About

My name is Van Vo. I am a Master student in Computer Science at University of Massachusetts Amherts.

Projects

Deep Learning for Estrogen Receptor Alpha enhancer analysis

Description:

Can deep learning help detecting enhancers contains MegaTrans? In this project, we investigate
if deep learning can classify different classes of estrogen receptor enhancers and identify
MegaTrans-containing enhancers. Furthermore, we want to know if DL models can identify useful DNA motifs.

Method:

- We train Transformer models on classification and regression tasks.
- In classification task, we train the models to classify between different classes of enhancers.
- In regression task, we train the model to predict the values of markers’ signals.
- We analyze which segments of DNA sequence is the most important for the models.

work done at Feinstein Institute for Medical Research

Bivariate Shrinkage-inspired Class Activation Map

Description of the project

Class Activation Map (CAM) is a popular class method to understand Convolutional Neural Networks prediction.
However, the extracted map computed at the final layers are coarse, while map from earlier layers can be noisy
but detailed. Inspired by Discrete Wavelet Transforms for denoising, in this project, we investigate whether
CAM-based methods show dependencies between layers, and can we improve CAM-based methods by exploiting this
dependencies.

Method:

- We investigate faithfulness of LayerCAM and GradCAM results at each layer on ImageNet
- We investigate failfulness of LayerCAM and GradCAM results when combined from different layers
- We adapt Bivariate Shrinkage idea from Discrete Wavelet Transform denoising to improve results

work done at Feinstein Institute for Medical Research

Deep Learning for Alzheimer's Disease Identification from PET images

Description:

In this project, we investigate whether Deep Learning models can predict cognitive assessment scores from
PET images. We analyze which brain regions are the most important for predicting. Our goal is to create a
score from PET images that can quantify the severity of cognitive ability.

Method:

- We trained 3D-CNN models on ADNI dataset on regression tasks
- We analyze which brain regions are the most important

work done at Feinstein Institute for Medical Research

Other

Email: vdvo@umass.edu