Arna Ghosh

Brain age estimation using structural and functional neuroimaging data

McGill University, Montreal

November 2018 - Present

• A ML framework to predict brain age of healthy subjects from MRI and MEG recordings.
• Identified relevant subcortical structures for age prediction using a CCA-based approach.
• Used Graph Convolutional networks to identify age-dependent changes in functional brain networks.

Brexting - Brain Texting

McGill University, Montreal

November 2017 - August 2018

• Built a DL framework on an Intel FPGA board for decoding imagined motor movements from EEG recordings to enable real-time typing.
• Implemented the trained CNN on FPGA to achieve hardware acceleration and enable real-time use.
• Won the Silver award in regional finals and Iron award in grand finals of Intel Innovate FPGA 2018 challenge.

Convolutional Neural Networks for motor task EEG analysis

McGill University, Montreal

September 2017 - June 2019

• Built a Convolutional Neural network-based framework in Torch to identify exercise-induced EEG signatures.
• Developed a technique called ccCAM (cue-combination for Class Activation Maps) to visualize the features learnt by CNN.
• Introduced an adversarial training strategy to improve population-level generalization of deep learning methods in limited sample setting.

• Code
• Paper

Deep Semantic Architectures for Mitotic Figure Detection in Breast Histopathological Images

IIT Kharagpur

July 2016 - May 2017

• Used Torch to implement deep learning algorithms to count number of mitotic figures in breast histopathological images.
• Compared color decomposition techniques to segment nuclei from image.
• Implemented localisation followed by classification of mitotic nuclei in an image. Used this to successfully detect leukemic WBCs.
• Awarded the systems society best Bachelor thesis award.

• Mitotis detection code
• ALL detection code
• Paper

Mechanism Underlying sensorimotor networks

McGill University, Montreal

May-July 2016

• Selected as a MITACS Globalink intern to work on the project for 12 weeks.
• Designed the software & hardware-software interface for different experiments to study motor learning characteristics
• Interfaced EEG, EMG and fMRI setups with hardware triggers to align data collected by the different equipment in time
• Developed automated analysis software for data collected from tDCS experiment to detect learning curves in different accuracy parameters
Code

Slide-Scanning Microscopy on a Smart Phone by High-Frame-Rate Video Capture

University of British Columbia, Kelowna, British Columbia

May-July 2015

• Developed image processing module to extract microscopic images from video, sharpen and remove dirt.
• Worked on fast deblurring of images and developed the hardware-software interface using Arduino.
• Developed a preliminary IOS app to take image at high frame rate using OpenCV and modified it to allow 240 fps video capture.
Code

IndicView

IIT Kharagpur

October 2014 - December 2015

• Funded by Google under the Google-IIT Pilot program and accepted to be integrated into Google’s applications.
• Implemented Sauvola binarisation algorithm to threshold scanned images and segregate into text and background.
• Developed the algorithm to detect words from scanned image of a document written in an Indian language.
Learn more

Autonomous Ground Vehicle Research Group

IIT Kharagpur

March 2014 - April 2017

• Developed image processing module to detect obstacles based on color, lane detection and construct world map by sensor fusion.
• Worked on efficient road segmentation for off-road conditions and visual odometry problem using CNN. Paper
• The team qualified for the prototype round of Mahindra Rise Prize - Driverless Car Challenge .
• The team qualified 2/3 rounds of Auto-navigation challenge in IGVC 2015 at Oakland University in Rochester, Michigan.
Learn more

Deep Learning for Computer Vision

Worked on biologically-plausible algorithms for computer vision. Most of the projects are inspired from the mechanisms of the human eye and the knowledge of perception in human beings. The titles of the projects and conferences where they have been accepted are mentioned below.

• Generative adversarial Motion to Image Synthesis. ECCV Workshop on Biological and Artificial Vision 2016
• Handwriting Profiling using Generative Adversarial Networks. AAAI Student Abstract 2017
• SAD-GAN: Synthetic Autonomous Driving using Generative Adversarial Networks. NIPS 2016: Deep Learning for Action and Interaction
• AdGAP: Advanced Global Average Pooling for Simultaneous Localisation and Classification