Embedded Files
Cancer Prediction of Genes using Graph Neural Networks
Cancer is a very complex disease which involves interactions between multiple genes, aberrant expression of regulatory genes (e.g. lncRNAs), etc. Recent popular approaches in predicting gene-disease association utilize heterogeneous data sources and network analysis. This project use heterogeneous protein-protein, protein-gene and gene-gene interaction data and use graph neural network models to predict cancer from genes. https://gitlab.com/libphy/cancernet (disclaimer: as a short-term personal project done in a week, the codes are not super tidy, but can show some ideas on how GNNs can be used in cancer prediction tasks and data sources)
Brain Tumor Segmentation using Deep Unsupervised Learning
Gliomas are the most common brain tumors in adults. Depending on their aggressiveness class (high-grade vs. low-grade), the median survival rate of a patient may be shorter or longer than 2 years, thus different treatment plans are used. For accurate diagnosis and corresponding treatment plans, a patient undergoes extensive neuroimaging, and doctors and radiologists perform image segmentation to figure out the tumor substructure and their volume and location in the brain which are important information for diagnosis and treatment plans. As the image segmentation (labeling each pixel of the MRI image) is time consuming and costly process, automated image segmentation would be a great help. We develop automated brain MRI segmentation algorithm using deep unsupervised learning, which is expected to have a better performance and require less labeled data for training than supervised deep learning approaches. https://doi.org/10.1007/978-3-031-09002-8_22
Predicting Concrete Mechanical Strength using Machine Learning
Every batch of concrete used in a construction site is required to pass mechanical strength test due to safety and quality assurance purpose. Current concrete strength test procedure is to send a sample from a concrete batch to a lab and compress to destroy the sample to measure the strength. The whole process takes about a week and it is costly. We develop a machine learning and deep unsupervised learning algorithm to determine the mechanical strength of concrete from the images of the concrete to reduce the time and cost.
aMazeGNN: A Maze clustering GNN
In robotic visual navigation problem, since the search space on state and action can be very large, a naive reinforcement learning approach does not always work well when navigating in a complex environment. By an abstraction of the environment to an abstraction graph, and then searching for a path in the abstraction graph can help navigating in a complex environment. This project is about using graph neural network to cluster on maze like grids to coarsen the graph and see if our algorithm can group similar room-like structures together. [Project git]
Predicting Fractional Flow Reserve in Coronary Arteries using Machine Learning
Heart disease patients develops plaque in their coronary arteries. If the blood pressure ratio between before and after the lesion (fractional flow reserve- FFR) is below certain level, a patient has a high chance of heart attack, thus a STENT placement surgery is required. Measuring the pressure in coronary arteries is and invasive process (with some health risk to the patients) and it incurs additional cost. We develop a deep learning algorithm to detect lesion in ultrasound images and used machine learning to predict FFR directly from the images to avoid invasive measurement. Published in ISBN: 978-3-030-01364-6
Brain Tumor Segmentation using Curriculum Learning
Curriculum learning is a staged training strategy to make the deep learning model learn efficiently, just as how students learn complex concepts by following a curriculum. We improve automated brain tumor segmentation algorithm using curriculum learning.
Brain Tumor Segmentation using Deep Learning
We developed automated brain tumor segmentation algorithm in MRIs using various modified U-Net. Published DOI: 10.1007/978-3-319-75238-9_30 [Docker]
Creating Bird Sounds from Images using Deep Generative Models
Children can learn a concept by association between different sensory information. For example, a child can learn that a dog can bark by associating the sounds and the image of a dog. We study different approaches in deep generative models (GANs and complex recurrent models using attention and gating) to create bird sounds from bird images.
Page updated
Google Sites
Report abuse