University of Toronto Computer Engineering Capstone Project 2023.

Built with React, Tensorflow, and publicly available U-Net architectures.

Potential video upload coming soon.

Project Report

System Overview

(a) Surveyed top three state-of-the-art models for volume segmentation from the MICCAI-ACDC challenge based on top average dice (IoU) scores for the left atrium, left ventricle, and myocardium

(b) Experimented with code repositories such as TransUNet, SSL4MIS, and w-net to evaluate initial feasibility of implementation in our project

(a) Integrated pydicom library with test code in Python to remove patient metadata from MRI images and sample according to underlying patient and MRI machine metadata

(b) Created data visualization script with mito and pandas to identify biases in the dataset, such as the lack of sex-disaggregated data, as well as variances in weight, height, age and sex

(c) Implemented a more ethical data sampling process by sampling an equal amount of men and women from the labelled samples in UK Biobank

(a) Learned to manually label cardiac MRI images with Scale.ai

(b) Created a batch image export script in Python to download all labelled images from Scale.ai for significant speedup in data labelling process

(a) Initial baseline model test using the keras segmentation library on Google Colab 

(a) Created functional requirements and sampling diagrams for test report, final report and demo presentations

(b) Leveraged Figma to create frontend website design for demoing different model performance and volume segmentation

(c) Created a frontend website using React to realize Figma design 

(a) Recorded team meeting minutes and next steps

(b) Drafted, reviewed and sent emails to the supervisor