SURG70004 Image Guided Intervention
IGI Coursework 1
Due date Thursday 11
th November 2021 at 2pm (via Blackboard)
During this exercise you will investigate the detection of edges and segmentation of skin lesions on
dermoscopic images. The image data used for this coursework is part of the 2018 International Skin
Imaging Collaboration (ISIC) Archive [1] and it includes skin lesion images (“Lesion_Image.jpg” files),
and a ground truth segmentation map for each image (“Ground_Truth_Segmentation.png” files) as
shown in the figure below. Before processing the images, resize them to 0.1 times the size of the
original images. The Matlab command B = imresize(A,0.1) can be used for this purpose.
Figure: Images included in the dataset (Left) Skin lesion image (Right) Ground truth segmentation map.
The questions are provided in black font. Please use blue for your answers in the spaces indicated.
Task 1 (15%)
A binary image contains straight lines oriented horizontally, vertically, and diagonally at 45°
and −45°
.
Give a set of 3x3 kernels which can be used to detect one pixel breaks in these lines. Assume that the
lines are one pixel thick, they do not intersect and that the intensities of the pixels belonging to the
lines and the background are 1 and 0, respectively.
(max. 100 words)
Answer here (expand as necessary):
Repeat the same task but this time propose a technique based on the 8-neighbour connectivity
analysis rather than using kernels for detecting the gaps.
(max. 100 words)
Answer here (expand as necessary):
Task 2 (30%)
The Laplacian of Gaussian (LoG) function:
can be approximated by the Difference of Gaussians (DoG):
Find the value of σ (as a function of 𝜎1 and 𝜎2) for which the LoG and DoG have the same zero crossings.
(max. 100 words)
Answer here (expand as necessary):
Use the Laplacian of Gaussian operator to detect the edges of the provided skin lesion image.
Investigate how the size of the LOG kernel affects the edge detection result.
(max. 100 words)
Answer here (expand as necessary):
Task 3 (30%)
Apply threshold-based segmentation to identify the lesion in the provided image. Explain your
threshold selection and investigate how the selection of the threshold affects the accuracy of the
segmentation result. You need to compare your results to the provided ground truth segmentation
map.
(max. 100 words)
Answer here (expand as necessary):
Use ROC analysis to select the optimal threshold value for your segmentation and display the best
segmentation map. Compare the segmented image with the edge detection result.
(max. 100 words)
Answer here (expand as necessary):
Task 4 (25%)
Corrupt the lesion image by Gaussian noise of zero mean and standard deviation of 0.1.
Apply image smoothing using 2D convolution between the lesion image and a gaussian kernel of
standard deviation 𝜎 = 1 and size [(2 ∗ 𝑐𝑒𝑖𝑙(2 ∗ 𝜎) + 1) × (2 ∗ 𝑐𝑒𝑖𝑙(2 ∗ 𝜎) + 1)] . Plot the
smoothed image. By presenting the result of filtering, show what is the effect of changing the value of
𝜎 on the smoothed image?
(max. 100 words)
Answer here (expand as necessary):
What is the computational cost (number of multiplications required)? Is it possible to make this 2D
convolution into multiple 1D convolutions? How (hint: check the associativity property of
convolution)? What would be the computational cost?
(max. 100 words)
Answer here (expand as necessary):
Apply the same filtering process in the frequency domain. Explain the steps you followed. Plot the
Fourier representation (magnitude) of the Gaussian filter and of the image.