Summary of the ISBI 23 Grand Challenge on 3D Deconvolution Microscopy Cédric Vonesch & Stamatios Lefkimmiatis Biomedical Imaging Group EPFL, Lausanne, Switzerland April 7, 23
Thanks to our sponsors! ISBI 23 Grand Challenge Organizers: Stephen Aylward Bram van Ginneken Bio Imaging & Signal Processing Technical Committee IEEE Signal Processing Society Financial support for awards! Funding: 2
About the organizers Main organizers: Expert committee members: Cédric Vonesch EPFL Lausanne, Switzerland Stamatios Lefkimmiatis EPFL Lausanne, Switzerland Laure Blanc-Féraud CNRS Sophia Antipolis, France Rainer Heintzmann Friedrich-Schiller-Universität Jena, Germany Arne Seitz EPFL Lausanne, Switzerland Michael Unser EPFL Lausanne, Switzerland 3
Primary goal: promoting cross-fertilization GFP Biology & Medicine Independent/ communitysupported developers Commercial software providers CCD Optics & Photonics Academic algorithm designers FFT Signal Processing & Applied Math 4
Overview of the challenge Months 3 ISBI special session Final Invite participants with best results to present their method 2 Qualification Determine the performance of proposed methods Larger data, less oracle information Training Familiarize participants with data sets, forward model, metrics... Simplified setting
Typology of the phantom data Manifold Structural Mathematical Corresponding Stain Color dimension class representation biological objects example channel (nm) Point sources, sub- Dirac distributions, Single molecules, MitoTracker Yellow/ 6 resolution structures narrow B-splines vesicles, mitochondria Orange Curves Bézier polynomials Microtubules, actin GFP Green 525 filaments 2 Surfaces Deformable Cellular or nuclear DiD Deep 675 ellipsoidal contours membranes Red 3 Dense volumes Wide B-splines Condensing DAPI, Blue 45 chromatine, DNA Hoechst 6
Summary of the forward model y = Q P(Tx + b)+n (, 2 I) Notations: x: ground-truth fluorophore distribution T: block-toeplitz matrix b: background signal (constant vector) 2 : variance of Gaussian noise Q: quantization and clipping operator Q(x) = 8 < arg min n2n x n if x>, : otherwise 7
Simulated micrographs (Maximum-intensity projections) 8
Simulated micrographs (Maximum-intensity projections) 9
Performance Metrics MSE-based Metrics Increase in signal-to-noise ratio (ISNR) ky xk ISNR = 2 log 2 x kˆx xk 2 : ground-truth ˆx : affine regressed reconstruction Normalized mean integrated squared error (NMISE) NMISE = N NX n= [T (x ˆx)] 2 n [Tx] n Structure Similarity Index (SSIM) Better correlation with human eye perception than SNR Mean-SSIM, Minimum-SSIM (over all slices of the 3D volume) Wavelet sparsity index Measure of the number of nonzero coefficients in the wavelet domain
Performance Metrics Derivative-based metrics Relative total variation error R = P N n= k[rˆx] n [rx] n k 2 P N n= k[rx] n k 2 Relative structure-tensor error P N n= R = k[sˆx] n k k[sx] n P k N n= k[sx] n k Relative Hessian-Frobenius error Sx = G rx rx T 3X q k[sx] n k = [ k] n k= R = P N n= k[hˆx] n k F k[hx] n k F P N n= k[hx] n k F k[hx] n k F = q [ ] 2 n +[ 2] 2 n +[ 3] 2 n
Performance Metrics Fourier-based metrics Fourier shell correlation: measures the normalized cross-correlation coefficient between two 3D volumes over corresponding shells in Fourier space FSC (!) = P! i 2! ˆX (! i ) X (! i ) q P!i2! ˆX (! i ) 2 P! i 2 X (! i) 2 Relative energy regain G R (!) = P! i 2! ˆX (! i ) X (! i ) 2 P! i 2! X (! i) 2 G R (!) = G R (!) = : spatially frequency domain perfectly reconstructed : no available frequency information 2
Results (top submissions).8.6.4.2.8.6.4.2.8.6.4.2.8.6.4.2.2.4.6.8.2.4.6.8.2.4.6.8.2.4.6.8 3
Results (top submissions).8.6.4.2.8.6.4.2.8.6.4.2.8.6.4.2.2.4.6.8.2.4.6.8.2.4.6.8.2.4.6.8 4
Results (top submissions).8.6.4.2.8.6.4.2.8.6.4.2.8.6.4.2.2.4.6.8.2.4.6.8.2.4.6.8.2.4.6.8 5
Results (top submissions).8.6.4.2.8.6.4.2.8.6.4.2.8.6.4.2.2.4.6.8.2.4.6.8.2.4.6.8.2.4.6.8 6
Results (top submissions).8.6.4.2.8.6.4.2.8.6.4.2.8.6.4.2.2.4.6.8.2.4.6.8.2.4.6.8.2.4.6.8 7
Results (top submissions).8.6.4.2.8.6.4.2.8.6.4.2.8.6.4.2.2.4.6.8.2.4.6.8.2.4.6.8.2.4.6.8 8
Summary of the challenge X Measured This year s awardees (SPS-sponsored prizes): Y. Ferréol Soulez, Lyon University ($5) 2. David Biggs, KB Imaging Solutions ($3) 3. Hiep Luong, Gent University ($2) Next edition: Deconvolved Online submission system to reopen soon Take up the challenge & submit your result! Best entries will win ISBI 24 travel grants! Visit our website for updates: 24 bigwww.epfl.ch/deconvolution/challenge X Z Measured Deconvolved 9