Senior Design Projects
ECS193 A/B Winter & Spring 2022
Identification of imaging features associated with disease using deep learning
| ********** | |
| Gerald Quon | |
| Genome Center/Molecular and Cellular Biology |
Project's details
| Identification of imaging features associated with disease using deep learning | |
| The human genetics community has identified many genetic mutations in our DNA that are associated with disease. What is unknown is what their actual effect is on human development. | |
| For a number of disease-causing mutations, we have images of animals with and without a human disease mutation, which can be compared to figure out if there are any differences between the images of animals with and without a mutation. The idea is, if no differences can be detected between animals with and without a mutation (e.g. an image-based classifier cannot distinguish images of the two groups), then that means the mutation might not cause any detectable changes in the animal. On the other hand, the bigger the effect of the mutation, the easier it should be to classify the two image groups. | |
| - An image-based classifier using deep learning. There have been many such systems proposed in the literature based on CNNs, Transformers and other architectures, that can be implemented here, or adapted from code that is sometimes published with these papers. Importantly, we typically have few labeled images (per mutation, ~80 per class), but have more unlabeled images (thousands). - What typically is not implemented along with image classifiers is a corresponding feature attribution method (such as Shapley) for neural networks, that will allow us to inspect, for each prediction of an image label, what pixels were used for prediction. We need a human-interactive tool (either web app, or standalone app that can run on either Windows, Mac or Linux) that users can use to flip through predicted images and understand what image features are being used to distinguish animals with/without a disease-causing mutation. | |
| - Python experience is a must - Either PyTorch or TensorFlow experience is a must | |
| ********** | |
| 30-60 min weekly or more | |
| Open source project | |
| Attachment | N/A |
| Yes | |
| Team members | N/A |
| Rex Liu | |
| N/A |