Anwesha Mohanty

I am a final year PhD Candidate at Dublin City University (DCU) as a part of CRT d-real programme funded by Science Foundation Ireland. I am advised by Dr. Alistair Sutherland, Dr. Marija Bezbradica and Dr. Hossein Javidnia.

In the summer of 2022, I was a Research Intern in the innovation branch of the digital analytical and research team within NHS England (NHSx).

My PhD is focused on studying a particular skin disease, i.e. Rosacea analysis with limited data using Computer Vision and Machine Learning (/Deep Learning). The main objective of my study is to attain the optimum outcome in computer-assisted disease diagnosis, mainly when the quantity of available images for the study is limited. The central gap in advancing technology and research between the medical domain and computing is primarily attributed to the scarcity of data. I am studying the intricacies of the mathematical and engineering aspects of Deep Learning models to uncover ways to extract maximum benefit from a minimal amount of data and thus enhance the efficacy of the models and future computer-aided medical diagnosis.

Email  |  Curriculum vitae  |  Google Scholar  |  Github  |  LinkedIn

I am particularly interested in anything that involves generative models, aka Generative AI. Additionally, I am intrigued by the existing/non-existent possibilities around training complex ML/DL models with small data (which may include a subset of various publicly accessible Big Datasets).

I am on the 2023 Academic and Industry(R&D) Research Job market!!

• Delivered an invited talk on the topic: "Getting the most out of a little data" at the University of Manchester[February 2023].
• Delivered a Guest Lecture on the topic: "Computer Vision, Deep Learning and 'a little data' " to the final year BSc. in the School of Computing, DCU [January 2023].
• Delivered a Guest Lecture on the topic: "Application of Statistics" to the BSc. students in the School of Computing, DCU [November 2022].
• Completed my Research Internship on Explainable AI with NHS England (NHSx)[May 2022 - July 2022].
• Won the best student poster presentation in 33rd Irish Signals and Systems Conference 2022.
• Our paper "Towards Synthetic Generation of Clinical Rosacea Images with GAN Models" has been accepted at the 33rd Irish Signals and Systems Conference 2022.
• Our paper "Skin Disease Analysis With Limited Data in Particular Rosacea: A Review and Recommended Framework" has been published on IEEE Access.
• Oral Presentation on my PhD topic in the Special Open session in the DeepLearn2019.

In this study, for the first time, a small dataset of Rosacea with 300 full-face images is utilized to further investigate the possibility of generating synthetic data, utilizing StyleGAN2. The study scrutinizes the impact of model fine-tuning and experimental setting variation on feature fidelity, highlighting the role of R1 Regularization. Evaluations by the expert dermatologists and lay participants are incorporated, alongside a thorough quantitative validations. The discussion concludes with potential limitations and future trajectories in synthetic image generation with clinical data.

In this study, for the first time, developed a novel 3D parametric modeling technique to generate synthetic Rhinophyma images, addressing data scarcity. Generated 2000 possible unique deformations of Rhinophyma nose and rendered 20,000 synthetic images to train deep learning classification models. Released the synthetic dataset and 3D models on an open-source platform to aid further research and healthcare applications. Achieved a 95 % recall (sensitivity) rate in classification on 220 real-world Rhinophyma images using deep learning models trained on synthetic data. Validated the model's performance with extensive metrics such as accuracy, precision, F1 score, confusion matrix, GradCAM visualisation, demonstrating its potential for accurate classification.

In this study, for the first time, a small dataset of Rosacea with 300 full-face images is utilized to further investigate the possibility of generating synthetic data, utilizing StyleGAN2. The study scrutinizes the impact of model fine-tuning and experimental setting variation on feature fidelity, highlighting the role of R1 Regularization. Evaluations by the expert dermatologists and lay participants are incorporated, alongside a thorough quantitative validations. The discussion concludes with potential limitations and future trajectories in synthetic image generation with clinical data.

Considering state of-the-art StyleGAN2 with Limited Data, the purpose of this paper is to provide an attempt towards the generation of synthetic faces with Rosacea skin condition with a small dataset.

This study considers one of the key challenges in data acquisition and computation, viz. data scarcity. With data scarcity in mind, the possible techniques explored and discussed include Data Augmentation, Transfer Learning, Generative Adversarial Networks(GANs), Meta-Learning, Few-Shot classification, and 3D face modelling.