• Introduction to Image Alignment [Ref. 1]
  • Problem Definition
    • Choosing a representation
    • Choosing an alignment criterion
    • Optimizing the alignment criterion
  • Some simple representations and alignment criteria
  • Optimization methods
    • Exhaustive search
    • Keypoint methods
    • Gradient descent
• Multimodal Alignment [Ref. 2]
  • Non-parametric density estimation
  • Alignment by maximization of mutual information

• Multi-image alignment [Ref. 3]
  • Problem definition
  • A joint alignment criterion
  • The congealing algorithm
  • Congealing 3D brain volumes [Ref. 4]
  • Digital brain atlases

Systematic screening for mutants with certain phenotypic alterations is one of the most powerful genetic approaches to study unknown biological processes. This approach allows unbiased identification of important regulatory elements without knowing underlying molecular machinery. It has been widely used in lower organisms during the past century, and made significant contribution to our understanding of the biology. Recently, we and other groups have generated a large number of mutant mice, so that systematic screens would be performed in this popular mammalian model. Mutants with various developmental or disease phenotypes have been successfully identified. In a pilot obesity/diabetes screen, we have isolated dozens of single gene mutations, suggesting these quantitative traits could be controlled qualitatively. Comparing with other model animals, longer generation time and smaller litter size make mice less efficient for genetic screens. Thus, techniques allow high resolution and noninvasive recognition of the structural, morphological, and behavioral changes in a small number of animals are highly desirable for the future study.

• Brown SD, Moore MW. Towards an encyclopaedia of mammalian gene function: the International Mouse Phenotyping Consortium. Dis Model Mech. 2012 May;5(3):289-92.
• Kennedy AJ, Ellacott KL, King VL, Hasty AH. Mouse models of the metabolic syndrome. Dis Model Mech. 2010 Mar-Apr;3(3-4):156-66.

• Using the warp as an information source
  • Learning from one example [Ref. 5]
  • Learning about anatomy from joint alignment
• Broadening the notion of alignment
  • Removing multiplicative offsets in MRI
  • Removing slowly varying bias fields in MRI [Ref. 6]
• Congealing complex images [Ref. 7]
• Deep belief nets for congealing

One direction in medical image analysis is to effectively represent knowledge and efficiently extract biomedical information (such as a deformable shape) from medical images. In particular, machine learning techniques (supervised, weakly-supervised, and unsupervised) have played increasingly important role. The large-scale data learning and analysis have also recently played a significant role in medical imaging.

The goal of this lecture is to provide a comprehensive assessment of discriminative learning techniques used for medical imaging applications such as anatomical structure detection and segmentation, image categorization, etc. Learning from an annotated dataset the covers the uncertainties involved in the applications, these techniques are able to derive compact descriptions between the image and knowledge and gain improvements in performance and speed when compared with conventional algorithms without using learning.

Coverage: Basics about supervised and semi-supervised learning

Lets make a date! Break the ice, mingle cultures and research, and discuss problems in the fields of computer vision, biology, and everything inbetwen. We will be working in small groups, and changing partners to brainstorm and discuss.