Research

Summary

My research centres on health informatics: developing AI-based, non-invasive disease screening methods that can work in low-resource settings. I combine medical image analysis, passive sensing from everyday smartphones, and graph-based deep learning, and I also apply deep learning to scientific and engineering data. My so far research activities can be summarized by the following keywords:

  • Health Informatics
  • Medical Image Analysis
  • Self-supervised Learning (SwAV)
  • Graph Neural Networks (GCN, Vision GNN, GAT)
  • Passive Sensing & Human Activity Recognition (smartphone tri-axial accelerometer)
  • Computer Vision Models (EfficientNet, ResNet, CNN encoder-decoder)
  • Scientific Machine Learning (3D temperature field reconstruction)
  • PyTorch for model development

> I'm open to collaborating on any AI/ML work that helps make the world a better place.

Health Informatics

AI for early, non-invasive and affordable disease screening

Undergraduate Thesis

DiabSense: early diagnosis of non-insulin-dependent diabetes mellitus using smartphone-based human activity recognition and diabetic retinopathy analysis with Graph Neural Network

Published in
Journal of Big Data Q1 · impact factor 10.8
DOI
10.1186/s40537-024-00959-w
Supervised by
Dr. Abdul Kadar Muhammad Masum
  • Objective: To address the global challenge of mitigating undiagnosed non-insulin-dependent diabetes mellitus (type II diabetes).
  • Developed, fine-tuned, and assessed a Graph Convolutional Network (Vision GNN) and a Convolutional Neural Network-based model (EfficientNet-B5) to grade the severity of Diabetic Retinopathy (DR) from fundus images.
  • Leveraged smartphone tri-axial accelerometer (passive sensing) data to perform Human Activity Recognition (HAR) with a Graph Attention Network (GAT).
  • Integrated HAR with DR grading to build a system - DiabSense, and validated the system's outcomes on four experimental subjects by performing clinical A1C tests.
  • My contributions: leading the entire research cycle - conceptualization, methodology design, data curation, implementation (PyTorch), visualization (Matplotlib, Plotly), manuscript preparation (LaTeX), and major revisions.
Overview of the DiabSense system architecture.
Figure: Overview of the DiabSense system architecture.
Graph Convolutional Network (ViG) overall architecture for DR grading.
Figure: Graph Convolutional Network - Vision GNN overall architecture for DR grading.
Graph Attention Network architecture for Human Activity Recognition.
Figure: Graph Attention Network architecture for Human Activity Recognition.
Snippets of the Android Application.
Figure: Activity data collection snippets from android application.
23 chosen activities correlated with diabetic symptoms
Figure: 23 chosen activities correlated with diabetic symptoms.
A continuous eleven hours sample of unlabelled triaxial accelerometer sensor data from experimental subject-1.
Figure: A continuous eleven hours sample of unlabelled triaxial accelerometer sensor data from experimental subject-1.
CollaborationMedical Imaging

SwAV-driven diagnostics: new perspectives on grading diabetic retinopathy from retinal photography

Published in
Frontiers in Robotics and AI Q2 · impact factor 3.7
DOI
10.3389/frobt.2024.1445565

Diabetic Retinopathy (DR) is a serious eye condition that occurs due to high blood sugar levels in patients with Diabetes Mellitus. If left untreated, DR can potentially result in blindness. Using automated neural network-based methods to grade DR shows potential for early detection. However, the uneven and non-quadrilateral forms of DR lesions provide difficulties for traditional Convolutional Neural Network (CNN)-based architectures. To address this challenge and explore a novel algorithm architecture, this work delves into the usage of contrasting cluster assignments in retinal fundus images with the Swapping Assignments between multiple Views (SwAV) algorithm for DR grading. An ablation study was made where SwAV outperformed other CNN and Transformer-based models, independently and in ensemble configurations with an accuracy of 87.00% despite having fewer parameters and layers. The proposed approach outperforms existing state-of-the-art models regarding classification metrics, complexity, and prediction time. The findings offer great potential for medical practitioners, allowing for more accurate diagnosis of DR and earlier treatments to avoid visual loss.

Framework for Diabetic Retinopathy severity grading with SwAV model.
Figure: Framework for Diabetic Retinopathy severity grading with SwAV model.

Scientific Machine Learning

Deep learning for physical and engineering data

CollaborationComputer Vision

High-Fidelity Reconstruction of 3D Temperature Fields Using Attention-Augmented CNN Autoencoders With Optimized Latent Space

Published in
IEEE Access Q1 · impact factor 4.2
DOI
10.1109/ACCESS.2024.3512873

Understanding and accurately predicting complex three-dimensional (3D) temperature distributions are critical in diverse domains, including climate science and industrial process optimization. This study presents a sophisticated framework employing a convolutional neural network (CNN)–based autoencoder (AE) architecture augmented with attention mechanisms for the efficient compression and reconstruction of 3D temperature distribution datasets. The framework integrates Singular Value Decomposition (SVD) analysis to ascertain the optimal latent space dimensionality, thereby ensuring a judicious balance between model complexity and reconstruction fidelity. Moreover, the autoencoder is trained using a customized loss function designed to prioritize higher temperature values, enhancing the reconstruction accuracy in critical regions—mathematically defined as regions where T > 675 °C. This ensures enhanced reconstruction accuracy in areas of significant thermal importance, which are critical for the accuracy of the model. Through systematic exploration of the latent space dimensionality and the relative weighting of non-zero temperature data points, optimal parameters are identified that maximize the coefficient of determination score. Empirical results indicate that optimal performance is achieved with a latent space size of six, incorporating a relative weight value of 4.5 for non-zero temperature data points and appropriate handling of zero-temperature data points. After evaluating the model for both zero and non-zero temperature data, the R2 scores improved from 95.80% to 99.27%, demonstrating a significant enhancement in overall accuracy. This proposed methodology provides profound insights into the intrinsic structure of the data and offers highly accurate predictions for applications necessitating detailed spatial and temporal temperature analyses.

Overall framework of compressing 3D volumetric temperature distribution image data into an optimal latent space size.
Figure: Overall framework of compressing 3D volumetric temperature distribution image data into an optimal latent space size.
(a) Temperature distribution for case:0 3D image; Cross-sectional temperature distribution along the (b) x-axis, (c) y-axis, and (d) z-axis for case:0 3D image.
Figure: (a) Temperature distribution for case:0 3D image; Cross-sectional temperature distribution along the (b) x-axis, (c) y-axis, and (d) z-axis for case:0 3D image.