution for everything4

3D DCE DatasetA multi-scale low rank reconstruction to capture these structured dynamicsDCE contrast dynamics are structured:Fast and sparse blood vessel dynamicsLocally correlated kidney and liver dynamicsStationary smooth tissues# 0575 Multi-scale Low Rank MRI4Previous worksGlobally Low Rank [1]Captures globally correlated dynamicsVery sensitive to local dynamics5[1] Pedersen et al. 2009, Liang et al. ISBI 2006[2] Ricardo et al. MRM 2014, Chandrasekaran et al. SIAM J. Optim., Candes et al. J ACM 2011[3] Lingala et al. TMI 2011, Trzasko et al, ISMRM 2011, Zhang et al. MRM 2015

Time

SpaceLowRank

# 0575 Multi-scale Low Rank MRIPrevious worksGlobally Low Rank [1]Captures globally correlated dynamicsVery sensitive to local dynamicsLow Rank + Sparse [2]Captures global + sparse dynamicsMany dynamics are not sparse!6[1] Pedersen et al. 2009, Liang et al. ISBI 2006[2] Ricardo et al. MRM 2014, Chandrasekaran et al. SIAM J. Optim., Candes et al. J ACM 2011[3] Lingala et al. TMI 2011, Trzasko et al, ISMRM 2011, Zhang et al. MRM 2014

Time

SpaceLowRank

+# 0575 Multi-scale Low Rank MRI

Globally Low Rank [1]Captures globally correlated dynamicsVery sensitive to local dynamicsLow Rank + Sparse [2]Captures global + sparse dynamicsMany dynamics are not sparse!Locally Low Rank [3]Captures local correlated dynamicsDoes not use global correlation7[1] Pedersen et al. 2009, Liang et al. ISBI 2006[2] Ricardo et al. MRM 2014, Chandrasekaran et al. SIAM J. Optim., Candes et al. J ACM 2011[3] Lingala et al. TMI 2011, Trzasko et al, ISMRM 2011, Zhang et al. MRM 2014Space

LowRankLowRankLowRankLowRankPrevious works# 0575 Multi-scale Low Rank MRITime

Combines all three methodsCaptures all scales of dynamicsSparseGloballyLow RankThis Work: Multi-scale Low Rank MRILocallyLow Rank8# 0575 Multi-scale Low Rank MRI8Multi-scale Low Rank MRI Decomposition1x14x416x1664x64128x112Fully-Sampled Dataset9=

+++# 0575 Multi-scale Low Rank MRI# 0575 Multi-scale Low Rank MRI

Illustration of the iterative reconstruction

Timekspace

10# 0575 Multi-scale Low Rank MRI

Timekspace

11Can be formally derived using ADMMIllustration of the iterative reconstructionGoodies of the Multi-scale Low RankConvex program -> always convergesTheoretically guided thresholds:Low Rank + Sparse: ~1 for sparse, image size for low rankMulti-scale Low Rank: ~ block sizeComputationally efficient:At most 2x more computation than conventional low rank methods12# 0575 Multi-scale Low Rank MRIComparison Results13

Globally Low RankLow Rank + SparseMulti-scale Low Rank# 0575 Multi-scale Low Rank MRICompressed sensing (Poisson Disk) undersampling [1]Parallel Imaging (ESPIRiT) [2]Free-breathing Respiratory Soft-gated (Butterfly Navigator) [3]Resolution: 1x1.4x2 mm3 and ~8s[2] Uecker et al. MRM 2014, [3] Cheng et al. JMRI 2014, Zhang et al. JMRI 2013

Locally Low Rank

Comparison Results14Multi-scale Low Rank

# 0575 Multi-scale Low Rank MRI[1] Uecker et al. MRM 2014, [2] Cheng et al. JMRI 2014Globally Low RankLow Rank + Sparse

Compressed sensing (Poisson Disk) undersampling [1]Parallel Imaging (ESPIRiT) [2]Free-breathing Respiratory Soft-gated (Butterfly Navigator) [3]Resolution: 1x1.4x2 mm3 and ~8s

Locally Low Rank

Globally Low RankLocally Low RankLow Rank + Sparse15Multi-scale Low Rank# 0575 Multi-scale Low Rank MRIComparison Results----- Meeting Notes (5/22/15 11:53) -----temporal curves, bleeding15

Globally Low RankLocally Low RankLow Rank + Sparse

16

Multi-scale Low Rank

# 0575 Multi-scale Low Rank MRIComparison Results----- Meeting Notes (5/22/15 11:53) -----temporal curves, bleeding16Conclusion:Multi-scale low rank reconstructionCaptures the right dynamics at the right scaleGeneralizes low rank + sparseThank You!17Berkeley Advanced Reconstruction Toolbox (BART)http://www.eecs.berkeley.edu/~mlustig/Software.html1x14x416x1664x64128x112=

+++# 0575 Multi-scale Low Rank MRI