Stellaris FISH dataset #1 Deconvolution tests

Recently George McNamara, of M.D. Anderson Cancer Center posted some datasets and invited the community to share deconvolution results (description here) and compare to results he generated using a GPU algorithm.  The experiments below are done using a cropped version of the “GM 20131101Fri_StellarisFISH_1_w43 = Cy3 -N3” set. (cropped version can be downloaded here: GM 20131101Fri_StellarisFISH_1_w43 = Cy3 -N3- ROI)

Figure 1. shows an XY and YZ slice view of the original data, and figure 2 shows the Bruce/Butte deconvolution result (provided by George McNamara).  The YZ view of the original image shows spherical aberration consistent with the description of the experiment (“Imaging conditions are not perfect because the cells were in aqueous solution”).  Note that in the Bruce/Butte deconvolution result some residual aberration can still be seen in YZ.

In an attempt to address aberrations the ImageJ PSF Generator Plugin from BIG Group EPFL was used to generate an aberrated PSF model.   Following suggestions from Model etc. al the PSF model was set as Gibson-Lanni, the specimen RI was approximated to be 1.4, and the particle depth was set to be near the middle slice of the stack (6400 nm).  All other PSF parameters were set using the values  provided with the data.  The PSF with aberrations is shown in Figure 3.  The Deconvolution Lab ImageJ plugin (also from BIG group) was used to deconvolve the image using the aberrated PSF.  The Richardson Lucy algorithm was used with 600 iterations.  The result is shown in Figure 4 and the data set can be found here (Richardson-Lucy 600n_sri_1.4).    The ImageJ result using a PSF without aberrations (depth=0) was also generated as a comparison (Figure 5.).

Note that when using the aberrated PSF the deconvolution result is more symmetrical in YZ (Figure 4) as compared to the CUDA result (Figure 2) and the ImageJ result without considering aberration (Figure 5.).    Further improvement could perhaps be obtained by applying axial scaling (Model), an even better model of the PSF (better estimates of the sample parameters or by measuring it), perhaps by using a depth variant model (such as here), or by using a more sophisticated deconvolution algorithm (total variation RL and wavelet based are both available in the ImageJ plugin).

The speed reported for the CUDA algorithm (in the order of seconds!!) is very good (the entire process, PSF generation + deconvolution  takes minutes with ImageJ BIG lab components, comparison with other implementations would be interesting).   If options exist to include aberrations in the CUDA algorithm PSF model even better results may be obtainable with the CUDA deconvolution.

XY 32   YZ 136

Figure 1: Original Image Slice Views, XY (z=32) and YZ (x=136)

XY 32_cuda    YZ 136_cuda

Figure 2:  Bruce, Butte Cuda Deconvolution, Slice Views, XY (z=32) and YZ (x=136)

XY 32 PSF_d6000_ri1.4    YZ 128 PSF_d6000_ri1.4

Figure 3: PSF with aberration.

XY 32_RL600_d6000_ri1.4    YZ 136_RL600_d600_RI1.4

Figure 4: ImageJ Result, 600 iterations Richardson Lucy using PSF with aberrations

XY 32_RL600_d0_ri1.5   YZ 136_RL600_d0_ri1.5

Figure 5: ImageJ Result, 600 iterations Richardson Lucy using PSF at coverslip (depth=0)


Leave a Reply

Your email address will not be published. Required fields are marked *