Our automated wing analysis system, WINGMACHINE, successfully fulfils its intended purpose as a means of rapidly gathering repeatable high-dimensional phenotypic data. We have shown that the system is useful for characterizing variation among Drosophilid species, and that it facilitates artificial selection experiments on complex aspects of wing shape.
Dryden and Mardia [37
] divide image analysis into "low" and "high-level" operations. Low level analysis involves local operations on small numbers of pixels, such as filters and edge detection. High level analysis involves detection and fitting of large-scale features of an image. Our use of an a priori
model of wing shape that is deformed to optimize fit to each image is a simple example of high-level analysis.
Prior to developing this approach, we devoted considerable effort to developing a feature extraction system based entirely on low level analysis. These efforts were frustrated by several features of wings. The leading edge of the wing consists of a thickened vein that exhibits high contrast, while the trailing edge of the wing does not. Second, lighting across the image is uneven. Third, small flaws in the image, such as dust or hairs, or in the wing itself, such as small nicks, are hard to automatically disentangle from wing features. All of these frustrate simple edge detection and tracing algorithms. WINGMACHINE successfully splines wings that are both damaged and dirty. Similar complications are common in most biological imaging problems. Our success in implementing high-level analysis suggests that it could be useful in a large number of image analysis applications in basic biology.
More specifically, our approach may be directly extensible to other objects that can be summarized as a framework of intersecting lines, such as leaf veins and edges, scales or feathers. The specification of a model with different vein or edge topologies than in Drosophila wings is readily accomplished. While the precise algorithms in our software are specifically tailored to Drosophila wings, our general approach might be useful to fit models of very different structures.
In comparison with the more widely used hand-digitization of wing landmarks (e.g. [12
]) the WINGMACHINE approach has the advantage of great speed, both in handling the specimens, and recovering quantitative information from them. An experienced operator spends on average about 1 minute per specimen in total. This speed comes with some disadvantages. While the repeatabilities of most landmarks are quite high, human observers can in some cases do much better. If the goal is to characterize the mean of a population (such as a family or a species), there is a simple tradeoff between speed and accuracy: if it takes x
times as long to measure an image by hand, then it will be worthwhile to do so if the measurement error of the automated system is greater than x
times the measurement error achieved by hand.
The structure of the model chosen for fitting and the details of image processing determine the precise locations of the curves and interesections recovered. The result is that the landmarks, for example, are frequently not as a human observer would place them. For example point 11, the intersection of L2 and L3, has relatively low repeatability because it is recognized as the intersection of the curves along these veins, rather than as the sinus formed by the interior outline of the veins, as a human observer would naturally do. This feature of the model potentially creates bias if a particular feature of the wing is of primary interest.
Another limitation of the system is that it is restricted to dealing with the distal features of the wing. The attachment point of the wing to the body, and the complex arrangement of veins near that point are not included. These aspects of wing form may be important mechanically and aerodynamically.
A final disadvantage of the WINGMACHINE may fail for wings of species with highly melanized spots at vein intersections, for example the "picture-winged" Hawaiian Drosophila. Initial attempts to spline wings of D. grimshawi have such a high error rate that hand-digitization is simpler and less time-consuming. On the other hand, melanization seems to be dependent on rearing conditions, and we have had good success with lighter-winged individuals of another picture-winged species, D. gymnobasis.
Ultimately, our understanding of biological systems needs to encompass the relationships between molecular and phenotypic data. Much attention is now focused on high throughput genomic techniques such as sequencing, expression microarrays and proteomics. To take advantage of this avalanche of genetic data, comparable efforts will be needed to characterize the whole-organism phenotype, what might be called phenomics [38
]. The WINGMACHINE shows that high-throughput phenotyping is also feasible.