Artificial Intelligence Genetics and Imaging

Dr. Degui Zhi explores the synergy of AI, genetics, and medical imaging. The speaker discusses the limitations of traditional imaging genetics and introduces unsupervised deep learning approaches to define endophenotypes for imaging genetics. He discusses using autoencoders to T1- and T2-FLAIR weighted brain MRI images from the UK Biobank to discover genes relevant to brain structures. The presentation also includes diffusion MRI, showcasing novel phenotypes. Self-supervised learning methods based on contrastive learning for retina images is explored, leveraging the shared genetic components of left and right eyes. Rigorous genetic association studies validate the models, demonstrating the potential of AI in generating complex phenotypes. The presentation encourages collaborative research across diverse medical imaging types.
1. Introduction to AI, Genetics, and Imaging
a. Speaker acknowledges the previous talks on general AI and imaging.
Time Stamp: 00:10-01:18
2. Traditional Imaging Genetics Approach
a. Explanation of how traditional imaging genetics involves image processing, generating descriptors, and conducting GWAS for genetic association.
Time Stamp: 1:19- 2:19
3. Limitations of Traditional Approach
a. Discussion on limitations of human-defined phenotypes and biases in imaging data.
Time Stamp: 2:20-3:24
4. Supervised and Unsupervised Deep Learning in Imaging Genetics
a. Introduction to supervised and unsupervised deep learning, focusing on autoencoders and contrastive learning.
Time Stamp: 3:25-4:29
5. Autoencoder for Brain Imaging
a. Implementation of autoencoder using brain imaging data from UK Biobank, explaining the compression of information into an endotype.
Time Stamp: 4:30-5:24
6. Contrastive Learning for an Unsupervised Approach
a. Explanation of contrastive learning and its application in capturing similarities between different images.
Time Stamp: 5:25-7:38
7. Results and Discoveries
a. Visualization of embedding or endotypes in 2D, showcasing natural variations in human shapes.
b. Genome-wide Association Study (GWAS) results revealing new phenotypes not identified by traditional methods.
Time Stamp: 7:39-9:46
8. Application to Diffusion MRI
a. Extending the approach to diffusion MRI, demonstrating its effectiveness in highlighting structures relevant to white matter.
Time Stamp: 9:47-10:28
9. Self-supervised Learning for Retina Images
a. Application of self-supervised learning to retina images, emphasizing the use of left and right eye pairs.
Time Stamp: 10:29-11:24
10. Validation and Genetic Associations in Retina Images
a. Confirmation of model performance by showing good separation between matched and random pairs of eyes.
b. GWAS results indicating significant signals related to retina images.
Time Stamp: 11:25-12:18
 
11. Summary and Future Directions
a. Summary of the ability of deep learning AI to generate phenotypes from complex imaging data.
b. Emphasis on collaborative applications across various diseases, modalities, and -omics data.
Time Stamp: 12:19-13:55