Mathematical and Computational Engineering Stream

Development of efficient deep learning models for medical image processing

Faculty: Phaneendra K. Yalavarthy (CDS)  Ambedkar Dukkipati (CSA)

Cost Sensitive Mini-Convolutional Neural Network (ENet+) for Segmentation of the Abnormalities in COVID-19 Chest Computed Tomography Images [Ref: https://github.com/NaveenPaluru/Segmentation-COVID-19]

Traditional medical image processing requires feature engineering to perform tasks at hand, like segmentation of abnormalities, enhancement of images, as well as classification. These tasks require deep understanding of imaging physics as well as image formation process to achieve reasonable performance. The deep learning models that purely data-driven, especially self-supervised or unsupervised, have been able to achieve these specific tasks learning these important features in automated fashion via convolutional neural networks. These deep learning models are often data hungry and have tens of millions of parameters to estimate in the network. As there is a lot of variability in the medical imaging protocols, often getting large datasets acquired under same imaging condition is not plausible.

This project aims to develop efficient deep learning methods that have atleast one order magnitude less parameters and are easy to train on small datasets without compromising the accuracy and performance. The deployment of these efficient deep learning methods for real-time applications will also be taken up as part of Ph.D. thesis topic to know the real-time applicability of developed models.

Background needed: Linear Algebra (and/or) signal processing.

Basic Qualifications: B.E./B.Tech. in EE/ECE/IN/CS/IT/BME (or) M.Sc. (Mathematics/Physics)