Acta Neuropathologica

, 112:287

Minicolumnar abnormalities in autism

Authors

    • Department of Psychiatry and Behavioral SciencesUniversity of Louisville
  • Imke A. J. van Kooten
    • Department of Child and Adolescent PsychiatryUniversity Medical Center
    • Department of Psychiatry and Neuropsychology, Division of Cellular NeuroscienceMaastricht University, and European Graduate School of Neuroscience (EURON)
  • Andrew E. Switala
    • Department of Psychiatry and Behavioral SciencesUniversity of Louisville
  • Herman van Engeland
    • Department of Child and Adolescent PsychiatryUniversity Medical Center
  • Helmut Heinsen
    • Morphologic Brain Research UnitUniversity of Würzburg
  • Harry W. M. Steinbusch
    • Department of Psychiatry and Neuropsychology, Division of Cellular NeuroscienceMaastricht University, and European Graduate School of Neuroscience (EURON)
  • Patrick R. Hof
    • Department of NeuroscienceMount Sinai School of Medicine
  • Juan Trippe
    • Department of Psychiatry and Behavioral SciencesUniversity of Louisville
  • Janet Stone
    • Department of Psychiatry and Behavioral SciencesUniversity of Louisville
  • Christoph Schmitz
    • Department of Psychiatry and Neuropsychology, Division of Cellular NeuroscienceMaastricht University, and European Graduate School of Neuroscience (EURON)
Original Paper

DOI: 10.1007/s00401-006-0085-5

Cite this article as:
Casanova, M.F., van Kooten, I.A.J., Switala, A.E. et al. Acta Neuropathol (2006) 112: 287. doi:10.1007/s00401-006-0085-5

Abstract

Autism is characterized by qualitative abnormalities in behavior and higher order cognitive functions. Minicolumnar irregularities observed in autism provide a neurologically sound localization to observed clinical and anatomical abnormalities. This study corroborates the initial reports of a minicolumnopathy in autism within an independent sample. The patient population consisted of six age-matched pairs of patients (DSM-IV-TR and ADI-R diagnosed) and controls. Digital micrographs were taken from cortical areas S1, 4, 9, and 17. The image analysis produced estimates of minicolumnar width (CW), mean interneuronal distance, variability in CW (VCW), cross section of Nissl-stained somata, boundary length of stained somata per unit area, and the planar convexity. On average CW was 27.2 μm in controls and 25.7 μm in autistic patients (P = 0.0234). Mean neuron and nucleolar cross sections were found to be smaller in autistic cases compared to controls, while neuron density in autism exceeded the comparison group by 23%. Analysis of inter- and intracluster distances of a Delaunay triangulation suggests that the increased cell density is the result of a greater number of minicolumns, otherwise the number of cells per minicolumns appears normal. A reduction in both somatic and nucleolar cross sections could reflect a bias towards shorter connecting fibers, which favors local computation at the expense of inter-areal and callosal connectivity.

Keywords

Autistic disorder/pathologyChild development disordersPervasiveNeocortexNeuropilPyramidal cells

Introduction

Minicolumns are basic architectonic and physiological elements identified in all regions of the neocortex [16] and in all mammalian species thus far evaluated [37]. The minicolumnar circuit is an evolutionarily and ontogenetically conserved template adapted in the various cortical areas according to their specific developmental and functional requirements. The minicolumnar core comprises radially oriented arrays of pyramidal projection neurons. At the core and periphery of the minicolumn, combinations of GABAergic interneurons provide for a diversity of signaling properties that serve to dynamically modulate pyramidal cell inputs and outputs that perform area and task-specific information processing needs [22, 30, 38, 51].

Available neuropathological and structural imaging data suggest that autism is the result of a developmental lesion capable of affecting brain growth. One possible explanation for this is the recent finding of minicolumnar abnormalities in autism (i.e., minicolumns of reduced size and increased numbers) [20]. In this initial study measures of minicolumnar morphometry were obtained relative to pyramidal cell arrays in nine autistic cases and in an equal number of controls. The feature extraction properties of the algorithms were corrected for minicolumnar fragments, curvature of the tissue section, and 3D proportions (stereological modeling) [24]. Later on, the same patient population was used to confirm the presence of cortical radial abnormalities in a study using the gray level index (GLI), i.e., proportional area covered by Nissl-stained to unstained elements in postmortem samples [21]. Other studies have provided evidence that the minicolumnar alterations in autism are not a nonspecific effect of mental retardation. Investigators have found that minicolumnar width in Down syndrome patients reaches adult proportions earlier than normal, possibly as a result of accelerated aging [16, 17]. In these studies minicolumnar size was reported to be normal despite the small brain size of Down syndrome patients.

An increase in the number of minicolumns is thought to underlie the neocortical expansion accompanying human encephalization, i.e., the process by which the brain has increased in size to a degree greater than expected when taking body size into account [88]. Empirical evidence and theoretical models indicate that local circuit neurons increase in number, complexity, and proportion relative to projection neurons during primate encephalization [45, 86]. These trends reflect the emergence in primates of a distinct population of dorsal telencephalic-derived inhibitory interneurons [59] modulating activity of minicolumnar pyramidal cells. Furthermore, isocortical areas such as dorsolateral prefrontal cortex, lacking direct homologs in non-primates, contain a well-defined granule cell layer of excitatory interneurons and increased numbers of supragranular local projection neurons [80]. Multiple polymorphisms associated with autism may be a consequence of phylogenetically recent changes in genetic programs guiding the development of species-specific cytoarchitectonic features. The morphometric analysis of such features complements genetic analysis. This study therefore investigates minicolumnopathy in an independent sample of autistic patients. It also expands on previous findings by studying cortical cell size and density as related to pyramidal cell arrays. The changes in these parameters, early during development, would provide for basic alterations in corticocortical connections and information processing.

Materials and methods

Clinical dataset

The diagnosis for each autistic patient was established postmortem by the Autism Tissue Program (ATP). A certified rater and trainer arranged for a postmortem visit with the family to obtain, with written consent, medical and clinical information via a questionnaire that included the Autism Diagnostic Interview-Revised (ADI-R; 63).

The Harvard Brain Tissue Resource Center (HBTRC) questionnaire was modified to include autism-specific questions for ATP use. The information obtained included: donor and respondent identifying information; ethnicity, handedness and known exposure to hazardous materials; diagnostic information including dates and physician; genetic tests; pre- and postnatal medical history; immunization, medication, and hospitalization information; family history and additional information about donor participation in any training or research studies such as imaging, medication trials, and/or genetic studies. The supporting documents such as autopsy reports, death certificates, medical, clinical, and/or educational records were obtained at the time of the home visit or by sending written requests, signed by the legal next-of-kin, to named providers.

Brain specimens

Postmortem brains (one hemisphere per case) from six autistic cases (mean interval between death and autopsy 20.0 ± 2.9 h) and from six age-matched controls (mean interval between death and autopsy 24.0 ± 11.1 h) were analyzed (Table 1). Brains were obtained from several brain banks in the USA and Germany (see Acknowledgments). All autistic patients met the DSM-IV-TR [2] and ADI-R [62] criteria for autism. None of them exhibited any chromosomal abnormalities. In all of the cases, autopsy was performed after informed consent was obtained from a relative. The use of these autopsy cases was approved by the relevant Institutional Review Boards. Clinical records were available for all cases (see Appendix).
Table 1

Clinical characteristics of the cases included in this study

Patient

Sex

Hemisphere

Age (y)

Cause of death

Clinical history

History of seizures

Medication history

Brain weight (g)

PMI (h)

Section thickness (μm)

A1

M

L

4

Drowning

Asthma/bronchitis

No

Daily medication (not specified) for asthma/bronchitis

1,160

30

200

C1

M

L

4

Myocardial infarct—Takayasu arteritis

NH

NH

NH

1,380

5

500

A2

F

L

5

Car accident

Ear infections

No

Antibiotics for ear infections

1,390

13

200

C2

F

R

4

Lymphocytic myocarditis

NH

NH

NH

1,222

21

200

A3

M

R

8

Sarcoma

Syndactyly of the fingers and feet; colitis; high fever; neutropenia; metastatic alveolar rhabdomyosarcoma; large paravertebral mass extending from chest cavity to abdomen

Abnormal EEG; not diagnosed with seizure disorder

Depakote (1 year after EEG); chemotherapy; Peridex; Nystatin; GCSF; Benadryl; Pheergan; Dexamethasone; Morphine

1,570

22

200

C3

F

R

7

Status asthmaticus

NH

NH

NH

1,350

78

500

A4

M

L

13

Seizures

Severe hypotonia; ketogenic diet for 1.5 years

Yes

Dilatin (seizures); Anticonvulsants; Trileptal (seizures); Trazadone (sleep)

1,420

26

200

C4

M

R

14

Electrocution

NH

NH

NH

1,600

20

200

A5

F

R

20

Obstructive pulmonary disease

ADHD; microcephaly; epilepsy; schizophrenia

Yes (3 times)

Various psychotropic medications including Haldol, Ritalin, and Congentin; DepoProvera (birth control); Mellaril (sleep); Zoloft

1,108

15

200

C5

M

R

23

Ruptured spleen

NH

NH

NH

1,520

6

200

A6

M

R

24

Drowning

Pneumonia; bronchitis; behavioral problems

First seizure prior to death

Quetiapine (200 mg BID); Propanolol (400 mg BID); Thioridazine (50 mg HS)

1,610

14

200

C6

M

R

25

Cardiac tamponade

NH

NH

NH

1,388

14

500

A autism, C control, M male, F female, L left, R right, NH no history, BW brain weight, PMI postmortem interval, y years, h hours

Tissue processing

After immersion-fixation in 10% formalin for at least 3 months all hemispheres were mounted with celloidin and cut into entire series of 200-μm-thick coronal sections as described in detail elsewhere [43]. Three hemispheres were cut at 500-μm thickness. These differences did not influence the results of this study, since imaging of the tissue was done at high magnification, with a depth of field much narrower than 200 μm (see below, Image capture). Every third section (in one hemisphere: every second) was stained with gallocyanin.

Brain regions

Gallocyanin-stained sections were used by three of us (van Kooten, Heinsen, and Schmitz) to identify cortical areas M1, V1, frontal association cortex, and S1 (areas 4, 17, and 9 of Brodmann [13] and area 3b of Vogt and Vogt [101], respectively) according to anatomical landmarks and cytoarchitectural criteria (Fig. 1). Gross anatomical landmarks for M1 include the anterior wall of the central sulcus and adjacent portions of the precentral gyrus. Cytoarchitectonically, the region is clearly demarcated by its giant Betz cells and minimization of layer IV [89]. Area 17 (V1) is located along the walls of the calcarine sulcus in the occipital lobe and adjacent portions of the cuneus and lingual gyrus [19]. It is defined histologically by a broad lamina IV divided into three sublayers with numerous very small pyramidal cells in layers II and III. It is noted for the dense line of Gennari in myelin stains. Area 9 lies in the superior and middle frontal gyrus. Rajkowska and Goldman-Rakic [83, 84] found that it was located in the middle third of the superior frontal gyrus in all the cases they examined. It covered both dorsolateral and dorsomedial surfaces of the gyrus and extended, in some cases, to the depth of the superior frontal gyrus and portions of the middle frontal gyrus.
https://static-content.springer.com/image/art%3A10.1007%2Fs00401-006-0085-5/MediaObjects/401_2006_85_Fig1_HTML.jpg
Fig. 1

Primary sensory cortex of an 8-year-old autistic male. Automatic segmentation has classified pixels as background, shown in black, and neurons, in gray scale. Clumps of neurons have been further separated into individual objects using the morphological watershed transform

Image capture

Regions of interest were delineated with a stereology workstation, consisting of a modified BX50 light microscope with UPlanApo objectives (Olympus, Tokyo, Japan), motorized specimen stage for automatic sampling (Ludl Electronics, Hawthorne, NY, USA), HV-C20AMP CCD color video camera (Hitachi, Tokyo, Japan), and StereoInvestigator software (MicroBrightField, Williston, VT, USA). Delineations were performed with a 10× objective (NA = 0.40). Digital micrographs each measuring about 200 μm by 150 μm were produced using the stereology workstation described above and a 40× oil objective (NA = 1.0). A few hundred such images were captured per region of interest to cover the entire cortical thickness. These images were assembled into one mosaic using the Virtual Slice module of the StereoInvestigator software. Only slight adjustments of contrast and brightness were made, without altering the data of the original materials.

Depth of field dtot of the mosaic component images can be computed according to the formula
$$ d_{{{\text{tot}}}} = {\left( {\frac{\lambda } {{{\text{NA}}}} + \frac{e} {M}} \right)} \times \frac{n} {{{\text{NA}}}}, $$
where λ is the wavelength of the illumination, M is the magnification, e is the resolution of the CCD (twice the distance between detectors), and n is the refractive index of the medium. These last two were e = 16.2 μm for the HV-C20AMP camera and n = 1.5 for oil. Taking λ on the order of 1 μm, dtot was equal to about 2.1 μm, so images represent only a small virtual slice of the 200- or 500-μm-thick sections, and differences in section thickness were thus not considered as a confound.

Computerized image analysis

Multiple techniques reduced the images to sets of descriptive parameters. The minicolumn fragment method [24] produced estimates of the mean width (CW) of minicolumns in a region of interest, the relative deviation (VCW) of minicolumnar widths about the mean, and the mean distance (MCS) between neighboring neurons within the same minicolumn fragment (the “CW” and “MCS” notation being used for consistency with earlier publications [20]). The parameters mean grain size \( {\left( {\ifmmode\expandafter\bar\else\expandafter\=\fi{A}} \right)}, \) mean grain perimeter \( {\left( {\ifmmode\expandafter\bar\else\expandafter\=\fi{U}} \right)}, \) and intensity (λ) of a Boolean spatial model [99] were fit to each image. The GLI method [91, 92] reported the mean (D) and standard deviation (sdD) of distances between ridges of high stain intensity, mean (W) and standard deviation (sdW) of the width of these ridges, and their average height (A) above background. Lastly, the distribution of distances between neighboring neurons within each image was modeled as a mixture of two distributions with means (mnear and mfar) and standard deviations (snear and sfar, respectively) with mnear < mfar. Details of each of the four methods are outlined below.

Computerized image analysis of minicolumnar structure in laminae II through VI was performed with algorithms described in the literature [24]. The feature extraction properties of the program were corrected for minicolumnar fragments, curvature of the tissue section, and have been adapted to 3D proportions (stereological modeling). The resulting estimates of minicolumnar width have also been validated against physiological measurements using intrinsic optical signaling and against anatomical measurements using myelinated fiber bundles [24].

Color mosaics were converted to intensity images and adaptively thresholded using a scale space approach [58]. Overlapping objects in the thresholded images were further separated using the watershed transformation. Each connected region in the resulting image was further classified according to size. Objects smaller than 10 μm2, assumed to be glia, neuron fragments, or noise, were discarded altogether. Of the remaining objects, those with areas above the 15th percentile were used for minicolumn fragment detection, while the smaller objects were classified as interneurons based on the fraction of all neurons estimated to be inhibitory by Braitenberg and Schüz [12].

Local neuron density was computed as the sum of equal contributions from large objects above the 15th percentile in the cross section. Each large neuron produced a density hump, peaking at the object’s centroid, with elliptical contours oriented so that the major axes were parallel to the local radial direction, i.e. along the axis of the minicolumn, at each point in an image. The ridges in the neuronal density indicated the cores of minicolumn fragments, and objects including those classified as interneurons were parcellated into clusters according to the nearest fragment core. The average distance between neighboring fragment cores was addressed as the minicolumnar width CW (Fig. 2). Standard deviation of the logarithm of these distances was the scale-independent measure variability in minicolumn width (VCW). Interneuronal distance (MCS) was the average distance between centroids of neighboring neurons within the same fragment. Only those fragments with more than ten neurons were considered when computing the means and standard deviations.
https://static-content.springer.com/image/art%3A10.1007%2Fs00401-006-0085-5/MediaObjects/401_2006_85_Fig2_HTML.jpg
Fig. 2

Minicolumns in Brodmann area 4, lamina III, in an autistic patient (bottom) and an age-matched control case (top). Insets highlight the cores of minicolumn fragments identified by our software, illustrating the reduction in minicolumnar width (CW). Scale bars measure 200 μm on left and 50 μm on right

From the adaptively thresholded images, computerized image analysis produced estimates of the relative amount of area occupied by Nissl-stained tissue (AA), the total boundary length of stained tissue per unit area (LA), and the planar convexity (NA+); these quantities are of no interest in themselves but suffice to compute the parameters of a Boolean model with convex primary grains using the method of moments [99]. Each realization of the Boolean model is the union of a number of convex random closed sets (grains)—independently and identically distributed with mean area \( \ifmmode\expandafter\bar\else\expandafter\=\fi{A} \) and mean perimeter \( \ifmmode\expandafter\bar\else\expandafter\=\fi{U} \)—located, or centered, at the points of a Poisson process with intensity λ. These three parameters, which completely determine the model, were obtained algebraically from the following equations [99], substituting for AA, LA, and NA+ their respective estimates:
$$ \begin{aligned}{} A_{A} & = 1 - {\text{e}}^{{ - \lambda \ifmmode\expandafter\bar\else\expandafter\=\fi{A}}} , \\ N^{ + }_{A} & = \lambda {\left( {1 - A_{A} } \right)}, \\ L_{A} & = N^{ + }_{A} \ifmmode\expandafter\bar\else\expandafter\=\fi{U}. \\ \end{aligned} $$

Considering that the foreground pixels of a thresholded image correspond to locations in the original image overlapped by one or more stained cell soma, \( \ifmmode\expandafter\bar\else\expandafter\=\fi{A} \) is the average area of a neuronal cross section and λ is their density, i.e. number of neuronal cross sections per unit area. This technique was used to obtain these values without any need to segment individual neurons. The method of minicolumn fragments, on the other hand, does attempt to segment individual cell somata but uses the size of segmented objects only to sort them into size classes as described above.

Thresholded images were also analyzed according to the GLI method [91, 92], slightly modified so that the GLI would vary smoothly across a region (Figs. 3, 4). Here, the inverse thresholded image, with foreground set equal to one and background set equal to zero, was convolved with a kernel measuring 11 μm by 110 μm at half-maximum. The long axis of the kernel was parallel to the axes of the minicolumns, assumed to be the image Y-axis. The result was an estimate of the GLI, or staining intensity, in the neighborhood of each pixel (Fig. 3). Each GLI image so produced was measured in profiles in the X direction to obtain the following parameters: (1) mean (D) and (2) standard deviation (sdD) in the distance between local maxima of the GLI, (3) amplitude (A) of the local maxima, or difference between peak GLI and the lowest GLI between the peak in question and the next peak in the profile, and (4) mean (W) and (v) standard deviation (sdW) in the width, i.e. full width to half maximum, of the peaks (Fig. 4).
https://static-content.springer.com/image/art%3A10.1007%2Fs00401-006-0085-5/MediaObjects/401_2006_85_Fig3_HTML.jpg
Fig. 3

Left: A 0.5 mm × 0.5 mm field from normal human primary motor cortex, lamina III. Right: Local GLI in the vicinity of each point in the field. Values are shown in gray scale from 1% (black) to 70% (white). For the significance of points X and Y, see Fig. 4. The gray level profile in this figure differs qualitatively from the example in Schleicher and Zilles [91] because we estimated GLI with a smooth, bell-shaped kernel and did not subsample the image. This allows for better localization of gray level peaks and troughs

https://static-content.springer.com/image/art%3A10.1007%2Fs00401-006-0085-5/MediaObjects/401_2006_85_Fig4_HTML.gif
Fig. 4

GLI parameters are illustrated on a profile from points X to Y in Fig. 3: amplitude A, the difference between peak GLI and average trough GLI, distance D between peaks, and full width at half amplitude W

A follow-up study considered the distance between neighboring objects. The objects were those classified as large neurons for the purposes of minicolumnar analysis (see above). Neurons were considered neighbors if (a) they were endpoints of an edge of the Delaunay triangulation of object centroids (Fig. 5), and (b) they were not further apart than the distance from either of them to the edge of the region of interest. Criterion (a) is equivalent to the geometric statement that three neurons are neighbors if the circle drawn through their centroids does not enclose the centroid of any other neuron. The criterion (b) is necessary to correct for a boundary effect where distant neurons are incorrectly labeled as neighbors because the true neighbors of one or more of them lie outside the region of interest. Now given that objects in a single lamina were clustered, as cells in minicolumns, an object’s neighbors may include the members of the same cluster or the members of a nearby cluster (Fig. 6). Accordingly, the distribution of distances between neighbors (d) would then be a mixture of intracluster or near distances, and intercluster or far distances. This was modeled as a two-component lognormal mixture distribution; that is to say, the log d were assumed to be drawn from either of two Gaussian distributions of near and far distances:
$$ \log \,d \sim \alpha N{\left( {\mu _{{{\text{near}}}} ,\sigma ^{2}_{{{\text{near}}}} } \right)} + {\left( {1 - \alpha } \right)}N{\left( {\mu _{{{\text{far}}}} ,\sigma ^{2}_{{{\text{far}}}} } \right)};\;0 \le \alpha \le 1. $$
https://static-content.springer.com/image/art%3A10.1007%2Fs00401-006-0085-5/MediaObjects/401_2006_85_Fig5_HTML.jpg
Fig. 5

Delaunay triangulation of larger neurons (gray scale) in the primary sensory cortex of an 8-year-old autistic male. Given a point set in the plane, neuron centroids in this case, the Delaunay triangulation is defined such that three points form the vertices of a triangle if and only if the circle through those points has no point from the set in its interior

https://static-content.springer.com/image/art%3A10.1007%2Fs00401-006-0085-5/MediaObjects/401_2006_85_Fig6_HTML.gif
Fig. 6

a Owing to the clustering of neurons within minicolumns, the triangulation will include edges between objects in the same cluster (solid lines) and edges between objects in other clusters (dotted lines). The distribution of edge lengths over the whole graph will be a mixture of these two subgraphs’ distributions. b We have estimated the locations and scales of distances within and between clusters by modeling the logarithm of edge lengths as a two-component Gaussian mixture (see text for details)

Nucleolar size

Nucleoli were identified visually in digitized images. Twelve random locations were selected, per mosaic, by computer. The user identified the nucleolus nearest to each random point, excepting those points that did not fall within laminae II through VI. Nucleoli were segmented from the surrounding cytoplasm by thresholding according to Otsu’s [73] method. If the thresholded nucleolus appeared to touch another object, the user manually removed pixels from that object until the nucleolus was separate. Nucleolar cross sections were measured by counting pixels in the thresholded nucleoli.

Statistical analysis

Autistic and control cases were sorted into age-matched pairs for the purposes of statistical analysis. Wilcoxon signed-rank tests were used to verify the absence of pairwise differences in brain weight or postmortem interval. Both tests were insignificant with P = 0.84. Otherwise, statistical analysis for all measurements employed repeated measures analysis of variance. Case pair (1–6) was the random factor, and fixed factors included diagnosis, sex, cortical area, and hemisphere. Interactions of the main effects involving both sex and cerebral hemisphere were excluded from the model due to multicolinearity. For the minicolumn fragment data only, a follow-up test was performed using a second model comprising the original model plus a factor for cortical lamina: III or V/VI. The model fitting and inferential statistics were performed with Matlab (The MathWorks, Natick, MA, USA).

Results

Minicolumnar width (CW) was greater in controls than in autistic persons by 1.46 μm or 5.54% of the mean [Student t = 2.466 with 19 degrees of freedom (df); P = 0.0234] (Table 2). CW also varied with cortical area (F = 17.50 with 3 numerator df and 19 denominator df; P < 0.0001), but no other effects were statistically significant. MCS exhibited a sex dependence in that the difference between control and autistic cases was greater in females than in males (t = 2.227 with 19 df; P = 0.0382); mean MCS was 18.3 and 18.9 μm in autistic females and males, respectively, and 19.7 and 18.6 μm in normal females and males, respectively. Again the only other significant effect was cortical area (F = 16.09 with 3 numerator df and 19 denominator df; P < 0.0001) (Table 2). There were no significant findings in VCW. The follow-up analysis on those minicolumn fragments lying outside of lamina IV, using the supplemental model with lamina included as a factor, found no significant dependence of any measurement on lamina or lamina by diagnosis interaction.
Table 2

Average minicolumnar morphometry, broken down by diagnosis and cortical area

Area

CW (μm)

MCS (μm)

VCW (%)

Autistic

Control

Autistic

Control

Autistic

Control

3

25.8

27.0

18.4

18.7

15.1

14.8

4

27.5

28.2

19.9

20.0

14.2

14.6

9

26.5

29.4

19.1

20.7

15.0

14.9

17

23.0

24.1

17.0

17.3

14.4

14.7

sd

1.2

0.8

0.9

Scale-independent variation in minicolumnar width VCW is expressed as percentage of the mean

CW minicolumnar width; MCS mean cell spacing (interneuronal distance); VCW variability in minicolumnar distance; sd within-group standard deviation

Simultaneous multivariate analysis of neuron density and neuron profile area and perimeter (Table 3) revealed significant diagnosis dependence (F = 5.47 with 3 numerator df and 17 denominator df; P = 0.0081) and diagnosis by cortical area interaction (F = 3.65 with 9 numerator df and 41.5 denominator df; P = 0.0020). Mean particle cross section was 30.5 μm2 less in autistic cases compared to controls (t = 3.804 with 19 df, P = 0.0012), while particle density was greater in autism: 5.15 × 103 mm−2 versus 4.19 × 103 mm-2 in controls (t = -2.723 with 19 df; P = 0.0135).
Table 3

Estimated Boolean spatial model parameters of the texture of Nissl-stained material from each cortical area examined, in autistic and control individuals

 

Area

\( {\left( {\ifmmode\expandafter\bar\else\expandafter\=\fi{A}} \right)} \)(μm2)

\( \ifmmode\expandafter\bar\else\expandafter\=\fi{U} \) (μm)

λ (μm−2)

Autistic

3

128.0

44.2

0.0047

4

144.9

45.7

0.0047

9

131.9

44.5

0.0051

17

114.8

44.8

0.0062

Control

3

151.1

49.4

0.0042

4

154.7

48.3

0.0043

9

182.3

51.8

0.0038

17

153.4

53.3

0.0045

Standard deviation (within groups)

13.0

2.8

0.0008

\( {\left( {\ifmmode\expandafter\bar\else\expandafter\=\fi{A}} \right)}, \) mean soma cross section; \( \ifmmode\expandafter\bar\else\expandafter\=\fi{U}, \) mean soma perimeter; λ, neuronal density

Each GLI parameter (Table 4) was reduced in autism: D by 1.93 μm (t = 2.8465 with 19 df; P = 0.0103), sdD by 0.96 μm (t = 2.1402 with 19 df; P = 0.0455), A by 2.2% (t = 2.226 with 19 df; P = 0.0383), W by 0.86 μm (t = 2.6936 with 19 df; P = 0.0144), and sdW by 0.59 μm (t = 2.2643 with 19 df; P = 0.0354).
Table 4

Inhomogeneities in the gray level index, broken down by diagnosis and cortical area

 

Area

A (%)

D (μm)

W (μm)

sdD (μm)

sdW (μm)

Autistic

3

16.0

32.1

15.2

13.1

6.1

4

18.2

33.7

16.2

13.5

6.5

9

16.9

32.3

15.4

12.8

5.9

17

13.4

29.0

14.1

10.9

5.0

Control

3

17.4

33.3

15.8

13.5

6.3

4

19.1

35.2

16.7

14.3

7.0

9

19.2

35.0

16.7

14.0

6.8

17

17.7

31.3

15.1

12.4

5.8

Standard deviation (within groups)

1.9

1.3

0.6

0.9

0.5

A, peak GLI above background; D, mean distance between peaks; W, mean peak width; sdD, standard deviation of distance between peaks; sdW, standard deviation of peak width

Multivariate ANOVA, with dependent variable being the estimated mean logarithmic neighbor distances μnear and μfar (Table 5), found the diagnosis effect significant (F = 3.72 with 2 numerator df and 18 denominator df; P = 0.044). Near distances in autistic cases, averaged across all four areas, were 98.0% of those of controls [confidence interval (94.8%,101.0%)], and far distances were 94.8% of those of the normal comparison group [confidence interval (93.0%, 96.7%)]. Diagnosis by cortical area interaction was statistically insignificant.
Table 5

Average modes mnear = exp(μnear) and mfar = exp(μfar) of the two components of the neighborhood distance distribution, broken down by cortical area and diagnosis

Area

mnear (μm)

mfar (μm)

Autistic

Control

Autistic

Control

3

15.2

15.5

25.6

26.1

4

17.3

17.1

27.6

28.9

9

18.1

17.8

27.6

30.4

17

14.9

16.2

22.9

24.1

Within-group standard deviation is 7.1% of the mean

Nucleolar cross sections (Table 6) were converted to radii of circles with the same area (requiv) in order to stabilize the variance for statistical analysis. Univariate ANOVA of mean (within each mosaic) requiv, again with the same model, revealed significant differences between autism and normal patients (F = 19.52 with 1 numerator and 17 denominator df; P = 0.0004). While requiv varied with cortical area (F = 5.28 with 3 numerator and 17 denominator df; P = 0.0093), there was no significant interaction with diagnostic group (F = 0.89 with 3 numerator and 17 denominator df; P = 0.465).
Table 6

Mean nucleolar equivalent radius, i.e. the radius of a circle with area equal to the nucleolar cross section, by diagnosis and cortical area

Area

requiv (μm)

Autistic

Control

3

2.27

3.19

4

2.24

3.08

9

2.56

3.19

17

2.03

2.46

Within-group standard deviation is 0.31 μm

Discussion

The results of this study indicate that, within a limited brain sample, minicolumnar morphometry varies according to cortical region. Since traditional parcellation schemes have relied on subjective appraisals, the semi automated method used in this study may offer a rapid and accurate procedure to define cytoarchitectural brain regions. Results also corroborate a reduction in the width of minicolumns in the brains of autistic patients, using two different methods: one that measures pyramidal cell arrays [24] and another using the GLI [91, 92]. The Wilcoxon signed-rank test indicated essentially no (monotonic) difference between autistic patients and controls with respect to brain weight (P = 0.84) or postmortem interval (P = 0.84). The statistical design for all analyses included sex and cerebral hemisphere as fixed factors. Statistically significant, diagnosis-dependent differences between groups correspond, therefore, to the effect of diagnostic category independent of those other factors.

Given the lack of significant differences in brain weights between our comparison groups (Table 1), a generalized reduction in minicolumnar width in the autistic patients suggests their increase in total numbers. Although minicolumnar width was measured as distance between cell arrays in both the present and previous studies, differences in section thickness, tissue processing, and microscope objectives prevent us from making absolute comparisons. The feature extraction routines of our algorithms measure vertical clusters of neuronal profiles whose axis are parallel to the plane of tissue section. Depending on the depth of field and focus of an objective, the use of thicker sections allows for a substantial overlap of neighboring minicolumns (see Image capture). The result is a relative diminution in minicolumnar width for both patients and controls when comparing the present study to previous ones [20].

Our initial finding, now corroborated in this study, is that the brains of the autistic patients have minicolumns of reduced width and consequently of increased numbers per linear length of distance [20]. Minicolumnar width varied with cortical area but no other effects were statistically significant. Of the four sampled areas (areas 3, 4, 9, and 17), the dorsolateral prefrontal cortex showed the greatest reduction in minicolumnar width when comparing autistics and controls (Table 2). Since the lateral or granular prefrontal cortex is apparently unique to primates [82], the finding could have important implications regarding putative animal models for autism, especially in rodents. Further topographical inferences regarding a minicolumnopathy in autism would require a larger patient population and a greater number of regions of interest.

The presence of supernumerary minicolumns is said to account for the process of cortical expansion during encephalization [88]. Cortical expansion necessitates an increased number of neurons, but not proportionally [26, 47]. Although the human brain, largely the neocortex, is three times the size of the chimpanzee brain, there is only a 25% increase in the total number of neurons [47]. In our study evidence suggestive of an increased number of minicolumns per linear length in the brains of autistic individuals led us to examine neuronal density. A point process model indicated increased neuronal density but failed to discern the basis for the same, e.g., the presence of supernumerary minicolumns, an increase in the total number of cells per minicolumn, or both. A subsequent analysis based on a Delaunay triangulation addressed the aforementioned possibilities. The new algorithm indicated significant differences in the edges of intercluster distances but not within intracluster distances. Minicolumns appeared to be packed closer together in autism (reduction in far distances) but the total number of cells per minicolumn (near distances) appeared normal. Finally, neurons within the minicolumns of the autistic patients were smaller in size and had a reduction in the size of their nucleoli. In light of the small sample size, the fact that all the various measurements revealed statistically significant differences is surprising. There is a strong possibility, then, that the true effect size associated with some of these measurements is greater than we have observed. Further studies with a larger sample may put tighter bounds on the magnitude of these differences between autistic patients and controls. The following paragraphs discuss the previously mentioned results from the perspective of both neuropathology and possible clinical correlates.

Minicolumns (development and numbers)

Cell arrays are the first radially aligned structure appreciable during development. These modules comprise layer V pyramidal cells (whose dendrites form a bundle extending through layer II), clusters of layer II and III cells, as well as associated interneurons [78]. They have been investigated in a number of species, including barrel field cortex of the mouse [103], barrel field cortex of the rat [75], area 17 of cat [77], visual cortex of non-human primates [14, 76, 78, 79], and various human cortical areas [11, 94, 95]. Pyramidal cell arrays have been the focus of investigation in the studies of columnar development in a series of fetal and adult postmortem tissues. Krmpotić-Nemanić et al. [54] described the development of cell arrays in human fetal and postnatal auditory cortex in a series from celloidin-embedded postmortem material with dates ranging from 9 weeks gestational to 3 months postnatal. They identified continuity in the development of ontogenetic columns into pyramidal arrays with regional differences in the elaboration of the basic columnar structure. Pyramidal cell arrays therefore derive from the ontogenetic cell column and provide the matrix around which growing axonal and dendritic processes are organized. In both monkeys and humans, most of the founder cell divisions that account for the number of cortical columns occur before embryonic day 40 [85, 87]. The genesis for an increase in the total number of minicolumns in autism would therefore transpire at an early stage of gestation.

Minicolumnar size

The possible significance of smaller minicolumns in autistic patients can be gleaned from the results obtained in the comparative studies of columns in cortical area 17 (V1) [31, 7479]. In primates, the primary visual cortex contains small minicolumns when compared to many other mammals. This is the case even though the brain size of primates is many times greater than the comparison species. Studies using uniformed section thickness in rhesus monkeys have reported minicolumns to be about 23–31 μm on the basis of apical dendrite bundles. By contrast they range from 50 to 60 μm in other small-brained mammals [15, 18]. Researchers have concluded that the narrow minicolumns reflect the increased processing complexity of primate vision and interpreted the smaller minicolumns as being a more complex, interconnected system. In effect, reduced minicolumnar spacing may provide for an increased overlap of their dendritic trees making the function of neighboring minicolumns more interdependent [96, 97].

Cell size and number

Studies estimating cell counts and/or describing cytoarchitectural features in autism have been few in number and have revealed no consistent findings. Aarkrog [1] described “some cell increase” in a frontal lobe biopsy. Coleman et al. [28] found significant differences in six out of 42 comparisons when studying the brain of a 21-year-old autistic female and two controls. Kemper and Bauman [52] used qualitative means to describe disturbed lamination in the anterior cingulate gyrus in five out of six autistic patients. No abnormalities in cell counts were reported by Bailey et al. [4] in their six autistic cases. Previous results by Casanova et al. [20] indicate that the brains of autistic patients have smaller and more abundant minicolumns. The tighter packing of cortical modules suggested increased cellular density. A more recent case study examined three cortical areas in nine autistic patients and 11 controls [21]. The overall mean GLI in this series was 19.4% for the control group and 18.7% for the autistic group (P = 0.724) with diagnosis-dependent effects in D. The latter authors concluded that in autism a normal GLI and an increased number of modules indicate a reduction in the total number of cells per minicolumn.

In the present study the overall GLI did not differ significantly between autistic patients and controls. The differences in gray level inhomogeneity as described by D and W (Fig. 4) corroborate previous findings that minicolumnar width is reduced in autism [20, 21]. Significant differences in sdD and sdW, together with the lack of significant differences in scale-independent variability in minicolumnar width, imply a direct proportionality between the mean and standard deviation of minicolumnar widths within a cortical area.

Instead of cell loss the present study proposes an alternate explanation to the preservation of GLI and increased number of minicolumns: a reduction in neuronal size. Our results indicate the presence of diminished neuronal cell size and increased density in the brains of autistic patients. These findings varied according to brain region (diagnosis by cortical area interaction). The results counter the general notion that, with the exception of striate cortex, cell density is similar across different cortical areas and even among species [90]. The impression of cellular uniformity in cortical columns has been refuted by using modern unbiased techniques [7, 8].

The neurons in our study sample had well-defined nucleoli that were easily distinguishable from other components of the karyoplasm. Histological studies have characterized the nucleolus as an RNA organelle whose function is to regulate protein synthesis. As such the volume of the nucleolus is a constitutional factor adjusted to match the basal metabolic requirements of a cell [66]. The exception to this rule appears to be fast spiking neurons where increased metabolism appears coupled to smaller neuronal size and nucleoli [3]. Pathological conditions causing either increased or decreased cellular activity have an impact on pyramidal cells’ nucleolar size in accordance with their protein synthesis requirements [66]. Studies of this and similar morphometric indices have illustrated their utility in conditions such as Alzheimer’s disease [6466] and schizophrenia [29, 61]. In autism a significant diminution in nucleolar size, after thresholding interneurons, suggests a corresponding reduction in the neuronal metabolic activity/efficiency.

Clinical correlates

Brain growth causes the isolation of non-adjacent neurons by expanding their intervening distance. With longer distances the presence of smaller neurons imposes a metabolic constraint on connectivity. Longer connections necessitate larger and more active neurons, where each cell is networked into a dynamically controlled energy system [27, 56]. A cortex biased towards smaller neurons would facilitate signal delays and metabolic inefficiencies when linking disparate brain areas. The result would be reduced or impaired functional connectivity between distant cortical regions [9, 48, 50]. In such a system both sensory coding and motor output, the endpoints of networked chain translating sensory information into behavioral actions, are normal. It is the intervening step of information analysis that is abnormal.

The presence of smaller neurons in autism may help explain the fact that autism spectrum disorder (ASD) patients normally activate the fusiform gyrus when viewing faces as compared to non-face stimuli [39]. The data indicate that the face-processing deficits encountered in ASD are not due to dysfunction of an individual area, in this case the fusiform gyrus, but due to more complex anomalies in the distributed network of interconnected brain regions [10, 39, 81]. In effect, the level of synchronization during activation tasks involving distant brain regions suggests impaired connectivity in autism [25, 48, 50, 53, 104]. It may be, as Minshew et al. [69] have suggested, that autism is manifested as abnormalities in high-level tasks whenever an elevated demand is placed on information processing [33, 70].

Just et al. [50] have subsumed the evidence for a lower degree of information integration in autism under the rubric of the “underconnectivity theory.” However, the term may prove to be a misnomer when applied to shorter intra-areal connections (arcuate or u fibers). In autism, an increase in the total number of minicolumns requires a scale increase (roughly a 3/2 power law) in white matter to maintain modular interconnectivity [46]. This additional white matter takes the form of short-range connections which makes up the bulk of intracortical connections [23]. Recent structural imaging studies suggest that short association fibers are overrepresented in autism [44]. This fact may help explain the superior abilities of the autistic patients when performing tasks that require local information processing [41]. A diminution in neuronal cell size and a concomitant increase in the total number of minicolumns biases cortical connectivity in favor of local rather than global information processing [6]. The result is a “hyper-specific brain” [36], where segments of perception are retained at the expense of losing the “big picture” [98]. In autism, a computational perspective validates this framework and sees hyperspecificity as a possible framework of neural codes in charge of elaborating concepts [67].

It is noteworthy that an adaptive strategy for increasing the metabolic efficiency of a system driven by smaller neurons is to increase their total cell numbers while reducing the activity of each neuron [60]. This approach, called sparse distributed coding, achieves high representational capacity by distributing small amounts of activity over a large population of neurons [57]. In sparse coding, neurons have the potential to respond strongly to focal features of sensory inputs [32, 40]. Since neurons rarely encounter their feature, they will fire in short bursts, sparingly distributed in time [57]. Sparse-distributed coding is characteristic of the visual system [100, 102]. A brain whose neuronal population is biased towards small neurons and corresponding metabolic exigencies would therefore create and execute strategies that emphasize selective convergence of information among closely adjacent modules, as e.g. in the visual system.

Large system operations can be subdivided into task-specific modules where information processing proceeds along hierarchical stages [49]. Metabolic constraints facilitate the connectivity of closely adjacent cortical areas [35, 68] which represent similar features of perception [93]. The analogy is to an operon where related genes (e.g., coding for enzymes in the same metabolic pathway) cluster together so that outside influences can provide for simultaneous negative or positive control to all of its units. This type of assembly is efficient because perceptions commonly “overlap with one another, sharing parts which continue unchanged from one moment to another” [5]. By way of contrast information at higher echelons emphasizes complex conjunctions of perceptual attributes.

When comparing the hierarchy of perceptual networks, primary association cortices, modules become larger and representation more categorical [34, 42]. As a consequence, information at lower echelons of the perceptual hierarchy is more localized and lesions result in concrete deficits. At higher levels, information is more dependent on distributed networks. One good example of a physiological characteristic dependent on a distributed network is face recognition, i.e., nodes in temporal, fusiform, and prepiriform cortex [34]. Another example of greater relevance to autism is joint attention. More so than face processing, multiple studies relate joint attention to diagnosis and outcome in autism [72]. Development of joint attention involves the prefrontal cortical areas 8 and 9 and anterior cingulate area 24. These areas serve to integrate self-monitored information about social intentions with information about goal-related behavior in other people [55] processed in the parietal and temporal lobe. This so-called “social executive process” [71] may be especially prone to disturbance wherever an increase in number and proportion of small neurons facilitate the integration of information within a given region while attenuating distal connectivity and coherence.

Acknowledgments

We are grateful to the following institutions for the provision of the specimens: Mount Sinai School of Medicine (New York, NY, USA), University of Würzburg Morphological Brain Research Unit (Würzburg, Germany), University of Maryland Brain and Tissue Bank for Developmental Disorders (Baltimore, MD, USA), Harvard Brain Tissue Research Center (Belmont, MA, USA), New York State Institute for Basic Research in Developmental Disabilities (Staten Island, NY, USA) and the US Autism Tissue Program (Princeton, NJ, USA). We thank E.K. Broschk and A. Bahrkel for their technical support. This article is based upon work supported by the Stanley Medical Research Foundation (H.H., C.S., P.R.H. and M.F.C.), the Korczak Foundation (H.v.E.), the National Alliance for Autism Research (C.S. P.R.H. and M.F.C.), the McDonnell Foundation (P.R.H.) and NIMH grants MH61606 (M.F.C.), MH62654 (M.F.C.), MH69991 (M.F.C), NIH MH66392 (P.R.H).

Copyright information

© Springer-Verlag 2006