UNet Ensemble Learning

Scored 93% in advanced machine learning course where I used 3D U-Net and ensemble learning to segment lung tumors in MRI images (Medical Segmentation Decathlon dataset). To do this, I focused on optimization of hyperparameters, data augmentation, and ensembling methods

GitHub Repository

Multi-Modal Phenotyping for Tuberculous Meningitis

Developed, tested and refined multi-modal models designed to phenotype brain lesions in patients with Tuberculous Meningitis and HIV, and predict their clinical outcomes. Publication is forthcoming where my hope is that this research will eventually improve medical care and save lives

Surgical Path Planning with Robotic Operating System (ROS)

ROS for Surgical Path Planning. The project involves: 1. compute the trajectory that acheives critical structure avoidance, 2. minimizes the distance to the target 3. move the robot arm along the trajectory

GitHub Repository

VAE-GAN

Image example for the VAE-GAN project

VAE-GAN

The aim of this project is to implement and validate a VAE-GAN as per the original paper by ABL Larsen et al. (2015) and to apply it to X-ray images of hands

GitHub Repository

Predicting gestational age at scan and birth for neonates from the developing Human Connectome Project

Predict the gestational age at scan and birth for neonates using data from the developing Human Connectome Project. A pre-activation ResNet architecture (He, 2015) is used, integrating confounds to enhance prediction performance. The training process included data augmentation, weighted sampling to manage class imbalances, and early stopping for training

GitHub Repository

Classification of prematurity using structural connectivity using PCA

Predict neonatal prematurity using brain connectivity matrices as features. Initial logistic regression models showed high specificity but low sensitivity due to dataset imbalance; however, adjusting for class weights improved the model's performance in predicting preterm births. Visualization using PCA demonstrated the separation between term and preterm samples, both for true and predicted labels, with the updated model showing better alignment with the actual data

GitHub Repository

MindSet

Contributed to development and testing of datasets to determine whether neural networks exhibit visual illusions (MindSet). This will enable the understanding of the difference and similarity between human vision and artificial system.

GitHub Repository

Publication

Measuring Perceptual Similarity in Neural Networks

A schematics animation made using manim, for MindSet project and the Orthographic DNN project. Click play to watch the animation

Orthographic Deep Neural Networks

Convolutional Neural Networks accounts for priming effect surprisingly well - better than dedicated psychological models

GitHub Repository

Publication

LLMs for Healthcare Records (Details are Confidential)

Current electronic healthcare record (EHR) systems encounter several issues, notably concerning accessibility and mobility of data, and the effective use of emerging technologies. A promising solution to these challenges is the application LLMs

Tensor Spectacle

A python package for visualizing a Sequence of Tensors

GitHub Repository

Clustering-Based Approximation for the Travelling Salesman Problem (TSP)

A hierarchical clustering-based approximation that proves to be significantly faster than genetic algorithm, especially with larger datasets - 974 times better solution (path length) and 1.22 times less computation time

GitHub Repository

Generative Model Art Plotter Pipeline

Using generative model for producing art, and reinforcement learning agent and a plotter for drawing it

Gesture Code

A project that aims to create simple, pythonic and intuitive syntax for gesture recognition

GitHub Repository