Convolutional Neural Networks with Applications in Medical Image Analysis, 7.5 Credits

Contents

This course covers deep convolutional neural networks (CNNs) for computer vision, with applications in medical image analysis. The course provides an introduction to fundamental concepts in machine learning, describes neural networks and the field of deep learning, and goes into detail about deep CNNs. The course describes the different parts that are used when building deep CNNs, such as filters, activation functions, loss functions; regularization techniques such as e.g. batch normalization and dropout; explains several of the different non-linear optimization algorithms that are used when training the networks, such as stochastic gradient descent, Adam, etc.; and describes popular network architectures, such as e.g. the U-Net, ResNet, and DenseNet, and discusses their pros and cons. The course also covers generative models, such as variational autoencoders (VAE) and generative adversarial networks (GANs).

Students in this course will learn to implement and train modern network architectures and deep learning methods, and apply these to large image datasets with medical and other images.

The course has two modules:
Theory and method, 5.5 ECTS credits
Practical assignments, 2.0 ECTS credits

Expected learning outcomes

After having completed the course, the student should be able to:
 
Knowledge and understanding

• Explain and describe central concepts in machine learning, such as e.g. training and validation data, classification and regression, supervised and unsupervised learning, bias and variance, loss function, generalization error, accuracy, precision, recall, etc. (FSR1)
• In detail explain the central concepts in deep learning, such as e.g. depth, learning rate, hyper parameter, early stopping, overtraining regularization, etc. (FSR2)
• In detail explain the different parts of a convolutional neural network and their purpose, such as e.g. layer, nodes, activation function, dropout, loss function, residual connection, etc. (FSR2)
• Describe the stochastic gradient descent algorithm, its benefits, and the main differences between it and other non-linear optimization algorithms that are used when training neural networks. (FSR4)
• In detail explain some of the most common architectures and models used, and describe the main benefits of the different architectures. (FSR5) 

Skills and abilities

• Independently design and implement common network architectures, and use these in applications for e.g. classification and regression. (FSR6)
• Use modern software libraries for building and training deep convolutional neural networks to solve specific tasks. (FSR7)
• Evaluate and compare performance and generalization of deep convolutional neural networks. (FSR8) 

Judgement and approach:

• Discuss how modern machine learning and deep learning methods affect humans, companies, and the society. (FSR9)
• Discuss the state of the field of deep learning and its place in machine learning, and discuss the general capabilities of artificial intelligence in our time. (FSR10)
• In a scientific way evaluate different solutions pros and cons. (FSR11)
• Critically review results or statements from other studies, or e.g. products that either are based on deep learning or otherwise is using it. (FSR12)

Required Knowledge

To be admitted you must have 90 ECTS credits in one of the main areas of computer science, physics, electronics, chemistry, mathematics or mathematical statistics, or two years of completed studies (120 ECTS credits).  To be admitted to the course, the credits above must include at least 7.5 ECTS credits within Programming methodology (e.g. 5DV104, 5DV157, 5DV158, 5DV176, or 5DV177), at least 7.5 ECTS credits within Data structures and algorithms (e.g. 5DV149, 5DV150, or 5DV169), at least 7.5 ECTS credits within Linear algebra (e.g. 5MA019 or 5MA160), at least 7.5 ECTS credits within Calculus (e.g. 5MA009 or 5MA153), at least 7.5 ECTS credits within Mathematical statistics (e.g. 5MS005, 5MS045, 5MS043, 5MS068, or 5MS069), or equivalent knowledge.
Proficiency in English equivalent to Swedish upper secondary course English A/5. Where the language of instruction is Swedish, applicants must prove proficiency in Swedish to the level required for basic eligibility for higher studies

Form of instruction

The course consists of lectures and practical assignments. In addition to scheduled activities, the students are required to work with the material individually. The students are required to cover the material and prepare questions in order to obtain appropriate feedback

Examination modes

Module 1 (Theory and method, 5.5 ECTS credits, ELO 1-5, 9-12): The module is assessed through a written exam. The grades for this module are: Fail (U), Pass (G), or Pass with merit (VG).
Module 2 (Practical assignments, 2.0 ECTS credits, ELO 6-8): The module contains three mandatory assignments that results in written reports. The grades for this module are: Fail (U) or Pass (G).
On the course as a whole, the grades are: Fail (U), Pass (G), or Pass with merit (VG). To pass the course, both the exam and all the assignments must be assessed and completed.

A student who has taken two tests for a course or a segment of a course, without passing, has the right to have another examiner appointed, unless there exist special reasons (Higher Education Ordnance Chapter 6, Section 22). Requests for new examiners are made to the head of the Department of Radiation Sciences.

A student who has passed an examination may not be re-examined.
For all students who do not pass the regular examination there are additional opportunities to do the examination in accordance with the University regulations (FS 1.1.2-553-14).

Other regulations

Examination based on this syllabus is guaranteed for two years after the first registration of the course. This applies even if the course is closed down and this syllabus ceased to be valid.