Visual Generative Model and Its Applications

  • 2-2 ECTS credits
  • Academic year 2024-2025
  • DP00AY99-3001
This short course is designed for research students who are interested in image generation, especially medical image. It is well known that the acquisition of the medical images is expensive and time consuming while deep learning methods for medical images require large amounts of data. This course will focus on how to generate anatomically plausible medical images and apply such synthetic data for different applications. After completion of this short course, students will have a good understanding of the problems, existing approaches, and current state of the art of generative models for medical image generation.

Education information

Implementation date

07.05.2025 - 12.05.2025

Education type

Field-specific studies

Alternativity of education

Optional

Location

Linnanmaa

Venue location

The twelve face-to-face lectures will be conducted in four mornings as follows:
• Date: 07 May 2025 Time: 8:30 - 10:00, 10:15 - 11:45 Venue: Lo124
• Date: 08 May 2025 Time: 8:30 - 10:00, 10:15 - 11:45 Venue: SÄ102
• Date: 09 May 2025 Time: 8:30 - 10:00, 10:15 - 11:45 Venue: Lo124
• Date: 12 May 2025 Time: 9:00 - 11:00 Venue: Lo124

Enrollment and further information

Education description

This class will provide a comprehensive overview of visual generative models, covering foundational concepts and their practical applications across diverse domains. We will begin by exploring the basics of diffusion models, delving into their underlying principles and mechanisms. The course will then focus on the role of generative models in medical imaging, emphasizing techniques to ensure anatomical plausibility in generated images. We will discuss how these methods can be effectively applied to downstream tasks such as diagnostic support, segmentation, and anomaly detection. Additionally, the class will highlight the use of generative models in art creation, demonstrating innovative approaches to incorporate art evaluation metrics into the training loop, enabling more refined and aesthetically pleasing outputs. A key aspect of the session will address strategies to accelerate the sampling process, introducing both training-free and low-cost optimization techniques to enhance efficiency without compromising performance. This lecture aims to bridge theoretical foundations with hands-on applications, catering to learners interested in both technical advancements and real-world implementations of visual generative models.

The invited lecturer is Wei Peng from University of Stanford. More information about the course from Professor Guoying Zhao.

Last updated: 26.3.2025