CSE 594 - Medical Imaging
Home
Schedule
Labs
Links
Policies
Grades

Grades:

Lab1:

ID avg box gauss norm img filter fft img fft total comments
3525 10 10 10 10 10 6
 10
 66
 no study with different sigmas
9880 10 10 10 10 10 10 10 70
9960 10 10 10 10 10 10 10 70
170 10 10 10 10 10 10 10 70
2825 10 10 10 10 10 10 10 70
5405 10 10 10 10 10 10 10 70
6965 10 10 10 10 10 6 10 66 fft of box has very high DC
1690 10 10 6 10 10 6
 10 62 Gauss exponent has no sigma
no study with different box sizes
6127 10 10 10 10 10 10 10 70
3516 10 10 10 10 10 5
 10 65
 no study with different box sizes / sigma
4821






0
5774 10 10 10 10 10 7 10 67 Gaussian fft is really the shifted Gaussian
5423 10 10 10 10 10 10 10 70


Lab2

ID gconv bconv fft med filt edge unsharp window pyramid report total comments
3525 8 10 7 10 7 10 10
2 64 gauss blur almost not noticeable
should see lobes for box-fft
edge image should be grey
9880 10 10 10 10 10 10 10
4 74
9960 10 10 9 10 9 10 1 2 2 63 should notice the lobes for box fft
use the abs() function to combine the edge images
the window and pyramid routines are incorrect
170 10 10 10 10 10 10 10
10 80
2825 10 7 10 10 10 10 10

67 convolution with 1D box 
5405 10 10 10 10 8
10
6 64 not compared with native edge() fct
no unsharp fct
6965 8 10 10 10 8 7 5

58 clipped gaussian, need larger filter size for higher sigma
not compared with native edge() fct
unsharp mask not very effective
should use own windowing fct
1690 10
 10
 8
 10
 10
 8
 10

34 unsharp masking not very effective
gaussian frequency spectrum very similar to org
6127 10 10 10 10 10 7 10
3 70 unsharp mask images appear pseudo-colored
3516 10 10 7 10 10 10 10
2 69 should see lobes for box-fft
4821








0
5774 8 10 10 10 10 10 10
4 72 clipped gaussian, need larger filter size for higher sigma
5423 10 10 10 10 10 10 10 20 10 100

Notes for lab2:
1. If you received 8 pts in the gconv category, then you probably did not make the filter large enough for greater sigmas. This results in a Gaussian multiplied by a box, which when convolved, gives rise to the sinc-pattern in the frequency domain. In general, convolving with a Gaussian should NOT give rise to a sinc pattern in the frequency domain, because it falls smoothly in the frequency domain, with no side lobes. The sinc pattern that appears when convolving with the box filter is due to the multiplication of orginal spectrum with the sinc lobes of the box filter. The Gaussian does not have such lobes, therefore none appear.
2. Up to 10 extra points were given for reports that detailed their findings and/or explored the parameter space of the various methods, that is, provided more than just one image to proof that the routine works.

Lab3:

ID spread  spread/recon filter  filter/recon # of projs partial error total
3525 10 10 10 10 10 0 10 60
9880 10 10 10 10 10 10 10 70
9960 10 10 10 10 10 10 10 70
170 10 10 10 10 10 10 10 70
2825 10 10 10 10 10 10 10 70
5405 10 10 10 10 10 10 10 70
6965 10 10 10 10 10 10 10 70
1690 10 10 10 10 10 8 0 58
6127 10 10 10 10 10 8 0 58
3516 10 10 10 10 10 10 10 70
4821






0
5774 10 10 10 10 10 10 10 70
5423 10 10 10 10 10 10 10 70



















In the last question, the low-quality reconstructions (blurriness) is due to a prior smoothing of the sinogram (perhaps caused a detector of limited resolution). There is also some noise.
Nearly all people did not justify the need for the filtering in filtered backprojection (but no point deductions were given).. Please look at the notes again. There are two explanations: one intuitive (but only informal) and one formal, mathematical explanation.