COL780: Computer Vision

Assignment-1: Automated Evaluation of Micro-sutures

Deadline: 5th F ebruary, 2024

1 Introduction
Micro-Suturing is a task of performing suturing (stitching) under a microscope. It is an indis-
pensable neurosurgical skill. Automated evaluation of micro-suturing would lead to efficient,
fast and unbiased training of a trainee neurosurgeon. We propose to evaluate the effectiveness
(final image) of micro-suturing using various parameters. In this assignment, you will calculate
three parameters of the evaluation using the classical computer vision algorithms. Namely,
(a) Number of sutures

(b) Inter-suture distance

(c) Angulation of the suture

(a) Image-1 (b) Image-2

Figure 1: Sample images of microsutures from the dataset. Image-1 shows a good micro-
suturing outcome, where as Image-2 is a relatively bad outcome.

2 Implementation Tasks
Various algorithms that have been discussed in the class can be implemented in order to quan-
titatively perform the below mentioned tasks for a given image. Some examples include Canny
edge detection, Harris corner detection, Contour hunting, etc. The tasks to be implemented
are as follows:

1
2.1 Task-1: Number of micro-sutures
Having a count on the number of micro-sutures in an image can prove to be an important metric
in assessing the quality of the surgery. For example, in Image-2, there are 9 micro-sutures in
total. Counting the number of sutures in a micro-suturing image involves applying computer
vision algorithms to detect and recognize individual sutures. One way to estimate number of
sutures is as follows:
• Preprocessing: Enhance the image quality by performing operations such as resizing,
contrast adjustment, and noise reduction to improve the effectiveness of subsequent al-
gorithms. An important pre-processing step for our application would be to perform
thresholding in order to effectively separate the background from the sutures.
• Edge Detection: Apply edge detection algorithms to highlight the boundaries of sutures.
Canny edge detection or other suitable edge detection techniques can be used to
identify the edges of individual sutures.
• Filtering and Refinement: Apply filters or criteria to distinguish sutures from other
object details in the image. This may involve filtering based on size, shape, or other
characteristics specific to the sutures in the micro-suturing image.
• Suture Counting: Once the relevant edges representing sutures are identified, we can
have a good estimate on the total number of sutures in the image. You can estimate suture
boundaries by considering the connected components of the edge pixels. You can use
a connected components algorithm or a region-growing approach to group neighboring
edge pixels into clusters, each representing a potential suture. Alternatively, you may use
Harris corner detection to count the number of corners and subsequently, number of
sutures can be found out by dividing the number of corners by a fixed constant.
Note: You may need to fine-tune parameters such as the Canny edge detection thresholds
and other settings based on the characteristics of the micro-suturing images. Additionally,
consider using color information or other features if grayscale edge detection is not sufficient
for your specific images. You are free to experiment with any other approaches as well. There
may be other approaches to estimate the sutures, for example, using contour detection. Mention
the approach that you have used along with some experimentation details in the detailed report
that needs to be submitted.

2.2 Task-2: Inter-Suture Spacing

In micro-suturing, suture spacing refers to the distance between individual sutures or stitches
made during a surgical procedure involving small or delicate tissues. The mean and variance
of inter-suture distance (relative to the height of image) are important metrics to measure
in order to assess the suturing done by the neurosurgical trainee.

After we have estimated sutures in Task-1, one way to estimate the inter-suture spacing
by calculating centroids of the sutures as follows: For each connected component, calculate
the centroid, and then compute the distances between consecutive centroids to measure the
inter-suture spacing.

Note: You should visualize the calculated distances and centroids on the original image.
This approach is based on the assumption that connected components obtained from the edges
from Canny detection reasonably represent the suture boundaries. In case you have followed
any other approach, mention appropriate experimentation details in the report.

2
2.3 Task-3: Angulation of the suture
Ideally, we would like have sutures parallel to the x-axis. However, in reality, different sutures
may be at different angles to the x-axis. Also, more the alignment amongst the sutures, better
the surgery. Hence, we should also consider the mean and variance of the angle formed
by various sutures and the x-axis in order to assess the performance of the trainee.

One way to find the angulation of the sutures and assess their alignment with the x-axis
is as follows: Once we have found the centroid for each suture in Task-2, we can calculate the
angle formed by the line connecting the centroid to a reference point (e.g., the leftmost
point in the image) and the x-axis.

2.4 Task-4: Comparison of two micro-suturing outcomes

Once we are able to extract the above features from the image, we can quantitatively compare
two micro-suturing outcomes on each feature, and conclude which image is better amongst the
two with respect to that feature. This will help us guide if our implementation is good enough
or not.

For this task, you will be given multiple pair of images, and for each pair, you need to output
which image is ”better” with respect to both features (inter-suture distance and angulation of
sutures) individually. Better image is one which has lower variance with respect to
that feature.

3 Dataset and Evaluation

• Dataset: We shall be using a subset of a large micro-suturing dataset (more information
regarding the dataset can be found here). You can find the dataset to experiment on
inside the data directory, wherein you have been provided with ten sample images to test
your implementation. We have a separate held-out test dataset (which would be similar
to the sample images given). We shall test the performance of your algorithms on that
dataset.

• Entire code must be submitted in the form of python files, along with a main.py which
will be executed in two parts during evaluation/demo as follows:

python3 main.py <part id> <img dir> <output csv>

There are two tasks:

– For the first part, you are required to iterate through all the images in the given
directory and generate a csv file specifying the various parameters that you have cal-
culated. It should have the following columns : image name, number of sutures,
mean inter suture spacing, variance of inter suture spacing, mean suture
angle wrt x-axis, variance of suture angle wrt x-axis. main.py will be
executed as follows:

python3 main.py 1 <img dir> <output csv>

3
 – For the second part, you will be given pairs of images and you need to output which
image is a better suture outcome with respect to inter-suture distance and angula-
tion of sutures. main.py will be executed as follows:

python3 main.py 2 <input csv> <output csv>

The input csv file has the following two columns: img1 path, img2 path, where
as output csv should have the following four columns: img1 path, img2 path,
output distance, output angle, where output distance and output angle take
the values either 1 or 2, depending upon which image is better with respect to that
feature.

A sample output.csv file will been provided later for this part which has ground
truth labels in the output column for some of all possible pairs of images in data
directory. Note that these have been manually labeled by experts.

4 Note
• You can use the inbuilt opencv functions to read and write the images and for visualization
purposes only. You CANNOT use any other inbuilt image processing functions unless
explicitly exempted in the description of the assignment.

• Entire code must be submitted in the form of python files, along with a main.py which
will be executed during evaluation/demo as explained earlier.

• Grading rubrik (out of 100 marks):

Task Marks
Task-1 40
Task-2 25
Task-3 15
Task-4 10
Report 10

Table 1: Evaluation Criteria

• You are required to submit the code, data, and a detailed report for this assignment and
zip the contents with file name ENTRY NUMBER.zip (For example: 2019CS50768.zip).
The submission can be done using Moodle. Any update in the submission instructions
will be shared later.

• This is an individual assignment. Mention the names of all the students you have discussed
with in the report. Any sort of plagiarism will lead to zero in the assignment straightaway
and stricter punishment depending on the case.