Example of gross images of wild-type (normal) and mutant zebrafish at age 5 dpf (days post fertilization). Image source: http://web.mit.edu/hopkins/images/full/1532B-1.jpg

Example of histological images of wild-type (normal) and mutant zebrafish at age 7 dpf. Images taken from Penn State Zebrafish Atlas [28]; direct link to virtual slide comparison tool at http://zfatlas.psu.edu/comparison.php?s[]=264,1,382,172&s[]=265,1,474,386

An overview of the high-throughput histology workflow (modified from Sabaliauskas et al. [32])

Example of histological images of wild-type (normal) and mutant zebrafish eyes at age 5 dpf (days post fertilization)

Typical example of a histological section of a larval zebrafish array (left) and its computationally detected underlying lattice structure (right)

Illustration of the array image lattice detection procedure

When each 5-dpf larval zebrafish image is rotated to align to a common origin point (here, the mouth, located at the leftmost position on the midline in the above images) and also along a common horizontal axis, we find that the positions of the eyes across all images are largely overlapping, which allows us to use a relatively simple location-based method for extracting a 768 × 768 square region around each eye for input to the SHIRAZ system

Labeled layers of the 5-dpf zebrafish eye. In addition, the arrow indicates the direction of the implicit “rotational symmetry” in certain histology images such as this, where we typically find a constant and repeating pattern of cellular layers that revolve about the lens

Overview of frieze-like expansion of zebrafish retina. Note that because of the distortion resulting from re-scaling the shorter line segments (generally found near the marginal zones of the retina, on either side of the lens), we extract only the middle 50% of pixels (dashed line) from the frieze-expanded eye image

Illustration of subdivision of frieze-expanded eye “parent” image into 64 × 64 “child” block images (Selected image blocks from the subdivision of the parent image have been marked with dotted borders to show their corresponding positions in the “exploded” view)

An example to illustrate the rationale for using overlapping child block images

Example of wavelet packet expansion for a 64 × 64 “child” image block taken from a frieze-expanded “parent” image. The expansion proceeds through four steps, and following the initial expansion, the image corresponding to the highest-energy band is chosen for the next level of expansion. A total of 16 energy values are thus used as texture features for each child image block

Examples of actual values of texture features extracted from two different image blocks taken from the same frieze-expanded parent image

Illustration describing how the feature matrix extracted from each frieze-expanded image is reshaped into a series of sorted nine-element feature vectors, one for each feature-neighborhood correspondence

Typical example images spanning the range of histological phenotypes and artifacts that SHIRAZ is trained to recognize

Illustration of agreement test for a nine-element “query vector” (that is, corresponding to a given 3 × 3 child block neighborhood, as in Fig. 12) and its degree of match to the statistically determined feature signature corresponding to that texture feature (see Sections 3.4.2 and 3.4.3). In practice, we allow for up to two elements to be outside the boundaries before the query vector fails the “betweenness test ” for that texture feature

Screenshot of entry point to SHIRAZ Web-based demo site

Selected images and their three highest-ranking annotations as predicted by SHIRAZ, with degree of correctness of the annotation given in parentheses. Correct annotations are shown in boldface, with incorrect annotations shown in italics

An example walkthrough of the SHIRAZ “Array mode.” Following the eye image extraction step, the user can choose one of the eye images for phenotypic annotation and retrieval of similar images