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

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 3 / 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

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 6 / 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