This GitHub repository contains the notebooks we will go over for CSC509.

Number of Credits: 3

Course Level: Junior and senior undergraduates

Description: The interpretation of medical images is a major bottleneck in medicine. Traditionally, medical experts must train for years or decades to become proficient in distinguishing abnormalities such as lesions or cancerous growths from normal tissue. Nevertheless, expert radiologists often disagree in their interpretation of such images, resulting in delayed or erroneous diagnoses or uncertain measurements of disease progression. This course explores the application of state-of-the-art deep learning models to biomedical image analysis: the task of identifying objects such as features within an image and classifying images according to disease type. The course introduces key medical imaging technologies and data types, begins with an overview of topics central to medical image analysis and deep learning based image analysis and culminates in two hands-on case studies.

Course objectives:

1. Learn the fundamentals of biomedical imaging with focus on different imaging technologies and their applications in clinical practice and in clinical research. Students should be made cognizant of current challenges and some ways that machine learning can address them.

2. Learn how to apply machine learning tools to classify images and identify disease subtypes.

Learning objectives

• Understand the basic principles of medical images (X-ray, CT, PET, MRI, OCT, US) with a focus on:
  • How medical imaging is used in clinical practice and clinical trials
  • Principles and acquisition basics of different imaging technologies
  • Some example cases for disease diagnosis, prognosis and trial endpoints
• Understand why deep learning/CNNs are specifically useful for medical imaging problems
  • What kinds of problems are we currently using deep learning for in medical imaging? (e.g. diagnosis, prognosis, segmentation, automatic labeling and image retrieval, quantifying change)
• Understand the unique challenges in applying CNNs to medical images
• Understand convolutions and how they are uniquely suited to automatic feature extraction with image-based data

Goal: Understand 2 commonly used medical image formats (dicom, nifti) and visualize medical images

Goal: Reinforce the basics of convolutional neural network via PET vs MRI simple classifier development in Keras

Goal: build a classifier that can distingush the T1 vs T2 images. The notebook has already been pointed to the data folder and you will have to do is fill-in the code block Descriptions We learned how to build a basic classifier to identify PET vs MRI images. Recall from the lectures we can have variations in the contrast of the MRI utilizing T1 and T2 weighted signals. In this project we ask you to adopt the notebook and build a CNN model that can classify T1 and T2 MRI images. As part of this exercise, we want you to submit report in google doc/pdf of the following

• What are the differences between a T1 weighted vs T2 weighted MRI? How would it show up when you look at the images?
• Complete the notebook for T1 vs T2 classification using: 1. Adam optimizer and 2. SGD optimizer, report your performance when switching these two optimizers. Copy /screen capture, the training performance, confusion matrix, ROC and PRC curves as images in your report showing how the performance changed with different optimizers.