Research Assistant in the Medical Image Processing (MIG) lab headed by Prof.Phaneendra Yalavarthy of the Computational Data Science department at IISc, Bangalore. My research was in collaboration with the Royal Dutch Shell Technology Centre, Bangalore - India

Research Overview

The MIG lab led by Prof.Phaneendra Yalavarthy mainly focusses on deep learning and computational methods in medical imaging, medical image processing (reconstruction/analysis), physiological signal processing, photoacoustic imaging, and diffuse optical imaging and closely related works.

The Digital Rock and Image Analysis group in Shell Technology Centre, Bangalore develops and uses image processing techniques to study the microstructure of porous media and build physics models to accelerate research and development cycles for Shell businesses. Examples of the porous media systems include 2-D or 3-D images of rocks and catalyst pellets. They have active projects to develop capability in areas of image classification, segmentation, physics based modelling and image reconstruction for different imaging modalities.

Digital rock involves scanning three-dimensional rock volumes using micro-CT scanners and using these volumes to estimate petrophysical parameters. Accurate estimation of these petrophysical parameters relies on acquiring high-resolution reconstruction volumes. Acquiring high-resolution reconstruction volumes is time consuming and challenging to cover the entire field of view. Therefore, there is a major emphasis on developing computational technique to enhance the resolution characteristics of the rock volumes. Being a part of the novel project, I aim to develop a deep learning based approach to super resolve the 2D low-resolution digital rock images and reconstruct a final super-resolved 3D volume, and further develop a 3D model which directly super-resolves and segments a 3D low resolution volume.

My Contribution

I have been fortunate to be mentored by Prof. Phaneendra Yalavarthy from IISc Bangalore and Dr. Vishal Ahuja from Shell Technology Centre Bangalore

• Developed a Generative Adversarial Network (GAN)-based novel approach, Siamese-SR for the above problem statement and generated super resolved volumes with their texture and porosity values closer to those of high resolution volume.
• Regularly interacted with a team of geologists at Shell and worked on their suggestions
• Siamese-SR model was capable of improving the resolution by a factor of 2, whilst preserving the texture and providing optimal denoising.
• Siamese-SR approach was found to generate petrophysical parameters close to the acquired HR image volumes compared to other state-of-the-art super-resolution techniques like SRGAN, ESRGAN.
• Worked on developing a 3D Convolutional Neural Network (CNN) based model like 3D UNet, VNet, ResNext to simultaneously super-resolve and segment the 3D low-resolution digital rock volume.