EE $417$ Computer Vision $-$ Spring $2023-24$

Project $($optional$)$

This assignment is optional, its submission is not mandatory. In case it is however submitted

before the deadline, it will be evaluated and its grade will contribute to your overall course grade.

The premise: this project involves breast cancer. The task is to develop an accurate and rapid

software solution for detecting mitosis $($cell division$)$ in a given image. By counting the mitodic cells

within a unit surface, medical professionals determine the aggressiveness level of cancer, and

plan the treatment accordingly. This is normally done by hand $($i$.$e$.$ counting them one by one

under a microscope$),$ and takes an expert$$s valuable time. We$$d like to automate this process.

The data: you are provided with mitosis.zip which contains $5$ folders. Each folder contains the data

of one breast cancer patient. The $*.$bmp files are color images of tumor biopsies acquired through

microscope. The $*.$csv files contain the coordinates of each mitosis pixel in the bmp file

accompanying it$.$ The $*.$jpg files are the same as the $.$bmp files, except that the mitosis cases are

painted yellow to highlight them.

The task:

a$)$ Develop a learning based computer vision solution $($deep or shallow that$$s up to you$)$ to

accomplish this task. In order to evaluate it$,$ use $5-$fold cross validation across the $5$ patients and

report the mean performances. Use precision, recall and F$1-$score as performance metrics. You

are free to use additional datasets if you need to$.$ You must develop your own approach. Don$$t just

download somebody else$$s github repo for this, or you will not only get zero from this project but

your course grade will be affected negatively as well. $(40$ points$)$

b$)$ Compare your solution against existing methods in the state of the art $($at least $3)$ by

downloading and using them with your data. $(20$ points$)$

c$)$ Implement a variety of distinct augmentation methods $($at least $8)$ to augment your training data,

also use test$-$time$-$augmentation. How does it they affect performance? $(30$ points$)$

d$)$ Investigate single$-$domain generalization and implement it for this context. How does it affect

performance? $(10$ points$)$

Submission: submit your python notebook, and your report $($pdf$)$ elaborating your method and your

findings, prepared in latex using the IEEE conference format.

For questions, use the sucourse forum.

Good luck.