[General boards] [Winter 2023 courses] [Fall 2022 courses] [Summer 2022 courses] [Older or newer terms]

CSC494/495 Opportunity - Improving machine learning models’ performance on medical images with transfer learning

The Lab

Our lab, based out of Sunnybrook Research Institute, is interested in developing computational and machine learning tools for investigating neurological and cerebrovascular networks in a clinical and preclinical setting. Current diseases we are researching include traumatic brain injury, Alzheimer’s disease, and diabetes. To study these disease images are acquired in house and through collaborators with two-photon fluorescence microscopy, light sheet microscopy, and MRI. The diverse group consists of neuroscientists, computer scientists, engineers, and graduate students in the department of medical biophysics available to assist the student with the project. The student will be supervised by Drs. Maged Goubran (https://medbio.utoronto.ca/faculty/goubran) and Bojana Stefanovic (https://medbio.utoronto.ca/faculty/stefanovic)

Background

We are currently analyzing two-photon fluorescence microscopy and light sheet microscopy images from mice and rats featuring vasculature, neurons, and other support cells. These images include prominent levels of noise that lead to defects when utilizing traditional computer vision image processing techniques and requires hours of work to manually analyze. To overcome this challenge, we are developing convolutional neural networks (CNNs) to segment and analyze morphology of various components of the rodent brain in disease models, and functional alterations to these images during stimulation. For example, we recently developed a 3D CNN with Unet architecture featuring residual blocks and dilated convolutions for segmenting the hippocampus (Goubran et al. Human Brain Mapping 2020). These deep learning networks however requires copious amounts of data to train and validate. To minimize the amount of data required for training these models we aim to utilize transfer learning from models trained on synthetic images.

Project

The student will be involved with development of synthetic microscopy images and their corresponding ground truth segmentation masks. This pipeline will feature the ability to slightly perturb vascular network parameters such as slightly increasing volume of single blood vessels. These synthetic models will be used to:

  1. Assess the contribution of transfer learning on final model performance.
  2. Determine the sensitivity of the model to slight alterations in vascular diameter.
    These experiments will contribute to our model’s uncertainty assessment and reduce amount of training data needed for manual tracing. This project aims at familiarizing students with optimizing neural networks, simulating data and evaluation of modeling errors.

Student benefit

• Write a conference abstract and involvement on resulting papers
• Hands-on experience with developing and evaluating CNNs
• Hands-on experience with developing synthetic data/simulation
• Familiarity with acquisition and analysis of microscopy images
• Access to computational resources and collaborations with lab members
• Journal club and lab meetings discussions

Eligibility

One of CSC303, CSC311, CSC320, CSC336, MAT332, MAT344 (Recommended: CSC320)
Proficiency in python ideally familiar with NetworkX and scikit-image
Familiarity with bash and Compute Canada/SciNet
8-10 hours/week
CSC494/495 or volunteer with option to extend

Contact information

For further information and to apply please contact our graduate student Matthew Rozak: matthew.rozak@mail.utoronto.ca.

  [General boards] [Winter 2023 courses] [Fall 2022 courses] [Summer 2022 courses] [Older or newer terms]