3D Detection of Mutant Hairs in Drosophila Wing Images

Abstract

We developed an automated mutant hair counting system in Drosophila wing images. In drug discovery, Drosophila is used in a gentoxicity test: the Drosophila wing somatic mutation and recombination test (SMART)[3][1]. This test analyzes the development frequency of mutant hairs of Drosophila. Since the current mutant hair counting method is totally manual work done by human rare eyes through the microscope, its speed and accuracy is limited. In our previous work [4], we developed the image acquisition method using multi-focused image stack, hair separation method into upper and lower hair, and hair detection method using template matching. In this work, the hair detection and mutant classification algorithms are enhanced and extended to 3D, and the wing area segmentation is newly developed.

Full text

3D Detection of Mutant Hairs in Drosophila Wing Images
Jihoon Kwak 1,2, Chun-Taek Oh 3, Sung-Jun Han 3, Michael Hansen 2
and Myungjoo Kang 4
1) Interdisciplinary Program in Computational Science and Technology, Seoul National
University, Seoul, Korea
2) Image Mining Group, Institut Pasteur Korea, Gyeonggi, Korea
3) Drug Biology Group, Institut Pasteur Korea, Gyeonggi, Korea
4) Department of Mathematical Science, Seoul National University, Seoul, Korea
Corresponding Author :Jihoon Kwak, jkwak314@snu.ac.kr
ABSTRACT
We developed an automated mutant hair counting system in Drosophila wing images. In drug
discovery, Drosophila is used in a gentoxicity test: the Drosophila wing somatic mutation and
recombination test (SMART)[3][1]. This test analyzes the development frequency of mutant
hairs of Drosophila. Since the current mutant hair counting method is totally manual work
done by human rare eyes through the microscope, its speed and accuracy is limited. In our
previous work [4], we developed the image acquisition method using multi-focused image stack,
hair separation method into upper and lower hair, and hair detection method using template
matching. In this work, the hair detection and mutant classification algorithms are enhanced
and extended to 3D, and the wing area segmentation is newly developed.
INTRODUCTION
The Drosophila’s wing (Fig. 1 (a)) has tiny hairs. Normal hairs (Fig. 1 (b)) are regularly spaced
and their orientations are also uniform. But, mutant hairs (Fig. 1 (c)) are less regularly spaced
and less uniform in orientation. Usually, two or three hairs are closer than others. We want to
count such mutant hairs automatically. The first problem is out-of-focus. Since the wing surface
is uneven and bumpy, it is hard to take a clean image focused in all the hairs. If it were possible,
another problem would be arose: the overlap of upper and lower side hairs. So, we take multiple
images varying the focus plane (Fig. 1 (d)).
WING SURFACE RECONSTRUCTION
Applying maximum projection method to this multi-focused image stack, we get a synthetic
clear image in which every hairs are focused. However, even normal hairs might be misunder-
stood as mutant hairs because the upper and lower hairs are overlapped in the image as Fig. 2 (a).
So, we need to separate the upper and lower hairs. We use variation span as focussing criteria
as in [5] and estimate the depth map (Fig. 2 (b)). Using this depth map, we can reconstruct the
surface (Fig. 2 (c)), which separates the upper and lower hairs. After maximum projection on
each side, we get the upper and lower hair images (Fig. 2 (d)).

(a) Drosophila Wing (b) Normal Hairs (c) Mutant Hairs (d) Image Stack
Figure 1. Drosophila Wing Images. (a) is an image of a Drosophila’s wing. (b) and (c) are
images of hairs on the wing. Normal hairs of Drosophila are regularly spaced as in (b). But
mutant hairs are less regularly spaced and less uniform in orientation as the hairs in red circle
of (c). Since the wing surface is uneven, (d) multi-focused image stack is required.
(a) Overlapped Hairs (b) Depth Map (c) Wing Surface (d) Separated Hairs
Figure 2. Wing Surface. Since wings have hairs on both sides, (a) the maximum projection of
image stack has the overlap of upper and lower hairs. To distinguish them, (b) depth map is
estimated and (c) the wing surface is reconstructed. Using this surface, (d) upper and lower
hairs can be separated.
HAIR AREA SEGMENTATION
We need to exclude hairs on blood vessels and the boundary of wing (Fig. 3 (c)), because the
size and distribution of hairs there are quit different from normal hair area (Fig. 3 (d)). Also,
we don’t need to try hair detection on background. Simple intensity threshold might not work
since the background intensity could be bright on boundary of image. First, we segmented the
background (Fig. 3 (b)) using total variation, which does not depend on the smooth variation of
background intensity. Second, the blood vessels are segmented by Mean Curvature Threshold
after smoothing hair intensity. Finally, we can get the hair area (Fig. 3 (d)) by complement of
the background and blood vessel areas.
(a) Original Image (b) Background (c) Blood Vessel (d) Hair Area
Figure 3. Hair Area Segmentation. (a) A Wing image can be partitioned into (b) background,
(c) blood vessel, and (d) hair area. Hair detection is done only on hair area.

HAIR DETECTION IN 2D AND 3D
In our previous work [4], 3D image stack was projected into two 2D images: upper and lower
hair images. In each 2D image, hairs were detected using template matching algorithm [2]. In
this work, instead of time consuming template matching, we use Hessian matrix of image in-
tensity. The trace of Hessian matrix is the Mean curvature (Fig. 4 (a)) and it has local extreme
values inside of hairs. After merging process, we can get one location point for each hair. The
smallest Eigenvector of Hessian matrix (Fig. 4 (b)) is the direction of hairs. Since the Mean
curvature has zero values on the boundary of hairs, we can measure the length of hair by check-
ing the zero crossing points (Fig. 4 (c)). Location, direction, and length compose line segments,
which correspond well to hairs (Fig. 4 (d)).
Even though the new hair detection algorithm itself works fine, the 2D projection image
has critical artifacts: missing hair intensity and root intensity from other side. This is due to the
overlap of upper and lower hair at hair surface. So, instead of using 2D projection image, we
use directly 3D image data. Hair detection algorithm in 2D can be naturally extended to 3D.
Location, direction, and length of 3D line segments are detected (Fig. 5 (a), (b)).
(a) Mean Curvature (b) Eigenvectors (c) Boundary (d) Detected Hairs
Figure 4. Hair Detection in 2D. (a) the Mean Curvature of an wing image has local extreme
values inside of hairs, which give the location of hairs. (b) Eigenvectors of Hessian matrix
give the direction of hairs. The Mean Curvature has (c) zero values on the boundary of hairs,
which gives the length of hairs. Location, direction, and length compose line segments, which
correspond to hairs.
(a) Location (b) Direction, Length (c) Side Classification (d) 2D Projection
Figure 5. Hair Detection in 3D. Similarly as in 2D case, 3D Mean Curvature and 3D Eigenvec-
tors of Hessian gives (a) location, (b) direction and length of hairs in 3D. Using wing surface,
(c) upper and lower hairs can be classified. (d) is the 2D projection of 3D hair detection.

SIDE AND MUTANT CLASSIFICATIONS
After detecting hairs, we need two classifications. One is the upper and lower side classification
and the other is normal and mutant hair classification. The side classification (Fig. 5 (c),(d))
can be done easily by the wing surface information (Fig. 2 (c)). For the mutant classification,
we used three criteria: length, distance, and angle (Fig. 6 (a), (b)). The length is the length of
detected hair. The distance is from current hair root to the closest neighbor hair root. The angle
is the difference of current hair angle and the average angle of neighbor hairs. By threshold of
three criteria, we can classify mutant hairs (Fig. 6 (c), (d)).
(a) Mutant Criteria (b) Distribution (c) Mutant Classification (d) 2D Projection
Figure 6. Mutant Classification. Three criteria, distance, angle, and length, are used to classify
mutant hairs. (a) is the histograms of each criterion. (b) is the 3D distribution of criteria. (c) is
the mutant classification result. (d) is the 2D projection of 3D mutant classification.
ACKNOWLEDGMENT
This work was supported by the National Research Foundation of Korea(NRF) grant funded by
the Korea government(MEST)(No. 2009-0090095).
REFERENCES
1. H.H.R. de Andrade, M.L. Reguly, and M. Lehmann. Wing somatic mutation and
recombination test. METHODS IN MOLECULAR BIOLOGY-CLIFTON THEN TOTOWA-,
247:389–412, 2004.
2. R.C. Gonzalez and R.E. Woods. Digital image processing. 2002. ISBN: 0-201-18075-8.
3. U. Graf, FE W¨
urgler, AJ Katz, H. Frei, H. Juon, CB Hall, and PG Kale. Somatic mutation
and recombination test in Drosophila melanogaster. ENVIRON. MUTAGENESIS., 6(2):153–
188, 1984.
4. Jihoon Kwak, Chun-Taek Oh, Sung-Jun Han, Myungjoo Kang, and Auguste Genovesio. Hair
Separation and Detection in Mutant Drosophila Hair Images. Proceedings of the Korean
Society for Industrial and Applied Mathmatics, 5(1):153–155, 2010.
5. D. Martisek. The Two-dimensional and threedimensional processing of images provided by
conventional microscopes. Scanning, 24(6):284–296, 2002.