B. GLI and anisotropy.
Closely related problem is quite wrongly claimed independence of GLI from structural anisotropy. In 2000’s paper we read about volume density measure: “its major advantage is that no assumptions on the special orientation of cells is required”. However, this assumption could be true only if all required steps of design-based stereology are satisfied (see, for example, S.Tschanz, J.P.Schneider, L.Knudsen, “Design-based stereology: Planning, volumetry and sampling are crucial steps fora successful study”, Ann. of Anatomy, 2013, p. 5, section 2.2). Moreover, as A.Schleicher et all. admitted, GLI is “biased by projection“. Therefore, the value of this bias has to be different for anisotropic structures projected under different angles, as it is the case for pyramidal neurons. If viewed under oblique angle, the projected image of the pyramidal neuron will lose the apical dendrite, which will systematically decrease the area fraction estimate by hardly predictable amount. In addition, compared to the area fraction itself, the profile of area fraction of anisotropic structure changes more dramatically by section orientation differences. For example, the peak of the profile in the region of pyramidal layer (III or V) can significantly decrease or completely disappear just because of the gradual change of orientation of cortical plate. It goes without saying, that in the region of a section orthogonal to the orientation of the cortical columns the GLI method will not work at all.
Based on these examples, one can conclude that in all analyzed papers authors demonstrated some problems with understanding basic terminology used in stereology. Still, even if we admit that GLI is just “an area fraction”, and it might have something to do with stereology since it could be used to calculate volume fraction, the method of comparison of two profiles of area fraction in 20 microns-thick 2D sections cannot be called “stereological” by any extent of imagination.
C. Questionable robustness of image processing
The second problem is the omission of comprehensive explanation of two very important details: adaptive thresholding and image processing algorithms. Without explaining them one can hardly expect the possibility of reproduction or verification of the same result anywhere else. The earliest reference to used method is to the paper of Prewit, J.M.S. and Mendelsohn, M.L. (1965) “The analysis of cell images,” Ann. N.Y. Acad. Sci., 128: 1035-1053. (This source is presently unavailable to me, but will be added ASAP. Nevertheless, judging by the date, it is unlikely to be very sophisticated.) In 1990’s publication adaptive thresholding procedure was described as thresholding by the grey value of the image, obtained using low-pass filter by structuring element that matches the size of the biggest cell, plus a fixed offset (8 grey levels in that case). Probably, this method was acceptable at the time, however, nowadays it sounds quite primitive. Later, in 2000’s publication, adaptive thresholding procedure was referenced again, now by pointing to the book “Digital Image Processing” by K.R. Castelman, published in 1979. At the same time, this paper bluntly claims that “GLI data are not affected by local variations in staining intensity” without any experimental proof of such statement. This claim is misleading again: GLI data might not depend on staining intensity only if a robust adaptive thresholding algorithm is applied. Obviously, the description presented in 1999’s paper does not look as robust at all. References in later papers are not specific enough to understand was it modified later, and if yes – how it was implemented. Descriptions of image processing steps in studies papers is even more limited. It seems that nothing is done prior to GLI image creation in addition to thresholding. Smoothing of GLI image by two passes of 3×3 median filter to “eliminate small particles” was briefly mentioned in 2009’s paper.
Additionally, important series of experiments conducted between 1980 and 1984 in our group in Moscow clearly demonstrated that staining variability is a major factor of lack of reproducibility of area fraction measurements. I personally spent two years to develop a technology which allowed to obtain stable and reproducible measurement data regardless of staining variability. This technology was published in 1985 (V.V. Istomin “Fundamentals of adaptive staining of cortical sections for automatic structural analysis”, Kosrakov’s Journal of Neuropathology and Psychiatry, LXXXV, 7,1024-1032). Ten years later a lion share of there results were included in a paper by Amunts, Istomin, Zilles and Schleicher (“Postnatal development of human primary motor cortex: a quantitative cytoarchitectonic analysis”, 1995, Anat. Embryol, 192:557-571), even though nor K.Zilles, neither A.Schleicher had anything to do with development of this technology. Even though K.Amunts did not participate in development of this technology either, she used it for several years collecting data for her Ph.D. (defended in 1989, K. Fiedler, “Quantitative analysis of the cytoarchitecture of area 4 of brain cortex in human onthogenesis”). In other words, the claim of independence of area fraction measurements from staining intensity not only contradicts experimental results used by K. Amunts (K. Fiedler) for many years, but it also contradicts to the data published at least four times with K. Amunts (K. Fiedler) as co-author (in 1990, 1992, 1994 and 1995). So, how to explain that Amunts, Zilles and Schleicher in their GLI-related papers ignored this contradiction?
To say the least, considering information published during all these years, the quality of image processing methods employed in GLI profiling technology is a poor match to reproducibility requirements, let alone – to the capabilities of modern image processing algorithms.