Optimization for Radiology and Social Media

1. Optimization for Radiology and Social Media Ken Goldberg IEOR (EECS, School of Information, BCNM) UC Berkeley College of Engineering Research Council, May 2010

2. Outline • IEOR Dept, BCNM • Radiology • Social Media

3. UC Berkeley IEOR Department • The only IEOR department in the UC system • Ranked #3 in USA • 55 BS, 10 BA, 30 MS, 5-8 PhD degrees per year

4. IEOR Faculty:Ilan AdlerAlper AtamturkJon BurgstoneYing-Ju ChenLaurent El Ghaoui Ken GoldbergXin GuoDorit S. HochbaumRichard Karp Philip M. KaminskyRobert C. LeachmanAndrew LimShmuel S. OrenChristos PapadimitriouRhonda L. Righter (Chair)Lee W. SchrubenZuo-Jun "Max" ShenIkhlaq SidhuCandace Yano

5. Mission To critically analyze and shape developments in new media from trans-disciplinary and global perspectives that emphasize humanities and the public interest. bcnm.berkeley.edu

6. Humanities Philosophy Rhetoric Journalism Art History Education Architecture iSchool Public Health Film Studies Theater IEOR BAMPFA CITRIS Music EECS Art Practice ME Technology Art/Design BioE New Media Initiative

7. Radiology Ken Goldberg, AlperAtamturk, Laurent El Ghaoui (IEOR) James O’Brien, Jonathan Shewchuck (EECS) I.-C. Hsu, MD, J. Pouliot, PhD (UCSF)

8. Prostate Cancer 1 in 6 men will be diagnosed with prostate cancer over 230,000 cases each year in the US one death every 16 minutes

9. High Dose Rate Brachytherapy http://www.prostatebrachytherapyinfo.net/PCT21.html http://automation.berkeley.edu/projects/needlesteering/

10. Robot Motion Planning • Theorem (Completeness): A sensorless plan exists for any polygonal part. • Theorem (Complexity): For a polygon of n sides, the algorithm runs in time O(n2) and finds plans of length O(n). • Extensions: • Stochastically Optimal Plans • Extension to Non-Zero Friction • Geometric Eccentricity / constant time complexity • Part Fixturing and Holding

11. Dosimetry: Inverse Planning Dose Distribution Dosimetric Criteria

12. Inverse Planning with Simulated Annealing (IPSA) • Inverse planning software developed at UCSF by Pouliot group • FDA-approved: used clinically worldwide • Simulated annealing dose point penalty method

13. Inverse Planning with Linear Programming (IPLP) • LP formulation (UC Berkeley) • Guarantees global optima • Optimization of HDR Brachytherapy Dose Distributions using Linear Programming with Penalty Costs. Ron Alterovitz, Etienne Lessard, Jean Pouliot, I-Chow Joe Hsu, James F. O'Brien, and Ken Goldberg. Medical Physics, vol. 33, no. 11, pp. 4012-4019, Nov. 2006.

14. Limitations of Penalty Model • Only specifies dosimetry at dose points, not to organs • Not equivalent to dosimetric indices • Not intuitive for Physicians • Results not always clinically viable. • Results difficult to customize for special cases • Dosimetric index: if dose at x > R, then x = 1, x = 0 otherwise • Discrete Variables

15. Inverse Planning with Integer Programming (IP2) • Indices • s: organ • i: point in organ • j: dwell position • Variables • tj: dwell time at j • xsi: counting variable for s,i • Parameters: • Dsij: dose rate from j to s,i • Rs: Dose threshold for s • Ms: Max dose for points in s • Ls: Lower bound for dosimetric s • Us: Upper bound for dosimetric s • Model • Maximize Σ x0i • Subject to: • Σ Dsij tj ≥ Rs xsi • Σ Dsij tj ≤ Rs + (Ms – Rs) xsi • Ls ≤ Σ xsi ≤ Us • tj ≥ 0 • xsiє {0,1}

16. Initial Results: Comparing IPSA with IP2 • Average Runtime (sec): • IPSA: 5 • IP2 (heuristic 1): 23 • IP2 (heuristic 2): 900 • Compliance with all clinical criteria • IPSA: 0% of patients • IP2 (heuristic 1): 95% of patients • IP2 (heuristic 2): 100% of patients

17. IP2 for Needle Reduction • Minimize number of needles • Minimize trauma • Speed Recovery Possible Needles Optimal Needle Selection (example)

18. Conic Optimization Robust Optimization Model uncertainties in: Organ location, motion Edema Catheter displacement Future Work

19. Tissue Simulation http://graphics.cs.berkeley.edu/papers/Chentanez-ISN-2009-08/ Nuttapong Chentanez, Ron Alterovitz, Daniel Ritchie, Lita Cho, Kris K. Hauser, Ken Goldberg, Jonathan R. Shewchuk, and James F. O'Brien. "Interactive Simulation of Surgical Needle Insertion and Steering". In Proceedings of ACM SIGGRAPH 2009, pages 88:1–10, Aug 2009.

20. Superhuman Performance of Surgical Tasks by Robots using Iterative Learning from Human-Guided Demonstrations Jur van den Berg, Stephen Miller, Daniel Duckworth, Humphrey Hu, Andrew Wan, Xiao-Yu Fu, Ken Goldberg, Pieter Abbeel University of California, Berkeley

21. Method: • 1. Robot learns surgical task from human demonstrations • Knot tying • Suturing • 2. Robot learns to execute tasks with superhuman performance • Increase smoothness • Increase speed

22. Social Media Ken Goldberg, Gail de Kosnik, Kimiko Ryokai Alec Ross, Katie Dowd (US State Dept)

23. collaborative robot control: … … Batch … MultiTasking … Collaborative

26. Motivation Goals of Organization • Engage community • Understand community • Solicit input • Understand the distribution of viewpoints • Discover insightful comments Goals of Community • Understand relationships to other community members • Consider a diversity of viewpoints • Express ideas, and be heard

27. Motivation Classical approaches: surveys, polls Drawbacks: limited samples, slow, doesn’t increase engagement Current approaches: online forums, comment lists Drawbacks: data deluge, cyberpolarization, hard to discover insights

28. Related Work: Visualization Clockwise, starting from top left: Morningside Analytics, MusicBox, Starry Night

29. Related Work: Info Filtering • K. Goldberg et al, 2001: Eigentaste • E. Bitton, 2009: spatial model • Polikar, 2006: ensemble learning

31. Six 50-minute Learning Object Modules, preparation materials, slides for in-class lectures, discussion ideas, hand-on activities, and homework assignments.

32. Canonical Correlation Analysis (CCA) z • Observed variables: x, y • Latent variable: z • Learn MLEs for low-rank projections A and B • Equivalently, find inverse mapping that maximizes correlation between A, B x y Graphical model for CCA x = Az + ε y = Bz + ε z = A-1x = B-1y

33. z x y

34. Canonical Correlation Analysis (CCA) • CCA gives three posterior expectations • E(z|x) • E(z|y) • E(z|x,y) • E(z|x,y) is used to visualize the opinion space Opinion Vector x z y Textual Comment

35. Canonical Correlation Analysis (CCA) Each point in the Canonical representation has an expected list of words associated to it. A visualization of this list of words can be used to give users more information about their location

37. Opinion Space: Crowdsourcing Insights Scalability: n Participants, n Viewpoints n2 Peer to Peer Reviews Viewpoints are k-Dimensional Dim. Reduction: 2D Map of Affinity/Similarity Insight vs. Agreement: Nonlinear Scoring Ken Goldberg, UC Berkeley Alec Ross, U.S. State Dept

38. Optimization for Radiology and Social Media Ken Goldberg IEOR (EECS, School of Information, BCNM) UC Berkeley College of Engineering Research Council, May 2010

41. Capping Allocate dose budget to dose points that are likely to need it. : Solve LP relaxation Analyze solution and impose new constraints on hottest dose points. Resolve to feasible solution. Hard Cuts Apply custom cuts so that IP2 emphasizes dosimetric indices. : Solve LP relaxation. Add cuts to incorrectly counted dose points. Repeat until feasible for IP2 IP2 Heuristics

42. Hard Cuts x Hard cut 1 Fractional Optimal Solution (cut off by Hard cut) Constraints 0 dose

43. Dimensionality Reduction Principal Component Analysis (PCA) • Assumes independence and linearity • Minimizes squared error • Scalable: compute position of new user in constant time

44. Approach: Visualization

45. Approach: Level the Playing Field

46. Approach: Wisdom of Crowds

47. “We’re moving from an Information Age to an Opinion Age.” - Warren Sack, UCSC

48. Berkeley Center for New Media (BCNM): David Wong: EECS Undergraduate Student Tavi Nathanson: EECS Graduate Student Ephrat Bitton: IEOR Graduate Student Siamak Faridani: IEOR Graduate Student Elizabeth Goodman: School of Information Graduate Student Alex Sydell: EECS Undergraduate Student Meghan Laslocky: Outside Consultant on Content Ari Wallach: Outside Consultant on Content and Strategy Steve Weber: Outside Consultant on Content Peter Feaver: Outside Consultant on Content U.S. State Department: Alec Ross: Senior Advisor for Innovation Katie Dowd: New Media Director Daniel Schaub: Director for Digital Communications

49. Multidimensional Scaling • Goal: rearrange objects in low dim space so as to reproduce distances in higher dim • Strategy: Rearrange & compare solns, maximizing goodness of fit: • Can use any kind of similarity function • Pros • Data need not be normal, relationships need not be linear • Tends to yield fewer factors than FA • Con: slow, not scalable j δij i j dij i

50. Kernel-based Nonlinear PCA • Intuition: in general, can’t linearly separate n points in d < n dim, but can almost always do so in d ≥ n dim • Method: compute covariance matrix after transforming data into higher dim space • Kernel trick used to improve complexity • If Φ is the identity, Kernel PCA = PCA