Local Statistical Filtering via Domain Dissection for Medical - PowerPoint PPT Presentation

Local Statistical Filtering via Domain Dissection for Medical Imaging GTC 2016 – San Jose, CA, USA Alexandros-Stavros Iliopoulos 1 Dimitris Floros 2 Nikos Pitsianis 2 , 1 Xiaobai Sun 1 Fang-Fang Yin 3 Lei Ren 3 1 Department of Computer Science, Duke University 2 Department of Electrical and Computer Engineering, Aristotle University of Thessaloniki 3 Department of Radiation Oncology, Duke University School of Medicine April 6, 2016 Iliopoulos, Floros, Pitsianis , Sun, Yin, Ren (Duke|AUTh) Locally Adaptive Signal-Noise Analysis GTC-2016 Apr 6, 2016 1 / 33

1 Spatially variant signal-noise analysis: motivation needs & challenges LA-SAS contribution 2 Locally adaptive signal-noise analysis formulation example filters analytic advance & technical challenges 3 LA-SAS: design & development design principle: multi-layer configuration domain dissection: local adaptivity & global concurrency CUDA LA-SAS experimental results 4 Recap & discussion 5 References Iliopoulos, Floros, Pitsianis , Sun, Yin, Ren (Duke|AUTh) Locally Adaptive Signal-Noise Analysis GTC-2016 Apr 6, 2016 2 / 33

Spatially variant signal-noise analysis: needs & challenges • Noise is prevalent in medical images – multiple sources (acquisition, processing, ...) – multiple types (Gaussian, Poisson, scatter, ...) • Noise study: characterization & suppression – critical to high-fidelity analysis (noise propagation in processing pipeline e.g. gradient calculation) – need effective tools for systematic investigation – speed important for on-board imaging applications • Challenging conditions – valuable low-contrast content (especially in CT) – acquisition constraints (resolution, imaging dose) – motion: nonlinear intensity-deformation relationship pelvis cone-beam OBI ⋆ spatial variance (125 kV, coronal projection) (w.r.t. material, density, acquisition set-up) with spatially variant scattering Iliopoulos, Floros, Pitsianis , Sun, Yin, Ren (Duke|AUTh) Locally Adaptive Signal-Noise Analysis GTC-2016 Apr 6, 2016 4 / 33

Spatially variant signal-noise analysis: contribution 9000 8000 7000 6000 • LA-SAS Frequency 5000 4000 (locally adaptive signal-noise analysis system) 3000 – revealing local noise statistics and signal structure 2000 – filtering in adaptation to local structures 1000 0 – enabling effective noise suppression 0 0.5 1 1.5 2 2.5 3 3.5 Range bins 3500 500 450 3000 400 • LA-SAS design and development 2500 350 300 Frequency Frequency 2000 250 – basic operations 1500 200 150 1000 – versatile filter composition 100 500 50 – CUDA LA-SAS (efficiency) 0 0 1.6 1.8 2 2.2 2.4 2.6 2.8 3 1.9 2 2.1 2.2 2.3 2.4 2.5 2.6 2.7 Range bins Range bins range histograms: global region (top) vs. nested sub-regions (bottom) Iliopoulos, Floros, Pitsianis , Sun, Yin, Ren (Duke|AUTh) Locally Adaptive Signal-Noise Analysis GTC-2016 Apr 6, 2016 5 / 33

Locally adaptive signal-noise analysis: problem description • Dual task of an analysis/filtering mechanism ( F ) I ( x ) = ˆ x ∈ Ω ⊂ R D I ( x ) + η ( x ) , – detect/reconstruct unknown signal, ˆ I – estimate/suppress unknown noise, η • Adaptation to local variation ⎞ )︂ ˆ ∑︂ x ′ , I ( x ′ ); p 𝒪 ( x ) I ( x ) := F x ′ ∈𝒪 ( x ) – based on local statistics, p N ( x ) , over spatial neighborhood, N ( x ) (mean, median, deviation, range distribution, etc) – preserving signal structure (smooth subregions, discontinuities at region boundaries, etc) Iliopoulos, Floros, Pitsianis , Sun, Yin, Ren (Duke|AUTh) Locally Adaptive Signal-Noise Analysis GTC-2016 Apr 6, 2016 7 / 33

Locally adaptive filtering example: median ˆ I ( x ) = p 𝒪 ( x ) = median 𝒪 ( x ) { I ( x ) } • basic denoising & processing sub-module (regional dynamic range) median Ąlter output residual image (5 × 5) 3500 3000 2500 Frequency 2000 1500 1000 500 0 Chung et al . NSS/MIC , 2010 1.6 1.8 2 2.2 2.4 2.6 2.8 3 Range bins Iliopoulos, Floros, Pitsianis , Sun, Yin, Ren (Duke|AUTh) Locally Adaptive Signal-Noise Analysis GTC-2016 Apr 6, 2016 8 / 33

Locally adaptive filtering example: entropy p 𝒪 ( x ) = Pr 𝒪 ( x ) { I ( x ) } ∑︂ H ( x ) = − p 𝒪 ( x ) log( p 𝒪 ( x ) ) • multimodal registration • basic step for other processing modules (e.g. segmentation, histogram equalization) (regional dynamic range) local entropy map (9 × 9) 3500 3000 2500 Frequency 2000 1500 1000 500 0 Zhang et al . ICBBE , 2008 Pluim et al . IEEE TMI (22), 2003 1.6 1.8 2 2.2 2.4 2.6 2.8 3 Range bins Iliopoulos, Floros, Pitsianis , Sun, Yin, Ren (Duke|AUTh) Locally Adaptive Signal-Noise Analysis GTC-2016 Apr 6, 2016 9 / 33

Locally adaptive filtering example: histogram equalization (HE) p 𝒪 ( x ) = hist[ r , I ( N ( x ))] where hist: local histogram r : quantized ranges • local contrast enhancement • local + global (global dynamic range) global HE local HE ( adapthisteq ) distribution information to be replaced with overlapping LHE 9000 8000 7000 6000 Frequency 5000 4000 3000 2000 1000 0 Zhu et al . CVIA (73), 1999 0 0.5 1 1.5 2 2.5 3 3.5 Range bins Iliopoulos, Floros, Pitsianis , Sun, Yin, Ren (Duke|AUTh) Locally Adaptive Signal-Noise Analysis GTC-2016 Apr 6, 2016 10 / 33

Locally adaptive filtering example: bilateral filter (BF) p 𝒪 ( x ) = σ r ( x ) ⊗ ‖ x ⊗ x ′ ‖ 2 2 k s ( x , x ′ ) = e σ 2 s ⊗ ‖ I ( x ) ⊗ I ( x ′ ) ‖ 2 2 k r ( I ( x ) , I ( x ′ )) = e σ 2 r ( x ) (space- and range-kernels) • boundary-preserving denoising (regional dynamic range) global BF locally adaptive BF σ s = 1 . 5 , σ r = 0 . 157 σ s = 1 . 5 , • local adaptation 3500 σ r ( x ) = mad( 𝒪 9 × 9 ( x )) 3000 2500 to boundary “jumps” Frequency 2000 1500 1000 500 0 1.6 1.8 2 2.2 2.4 2.6 2.8 3 Range bins Tomasi & Manduchi. ICCV , 1998 Iliopoulos, Floros, Pitsianis , Sun, Yin, Ren (Duke|AUTh) Locally Adaptive Signal-Noise Analysis GTC-2016 Apr 6, 2016 11 / 33

Locally adaptive filtering example: bilateral filter (BF) p 𝒪 ( x ) = σ r ( x ) ⊗ ‖ x ⊗ x ′ ‖ 2 2 k s ( x , x ′ ) = e σ 2 s ⊗ ‖ I ( x ) ⊗ I ( x ′ ) ‖ 2 2 k r ( I ( x ) , I ( x ′ )) = e σ 2 r ( x ) (space- and range-kernels) • boundary-preserving denoising (regional dynamic range) global BF locally adaptive BF σ s = 1 . 5 , σ r = 0 . 157 σ s = 1 . 5 , • local adaptation 3500 (residual image) σ r ( x ) = mad( 𝒪 9 × 9 ( x )) 3000 2500 to boundary “jumps” Frequency 2000 (residual image) 1500 1000 500 0 1.6 1.8 2 2.2 2.4 2.6 2.8 3 Range bins Tomasi & Manduchi. ICCV , 1998 Iliopoulos, Floros, Pitsianis , Sun, Yin, Ren (Duke|AUTh) Locally Adaptive Signal-Noise Analysis GTC-2016 Apr 6, 2016 11 / 33

Locally adaptive filtering: analytic advance & technical challenges • Reveal and preserve spatially variant signal structure (same as in conventional methods with spatial adaptivity) • Permit spatially inhomogeneous noise behavior (often observed in medical imaging) • Depart from filtering algorithms with predetermined, global parameters (including histograms and some bilateral filters) Iliopoulos, Floros, Pitsianis , Sun, Yin, Ren (Duke|AUTh) Locally Adaptive Signal-Noise Analysis GTC-2016 Apr 6, 2016 12 / 33

Locally adaptive filtering: analytic advance & technical challenges • Reveal and preserve spatially variant signal structure (same as in conventional methods with spatial adaptivity) • Permit spatially inhomogeneous noise behavior (often observed in medical imaging) • Depart from filtering algorithms with predetermined, global parameters (including histograms and some bilateral filters) • Challenge traditional parallel primitives in multiple aspects (algorithmic complexity, concurrency, numerical behavior) Iliopoulos, Floros, Pitsianis , Sun, Yin, Ren (Duke|AUTh) Locally Adaptive Signal-Noise Analysis GTC-2016 Apr 6, 2016 12 / 33

Local Statistical Filtering via Domain Dissection for Medical - PowerPoint PPT Presentation

Local Statistical Filtering via Domain Dissection for Medical Imaging GTC 2016 San Jose, CA, USA Alexandros-Stavros Iliopoulos 1 Dimitris Floros 2 Nikos Pitsianis 2 , 1 Xiaobai Sun 1 Fang-Fang Yin 3 Lei Ren 3 1 Department of Computer Science,

Content A Polyreference Least Square Complex Frequency Introduction domain based statistical

Part II: Cow Eye Dissection Dissection of a Cow s Eye After the front of the eye has been

Statistical Filtering and Control for AI and Robotics Part I. Bayes filtering Riccardo Muradore

Chronic Aortic Dissection Treatment of Chronic Aortic Dissection Is Evolving 40% of repairs of

Statistical Filtering and Control for AI and Robotics Part II. Linear methods for regression

Least-Action Filtering L. C. G. Rogers Statistical Laboratory, University of Cambridge

Problem Domain Collaborative filtering (CF)-based recommender systems (RS). Issue:

Digital Image Analysis and Processing CPE 0907544 Filtering in the Frequency Domain Part I

1 An Filtering System that Monitors Document Search Engines Can Help, But Not Enough!

BBM 413 Image filtering: computes a function of a local neighborhood at Fundamentals of each

What is domain adaptation?

Statistical Filtering and Control for AI and Robotics Part III. Extended Kalman filter, Particle

Lesson 4. Iterated filtering: principles and practice Edward Ionides, Aaron A. King, and Kidus

Review-Based Cross-Domain Collaborative Filtering: a Neural Framework Thanh-Nam Doan, Sherry

Lecture 4 Filtering in the Frequency Domain Lin ZHANG, PhD School of Software Engineering

The Fourier Transform II Image Analysis Filtering in the frequency domain The Fourier transform

ECG782: Multidimensional Digital Signal Processing Filtering in the Frequency Domain

ECG782: Multidimensional Digital Signal Processing Spatial Domain Filtering

Domain Adaptation in Statistical Machine Translation Logic, Language and Computation Bart

Protecting Internet Threat Monitors: A Statistical Filtering Approach Yoichi Shinoda JAIST

Traffic Control Mechanisms Filtering Source address filtering Other forms of filtering

CS490W: What is Collaborative Filtering? Collaborative Filtering (CF): Making recommendation

Nonlinear filtering with local couplings Alessio Spantini Ricardo Baptista Youssef Marzouk

Type B Dissection Sub-Categories Acute Complicated Rupture Malperfusion Title