Copyright 2021 © University of Southampton Page 1 of 4
UNIVERSITY OF SOUTHAMPTON ELEC6213W1


SEMESTER 2 FURTHER ASSESSMENT 2020-21

IMAGE PROCESSING





This paper contains 1 question with 8 sub questions


Answer All sub questions.



An outline marking scheme is shown in brackets to the right of each
question.
















ELEC6213W1


Copyright 2021 © University of Southampton Page 2 of 4

Hypoxic Ischemic Encephalopathy (HIE) is a disease caused by the
shortage of oxygen during birth. In order to check if a new born baby is
affected by HIE, MRI scans of the brain are taken after the birth. Figure
(1) shows MRI scans of four patients affected by HIE. This figure shows
the MRI scans of the affected infant brains in the order of HIE severity by
depicting a healthy brain in figure (1.A), the least affected brain in figure
(1.B) and finally the most affected brain in figure (1.D). As observed from
these MR images, brain vessels become more distinctive and apparent
in MRI scans as HIE becomes more severe. We have a small dataset of
these MRI scans and you are tasked to analyse these images.


Figure 1

(i) Advance an algorithm to detect vessels in these MRI scans.

[10 marks]

(ii) In the presence of noise, explain what you can do to minimise the
effects of noise in these images
[5 marks]


ELEC6213W1


Copyright 2021 © University of Southampton Page 3 of 4


(iii) Propose an algorithm to detect the vertical central line separating
the two halves of the brain and remove that line in order to make
sure that the vertical central line is not considered as a vessel in
the feature extraction step. Explain how your method works.

[10 marks]

(iv) Propose features which can be extracted from these images to
classify these images in four groups. (i.e., Group A: Healthy brains,
Groups B, mildly HIE brains, Group C: Moderate HIE brains, and
Group D: Severe HIE brains)
[15 marks]

(v) Explain what the accuracy is for each group, if a classifier (random
classifier) randomly assigns an MRI scan to a group (assume each
group contains 10 patients). However our dataset is not balanced.
This means that there are 100 healthy brains in group A and
groups B, C and D each have 10 HIE brains. Find the accuracy of
a random classifier for this unbalanced dataset. Describe a
scenario where this unbalance dataset can produce a total
accuracy of more than 76% for even a very bad classifier which
assigns all images to only one of the groups.
[15 marks]

(vi) Propose a classification algorithm by using the features you
extracted in question (iii) to classify these images into these four
groups. Explain how you resolve the issue of the unbalanced
dataset.
[20 marks]

(vii) Explain how you set up your experiments in order to have a mean
and variance for the classification accuracies in each group.

[7 marks]


ELEC6213W1


Copyright 2021 © University of Southampton Page 4 of 4


(viii) If our dataset is very large and contains more than 4000 patients in
each group, then describe an alternative approach for this
classification task and explain why you would take such an
approach for this large dataset instead of implementing the
aforementioned steps?
[18 marks]








END of PAPER








欢迎咨询51作业君