Projects
Current and past projects that I have been a part. Click on each project to know more
Research Projects
Replay Learning for Multi Site Prostate Segmentation
Computer Vison Lab and Medical Imaging & Theranostics Lab @ IIT Kharagpur
May 2022 - Current
This work focuses on a deep neural network that is able to learn from multiple datasets, with the goal of being able to quickly adapt as new data arrives over time. The challenge with this type of learning is that it can be forgetful, meaning it loses some performance when switching tasks. To address this problem, replay-based continual learning methods have been developed that store experiences or representations from previously learned domains in a memory buffer. Our goal was to show that we could achieve competitive performance by employing effective sampling strategies for the replay buffer, even when using less than 5% of samples. We compared our method against other replay-based continual learning methods and showed superior performance across four prostate datasets [Promise12, ISBI, Decathlon, Prostate158].
Learning to Forecast for Endangered Water Bodies using Satellite Imagery
Independent Research
Dec 2022 - Current
This work focuses on forecasting water cover in endangered lakes and ponds a few weeks ahead, using satellite imagery. Such bodies of water are important ecosystems that provide both habitats for wildlife, and resources for human communities. As a result of climate change, as well as human activities, many such places are drying up worldwide; To address this issue, we introduce a dataset comprising satellite imagery from multiple endangered water bodies worldwide. As this data was insufficient to train a deep neural network directly, we made use of large synthetically generated linear and non-linear shapes in order to pre-train it for shrinking or expanding transformations; our model is thus a ConvLSTM network which is trained in an self-supervised manner will help predicting future water coverage accurately. This seemingly simple task could prove invaluable to climate activists and policy makers alike by providing them with critical insights into how these ecosystems might look weeks down the line.
Agriculture Application Robot
Robotics Lab @ RGUKT RKV and TechInAg
December 2019 - Current
As a part of my undergradute research, I've been working on agriculture robotics which later turned into a funded startup. I am in the forefront while designing and developing this project. AAR is a semi-autonomous robot which can travrse crop fields for disease identification, weed removal, watering plants and collecting critical soil & weather information for analytics.
Collaborators: Althaf Shaik, Aravind Gundlapalle
Minor/Class Projects
AI based Tele-Pathology
Tele Medicine Lab @ IIT Kharagpur
Jan 2022 - Current
As part of tele-medicine laboratory which involves design and devolpment of tele-medicine application, we choose to develop and deploy a AI-based Tele-Pathology application. This project involves devloping a dynamic website and deep learning for analyzing patient pathological images.
Collaborators: Sai Pavan, Abhilash
Diabetic Retinopathy Detection
IIT Kharagpur
Mar 2022 - Apr 2022
As part of Medical Image Analysis course, we choose work diabetic retinopathy detection which is a growing cause of blindness. We have worked with feature extraction and classification techniques to detect diabetic retinopathy. Model interpretation using saliency maps and integrated gradients is also a part of the project.
- Image extracted features(Morphological + Texture) and a classifier
- Image extracted features(using Inception V3) and ML classifiers
- End-to-end deep Learning based DR recognition
- Model interpretation using saliency maps and integrated gradients
Teammates: Sneha Singh( PhD Student)
Image-to-Image Translation with Conditional Adversial Networks
IIT Kharagpur
November 2021
As a part of my course work I've chosen Neural Networks and Application in Autumn 2021, taught by Prof Dedhasis Sen. As a part of my course capstone project I've re-implemented original paper Image-to-Image Translation with Conditional Adversial Networks in PyTorch. I've used P5000 16GB GPU from paperspace cloud for training the Networks.
GitHub Link : github.com/ummadiviany/Pix2Pix
Paper Link : Image-to-Image Translation with Conditional Adversial Networks
Malarial Image Classification
IIT Kharagpur
November 2021
Digital Image Processing and Applications taught by Prof Debdoot Sheet is a depth course for my department. As part of my course work, Professor asssigned me Malaria Image Classification as my term project. The restriction here is to not to use any of machine learning techniques and must be done with purely classical image processing techniques that were taught in the class.
GitHub Link : github.com/ummadiviany/Malaria-Image-Classification
Matlab File Exchange Link : Malaria Image Classification