Abhijit Adhikary

PhD Fellow in EPSRC Center for Doctoral Training (CDT) in Smart Medical Imaging at King's College London
Deep Learning | GAN | Computer Vision
image

I am a PhD Candidate in Biomedical Engineering and Imaging Sciences. My research focuses on applying machine learning models to computational cardiac models to extract biomarkers from medical images for disease stratification. More specifically, I am studying the role of the cardiac right ventricle for risk assessment and intervention planning for patients going through mitral valve replacement. I am jointly funded by the EPSRC Center for Doctoral Training (CDT) in Smart Medical Imaging at King's College London and Philips Research I have a Masters's degree in Machine Learning and Computer Vision from the Australian National University and an MRes degree from King's College London focusing on Medical Artificial Intelligence. I specialize in Deep Learning, specifically in Generative Adversarial Networks (GAN). I have interests in various Computer Vision techniques, i.e. Human Pose Estimation, Image Style Transfer and Human Centered Computing. I also have strong communication, management, leadership skills.

Education

King's College London, UK

Research Topic: The right ventricle’s role in risk prediction following mitral valve replacement: a combined imaging-modelling study.
Supervisor 1: Dr Adelaide De Vecchi, Lecturer in Computational Cardiovascular Modelling, School of Biomedical Engineering and Imaging Sciences, King's College London
Supervisor 2: Professor Pablo Lamata, Wellcome Senior Research Fellow, School of Biomedical Engineering and Imaging Sciences, King's College London
Clinical Champion: Professor Ronak Rajani, Professor of Cardiovascular Imaging, King's College London; consultant cardiologist in heart valve disease, Guy's and St Thomas' NHS Foundation Trust .
Industrial Supervisor 1: Dr Valentina Lavezzo, Team Lead Biophysics models of patients presso, Philips Research
Industrial Supervisor 2: Dr Caroline Balemans, Scientist Computational Biophysics - Philips Research

King's College London, UK

Supervisor 1: Dr Adelaide De Vecchi, Lecturer in Computational Cardiovascular Modelling, School of Biomedical Engineering and Imaging Sciences, King's College London
Supervisor 2: Professor Pablo Lamata, Wellcome Senior Research Fellow, School of Biomedical Engineering and Imaging Sciences, King's College London

The Australian National University, Canberra, ACT, Australia

GPA: 6.88/7.00 (Distinction)
Thesis: Privacy of Biometrics using Generative Adversarial Networks (GANs).
Supervisor: Professor Tom Gedeon, Professor of Computer Science and Head of the Human Centred Computing (HCC)
Details: I am using a conditional cycleGAN to analyze and filter out certain information (i.e. disease status, stimulus etc.) using a conditonal cycle consistent generative adversarial network.
Significant Courses: i) Neural Networks, Deep Learning and Bio-Inspired Computing, ii) Advanced Topics in Machine Learning, iii) Advanced Topics in Computer Vision, iv) Statistical Machine Learning, v) Advanced Topics in Mechatronics Systems (Convolutional Neural Networks), vi) Engineering Data Analytics, vii) Optimisation, viii) Document Analysis (Information Retrieval and Natural Language Processing)
Degree Certificate: View

Australian Centre for Robotic Vision, Canberra, ACT, Australia

Achievement: Best performing group in the Adelaide Node.
Certificate: View

BSc (Honours) Computer Science

09/2017 - 06/2019

Middlesex University, London, UK

Result: 1st Class Honours (Distinction)
Honours Project: Image to Image Translation Using Cycle Consistent Generative Adversarial Networks (CycleGAN).
Supervisor: Dr David Windridge, Professor of Data Science and Machine Learning
Details: I analysed the importance of various losses, i.e. cycle-consistency, identity etc. for Cycle-GAN training in the domain of unsupervised image style transfer.onsistent generative adversarial network.
Significant Modules: i) Artificial Intelligence, ii) Image Processing with MATLAB, iii) Industrial Networking
Appreciation Letter: View
Degree Certificate: View

North South University, Dhaka, Bangladesh

Status: Incomplete (Transferred to Middlesex University)

Work Experience

The Australian National University, Canberra, ACT, Australia

Supervised Courses: i) Neural Networks, Deep Learning and Bio-Inspired Computing, ii) Introduction to Machine Learning, iii) Computer Networks, iv) Software Engineering
Role: Conduct lab/tutorial sessions and grade students' homeworks, assignments and exams.

Middlesex University, London, UK

Supervised Modules: i) Web Applications and Databases, ii) Distributed Systems and Networking
Role: Conduct lab/tutorial sessions and grade students' homeworks, assignments and exams.
Year Book: View
Certificate: View

Publications

  1. Abhijit Adhikary, Namas Bhandari, Evan Markou and Siddharth Sachan. ArtGAN - Artwork Restoration Using enerative Adversarial Networks. In 2021 13th International Conference on Advanced Computational Intelligence (ICACI), pages 199-206, 2021
    Paper: https://doi.org/10.1109/ICACI52617.2021.9435888
    Code: https://github.com/namas191297/artgan
    More Details: Click here
  2. Abhijit Adhikary and Namas Bhandari. PosEmotion - Combining Real-Time 2D Body Pose Estimation and Facial Emotion Recognition to Analyze Human Behavior. In 2021 26th International Conference on Automation and Computing (ICAC21)
    Paper: https://doi.org/10.23919/ICAC50006.2021.9594155
    More Details: Click here
  3. Hao Wang and Abhijit Adhikary. StressNet: A Deep Neural Network Based on Dynamic Dropout Layers for Stress Recognition
    Paper: https://doi.org/10.1007/978-3-030-92270-2_43
  4. Siyuan Yan and Abhijit Adhikary. Stress Recognition in Thermal Videos Using Bi-directional Long-Term Recurrent Convolutional Neural Networks
    Paper: https://doi.org/10.1007/978-3-030-92270-2_42

Reviewer

Specializations

Achievements

  • EPSRC CDT PhD Fellowship Full scholarship towards tuition fees & monthly stipend. Funding Body: i) EPSRC Center for Doctoral Training (CDT in Smart Medical Imaging) at King's College London, ii) Phillips Healthcare, iii) King's College London
  • Junior Scholarship 2009 (National), Bangladesh
  • 2nd Place: 32nd National Science & Information Technology Fair 2011, Bangladesh

Voluntary Experience

Mentor

08/2021 - Present

Beta Tester

05/2021 - Present

News & Activities

  • Actively looking for part-time jons in Machine Learning/Deep Learning/Computer Vision.

Contact Me At

References

Available upon request.

Mentionable Presentations

Kernelizing a Support Vector Machine (SVM) - Primal vs Dual

This is a short presentation (8 minutes) I did for my Statistical Machine Learning course on how to kernelize a support vector machine.

Brief description: I tried to explain the formulation of SVM in a simple and elegant way while making use of as minimal maths and notation as possible (but sufficient). This should be sufficient to give the required intuition about SVMs to a person who has a basic understanding of how it works, but does not want to dive deep into the maths behind the actual formulation. I tried to present the differences between the soft and hard margin SVM and also how to obtain the Dual form from the Primal form. Finally, I presented when one variant would be beneficial over the other.

The images which I used from the internet are referenced at the bottom of each slide. I obtained a number of equations and formulations from the lecture slides of the Statistical Machine Learning course by the University of Tübingen. It can be accessed from:
https://www.tml.cs.uni-tuebingen.de/teaching/2020_statistical_learning/downloads_free/luxburg_statistical_learning_slides.pdf


Mentionable Projects

Privacy of Biometrics using Generative Adversarial Networks (GAN)

For my masters research project I am training a Conditional-GAN to filter out selective data i.e. identity, disease information etc. from EEG data. The raw EEG data is first preprocessed and converted to a special RGB image to facilitate CNN training. Although the project is in an intermediate stage, the results are already very promising.
Code: View on Github

ArtGAN : Artwork Restoration using Generative Adversarial Networks

ArtGAN introduces a novel generator architecture and follows the Pix2Pix architecture of the Conditional GAN for damaged artwork restoration. The generator consists of multiple stacked encoder-decoder blocks with skip connections between both intra and inter blocks. An efficient usage of 1x1 convolutions allowed us to build a very deep network while restricting the parameter count to 7.5 millions while achieving comparable performance to U-Net.
Code: View on Github

ArtGAN Network Structure Fig: ArtGAN Network Structure

CResNet Block Structure Fig: CResNet Block Structure

Sample Output 1 Fig: Sample Output 1

Sample Output 2 Fig: Sample Output 2

Sample Output 3 Fig: Sample Output 3

Sample Output 4 Fig: Sample Output 4


PosEmotion - Combining Real-Time 2D Body Pose Estimation and Facial Emotion Recognition to Analyze Human Behavior

The model first Jointly estimates the human pose using cascaded part confidence heatmaps and the activity of the person. Afterwards, the the facial region is extracted from the detected keypoints and the facial emotion as well as the facial keypoints are detected. The model can make accurate realtime predictions with 20-25 frames per second.
Code: View on Github

PosEmotion Network Pipeline Fig: PosEmotion Network Pipeline



Analysis of Network Pruning with Variable Finetuned Layers on a Pretrained AlexNet Model

Although a pretrained neural network provides a better starting point in most cases, choosing the number of layers to fine tune is a challenge, specially when the dataset is small. I observed the role of fine tuning difierent number of pretrained layers of the AlexNet model and concludes the ideal number of layers for small datasets.
Code: View on Github

Test ccuracy over 5 runs when different numbers of layers are frozen Fig: Test ccuracy over 5 runs when different numbers of layers are frozen
Test ccuracy of the network as different number of layers are frozen. (Blue – Actual values, Red – Mean values) Fig: Test ccuracy of the network as different number of layers are frozen. (Blue – Actual values, Red – Mean values)
The networks performance when different angular separation degrees are used Fig: The networks performance when different angular separation degrees are used


Image to Image Translation Using Cycle Consistent Generative Adversarial Network (Cycle-GAN)

For my undergraduate Honours research project I analysed the importance of various losses, i.e. cycle-consistency, identity etc. for Cycle-GAN training in the domain of unsupervised image style transfer.
Code: View on Github

Semantic Segmentation of Vehicle Dashcam Video Using Deep Convolutional Networks

This project utilized U-Nets to segment different regions of an image based on their semantic meaning. The model was trained on CamVid the car dashcam video and achieved an IOU of ~88\% and ran around 30 frames per second on a real-life setting.
Code: View on Github

Analysis of Street Image Segmentation Performance using Spectral Clustering via Nystrom Approximation

I analysed the speed and segmentation quality of the Nystrom approximation to perform spectral clustering on the IDD20k dataset, containing vehicle dashcam image sequences, and compared it with SOTA algorithms.
Code: View on Github

Transcutaneous Electrical Nerve Stimulation (TENS) for Hypertension Management

In partnership with Brun Health Ltd and Afferent Medical Solutions Ltd I undertook this medical product design project for a TENS product for hypertension management.