Phenotying Mutant Mouse using Deformation Features


Sharmili Roy Xi Liang Asanobu Kitamoto Masaru Tamura Toshihiko Shiroishi Michael Brown

(MICCAI'13)





Introduction

Completion of the human genome project brought comprehension of location and sequence of each human gene. Scientists now want to map these individual genes to their corresponding physiological functionalities. Mouse is chosen as the principal study model for this gene mapping due to its 99% genetic similarity with humans. Gene targeting technology is being actively employed by many international organizations to generate transgenic mouse lines by knocking out each of the approximately 25,000 mouse genes (i.e., systematically removing each gene one by one and growing the mouse). High-throughput phenotypic assessment systems are necessary to systematically analyze and interpret the genetic information generated by these large-scale mutagenesis programs. A significant proportion of the generated mice strains are embryonic lethal resulting in the shift towards embryo-centric phenotyping. Mouse phenotyping still largely relies on microscopic evaluation of mouse sections which is not only extremely low throughput but is also highly inefficient. We aim to enhance mouse phenotyping by proposing a generalized defect detection framework that automatically identifies phenotypic areas in micro-CT images of mutant mouse embryo using registration and deformation-based features.




Contributions

  • In this project, we propose a deformation-based feature, named, deformation stress, to identify candidate defective areas in micro-CT images of pre-natal mutant mice.

  • Using deformation stress and other deformation features like we design an automated and generalized defect detection framework that identifies and highlights defects in mouse micro-CT images.

  • Unlike other defect detection algorithms designed to detect specific known defects, our system also highlights candidate novel defects that may not be readily recognized by human experts due to absence of significant visual features. This greatly enhances phenotyping throughput by reducing the vast search space of novel phenotypes.




Method

Defect detection consists of two main steps; the first step is construction of a mean for the normal mouse population and the second step comprises of computing defects in mutants using deformation features obtained by registering them to the normal mean.

Step I: Construction of normal mean

This step consists of extracting wild mouse images from their acquisition volumes and group-wise registering them via rigid, affine and B-spline registration steps to compute a normal consensus average image.



Figure 1: This figure illustrates the steps in the computation of normal mean image. (a) Acquisition volume (b) extracted normalized embryo images (c)-(e) consensus average images at rigid, affine and B-Spline registration stages respectively.

Step II: Computation of deformation features

Mutant mice are registered to the normal mean and deformation features are extracted from the respective deformation fields.



Figure 2: A mutant mouse is registered to the normal mean and the corresponding deformation vector is computed.

For each voxel we have:

Transformation in x-direction – Tx , θx
Transformation in y-direction – Ty , θy
Transformation in z-direction – Tz , θz

We compute three deformation features from this deformation field as follows:

1. Determinant of spatial Jacobian of transformation at a voxel is given as:

1+∂(Tx)/∂(x) ∂(Tx)/∂(y) ∂(Tx)/∂(z)
∂(Ty)/∂(x) 1+∂(Ty)/∂(y) ∂(Ty)/∂(z)
∂(Tz)/∂(x) ∂(Tz)/∂(y) 1+∂(Tz)/∂(z)

We define IJ as the volumetric regions that have high determinant of spatial Jacobian of deformation. IJ identifies regions with high local expansion and compression.



Figure 3: IJ overlaid on a mutant mouse image.

2. Let Θ = [θx θy θz]T. We divide the deformation field into small blocks and find blocks which have high entropy of Θ. IS, called deformation stress, identifies the volumetric regions with high entropy of deformation direction (Θ). IS is meant to capture the incoherency in deformation directions within small neighborhoods.



Figure 4: IS overlaid on a mutant mouse image.

3. For a population of NM mutant images registered to the normal mean NAvg, we compute the voxel-wise intensity variance of the mutant population as follows:

We define IIV as the volumetric regions with high intensity variance. IIV in essence captures voxels that have low registration accuracy.



Figure 5: IIV overlaid on a mutant mouse image.

We combine the three regions defined above to detect defects as follows:

IDefect=(IIV ∩ IJ) ∪ (IIV ∩ IS) ∪ (IJ ∩ IS)

IDefect identifies the resulting defective areas.



Figure 6: IDefect overlaid on a mutant mouse image.




Results



Figure 7: This figure illustrates some of the defective areas identified by the proposed defect identification rule.



Figure 8: This table describes the detection performance for known defects, such as polydactyly (presence of extra fingers and toes) and ventricular septum defect (VSD, presence of a hole between the two ventricles of the heart). VSD is assumed detected if ventricular area is highlighted.




References

  1. Sharmili Roy, Xi Liang, Asanobu Kitamoto, Masaru Tamura, Toshihiko Shiroishi, Michael S. Brown, "Phenotype Detection in Morphological Mutant Mice using Deformation Features," Medical Image Computing and Computer Assisted Intervention (MICCAI'13), Sep 2013 [PDF] [Poster] [Slides]


© Copyright 2014, Sharmili Roy