CS463/516 assignment 1
Due: Thursday, May 27th at 11:59 PM (Eastern time).
Submit: a pdf report showing your source code, displayed images, and any explanation/notes.
Topic: Basic numpy, medical imaging modalities, SNR, and denoising
In this assignment we will familiarize ourselves with basic numpy array manipulation, different imaging
modalities, and gain understanding of some basic image features (contrast and SNR).
Setup
First, install spyder (or any python IDE). If you are on windows, you will probably want to install
anaconda first: https://docs.anaconda.com/anaconda/install/
• Note – you don’t need to use spyder (any python IDE is fine) but it will help a lot in this course
because the ipython console is very useful for debugging and visualization of intermediate
results.
Second, get the images (link below). This directory contains 8 images:
The images were all acquired using MRI scanner, with the exception of ct.nii.gz and meanpet.nii.gz,
which are from CT and PET scanner, respectively. Most of the images are 3D, with the exception of
cardiac_axial.nii.gz, cardiac_realtime.nii.gz, and fmri.nii.gz, which are 4D.
https://drive.google.com/file/d/10_bOHQmfe9WkTlK9Td5IhjVHhBp-s2ni/view?usp=sharing
Part 1: (25%): simple plotting with matplotlib
a) Display all images middle z-slice (3rd dimension, axis=2) (as seen below). Use the ‘jet’ color scale
(instead of the gray color scale that I use below in my example). Above each image, show the title of the
modality. Remove the x/y axis labels (as I did below). Use plt.subplot
b) display the minimum intensity projection (MIP) for the swi.nii.gz, and the maximum intensity
projection (MIP) for the TOF (in jet color map).
Left: part (a) – replicate this
but use jet color map
instead of gray.
Left: part (b). note how
the blood vessels are
displayed prominently
due to the projection.
You will need to restrict
the z-slices from the SWI
to achieve a good MIP.
Use np.min and
np.max.
Part 2 (25%): contrast estimation
Using numpy, get 3 different contrast measures for each image (root mean square, Michelson, and
entropy, see lecture 3 slide 4, 5). Report the contrast (all 3 versions) in the title of the plots in figure 1a.
base your contrast estimation on the entire 3D or 4D image (not just the slice shown in the figures).
Part 3 (25%): SNR estimation, quantifying noise
Using the method outlined in the lecture 3 slide 7, report the SNR for each of the modalities. Which
modality has the highest SNR and which has the lowest?
Plot histograms of the noise in each image. What type of distribution does the noise follow?
To display the solution to part 3, create a new figure (as in part 1) and display the noise histogram of
each image (instead of the image itself) in each sub plot. Show the SNR as the title above each histogram
(along with the image name).
*caution – when selecting your noise patch, be sure the patch isn’t all zeros, otherwise your noise will
be estimated as 0 and the SNR will be infinite*
Part 4 (25%): linear filtering
Using the Fourier transform method shown towards the end of lecture 2 video, apply linear filtering to
each image for 𝜎 = 2, 𝜎 = 4, and 𝜎 = 15. Create 3 new versions of the figure in part 1a, one figure for
each sigma. Show the middle slice of the filtered image in all subplots.
Submission: Put all your figures in your pdf, along with the code used to generate them and any
comments you have about the work.
Bonus +5%:
Make a python class that can display 3d and 4d images (scroll through the slices and time points), similar
to AFNI’s method for time series display (see lecture 1 at time 58:12).