Internship Report(LVPEI)

This blog is about my internship experience in LV Prasad Eye Innovations. I am pursuing my 4th year at Bennett University. We were allowed to do an internship in our 8th semester, despite doing a project in the university, which enhances our skill and boosts our resume. I am happy that I got an internship in LVPEI, located in Hyderabad, Telangana. This is an online internship because of the cov-19 situation. It would have been great if I perform different functions using the device in the institution, as it is a new device.

I have worked on a project to structure a device called pupil+. My mentor's name is Divyank, who is very helpful to get the results accurately.

This is a device named pupil+, which detects the RAPD(Relative Afferent Pupillary Defect) in human eyes.

Let us get into a detailed explanation about the use of the device. Nowadays, ophthalmologists(eye doctors) use the only lens to observe a patient’s eye by focusing light on the person's two eyes, which gives results less accurately. To get accurate results of RAPD, the organization is designing this device.

The technique used by doctors to detect RAPD.

The eyeball contains iris in it. Iris contains dark-colored openings at the center of a person's eye, which allows light known as the pupil. When exposed to bright light pupil shrinks and expands to the maximum extent when the surroundings are dark. The movement of the pupil of both the eyes should be almost the same, if that differs it leads to Relative Afferent Pupillary Defect(RAPD). RAPD is present if the sum of absolute scores of amplitude and latency are >= 0.5. It is negative if the score is <=0.2 .

Coming to the work I need to do. This device is already designed but not brought into the market. Our teamwork is to test for different types of images in the data set and to implement an algorithm to get results accurately. After the result of every algorithm, it was taken to the doctor and a system check is performed to satisfy all the conditions. Basically, our work is to track the pupil properly and detect the amplitude and latency as much accurately as possible.

When we perform SFT for 2 cycles, amplitude and latency are detected accurately.

Step 1:The below shown image is an actual image from the data set we have. I have used certain dimensions to crop the image to our required level.

Image frame[269:795,537:1416]

Step 2: Converting BGR image(color image) to Grayscale image. We can undergo different functions only once we have the image in the binary frame, i.e, only black and white frames. The image shown below is changed into a grayscale.

Grayscale image

Step 3:Let us talk about different types of technologies used as we had changed the models two times to date. Initially, we have cropped the image to the required size using different image processing techniques. Using the ReLU layer we tried to shrink the image as per our requirement. But it did not go as per our expectations. Thereafter, using the Gaussian filter we have reduced the noise in all the images given as a data set. It worked well, i.e, we can go through the next steps.

The below is the image after converting the BGR image(color image) into a Grayscale image and using a 7*7 Gaussian filter to blur all the noise. Now we can only see the pupil part which will be highlighted in white.

All noises are erased and the pupil is displayed.

Step 4: This will the final step. Now we only require the pupil circle to be highlighted in the actual picture. For that, we can plot the contours for the image and mark the contours border with any of the 3 colors blue, green, or red. Below shown is the image detecting the pupil of an eye.

Final output.

So this concludes the task of detecting the pupil. This process is done to all the images present in the data set, separately for each individual's eyes. The task given to me as an intern is to crop all the images to the required size. Now, these are used to train. The output after using all the images and using the algorithm they already have is with good accuracy. They have recently tested this on a person. The below are the results obtained using pupil+.

As this is a continuous process, the results will be carried out until the LV Prasad Eye organization reach the expected accuracy. If the expected accuracy is met, the product will be brought into use.

Thanks to Bennett University CSE department for giving me the opportunity to do intern in the field I am interested in. Special thanks to Deepak Garg, HOD, CS Department; for encouraging us to do different projects outside the university.