Effects of insertion conditions on tissue strain and vascular damage during neuroprosthetic device insertion

INSTITUTE OF PHYSICS PUBLISHING JOURNAL OF NEURAL ENGINEERING

J. Neural Eng. 3 (2006) 196–207 doi:10.1088/1741-2560/3/3/002

Effects of insertion conditions on tissuestrain and vascular damage duringneuroprosthetic device insertionC S Bjornsson1,5, S J Oh2, Y A Al-Kofahi3, Y J Lim4, K L Smith1,J N Turner1, S De4, B Roysam3, W Shain1 and S J Kim2

1 Laboratory of Nervous System Disorders, Wadsworth Center, New York State Department of Health,Albany, NY 12201-0509, USA2 Nano Bioelectronics & Systems Research Center, Bldg. #104, Seoul National University, San 56-1,Shillim-dong, Kwanak-gu, Seoul 151-742, Korea3 Department of Electrical, Computer and Systems Engineering, Rensselaer Polytechnic Institute, Troy,NY 12180-3590, USA4 Department of Mechanical Engineering, Rensselaer Polytechnic Institute, Troy, NY 12180-3590, USA

Received 20 January 2006Accepted for publication 12 May 2006Published 21 June 2006Online at stacks.iop.org/JNE/3/196

AbstractLong-term integration of neuroprosthetic devices is challenged by reactive responses thatcompromise the brain–device interface. The contribution of physical insertion parameters toimmediate damage is not well described. We have developed an ex vivo preparation to capturereal-time images of tissue deformation during device insertion using thick tissue slices fromrat brains prepared with fluorescently labeled vasculature. Qualitative and quantitativeassessments of damage were made for insertions using devices with different tip shapesinserted at different speeds. Direct damage to the vasculature included severing, rupturing anddragging, and was often observed several hundred micrometers from the insertion site. Slowerinsertions generally resulted in more vascular damage. Cortical surface features greatlyaffected insertion success; insertions attempted through pial blood vessels resulted in severetissue compression. Automated image analysis techniques were developed to quantify tissuedeformation and calculate mean effective strain. Quantitative measures demonstrated that,within the range of experimental conditions studied, faster insertion of sharp devices resultedin lower mean effective strain. Variability within each insertion condition indicates thatmultiple biological factors may influence insertion success. Multiple biological factors maycontribute to tissue distortion, thus a wide variability was observed among insertions madeunder the same conditions.

M This article features online multimedia enhancements

1. Introduction

Many techniques developed to study brain function or to treatneurological disorders involve inserting devices into the brain,yet very little is known about the immediate damage producedby these insertions.

5 Author to whom any correspondence should be addressed.

Refinements in neuroprosthetic design and fabricationhave resulted in devices with multiple electrodes capable ofrecording large volumes of information [1–6], stimulatingdiscrete brain regions to alleviate symptoms of neurologicaldisorders [7–9], or restoring peripheral sensation [10, 11]. Thepotential these devices hold is presently limited by biologicalreactive responses to their insertion and chronic presence inthe brain [6, 12–14]. Reactive responses are triggered bydevice insertion and culminate weeks later in the formation

1741-2560/06/030196+12$30.00 © 2006 IOP Publishing Ltd Printed in the UK 196

Effects of insertion conditions on tissue strain and vascular damage

and persistence of a fibrous and cellular sheath encapsulatingthe device that impedes recording and stimulation of neuronalactivity. The early events leading to sustained sheath formationstem from damage to both the nervous parenchyma and theneurovasculature during the initial insertion [15, 16].

Micromachined silicon devices inserted into the brainencounter the same difficulties as conventional glassmicroelectrodes and microwires [17]. Device volume,geometry and rate of insertion may all affect the abilityto penetrate and travel through the cortex with minimaldisruption. While it has been shown that the amountof insertion-related damage is influenced by device size[15], the role of other physical factors including devicegeometry and insertion speed are unknown. We propose thatinsertion conditions can be established that will produce lessdamage, thereby minimizing both initial brain damage and theconsequent reactive responses.

Computer modeling of imaging-based brain simulations[18, 19], forces acting on the surface of fresh brains [17, 19,20] or forces generated during insertion into soft materialsincluding agarose [21] provide some insight into the strainsintroduced by device insertion. However, these typicallydo not incorporate the heterogeneous composition of neuraltissue, including an analog to the pia. Both of these tissuecharacteristics will have considerable impact on the dynamicsof insertion and tissue deformation. Studies using biologicalmaterial have typically focused on surface deformation and theforces generated during insertion [17]; however, these oftendo not directly address damage and strain occurring within thecortex itself.

The neurovasculature contributes greatly to theheterogeneity of nervous tissue and of the pial surface. Thesimple pia consists of one or several layers of fibroblasts(leptomeningeal cells) separated from the underlying nervoustissue by a thin basal lamina [22, 23]. In the rat, the piais continuous with cells of the arachnoid and the subduralmesothelium over much of the cerebral cortex [24]. Arteriesand veins travel across the pial surface to supply the cortexand deeper brain structures. Large conducting arteries(!700 µm diameter in humans) branch into distributingarteries (!200 µm diameter) that in turn give rise to precorticalarteries (!60 µm diameter) that eventually penetrate the cortexperpendicular to the surface [25]. Penetrating intracorticalarteries supply superficial, middle, or deeper regions of cortex,organized in a regular pattern that reflects their ontogeny [26].Unlike intracortical arteries, larger intercortical arteries passthrough the cortex to supply deeper brain structures. All ofthese vessels are composed of thick connective tissue andsmooth muscle layers. Intracortical arteries branch laterallyto form primary arterioles, and these eventually lead to thin-walled capillaries 4–5 µm in diameter. Veins travel backthrough the cortex and across the pia to reach the dural sinuses,and are characterized in part by thinner walls relative to arteriesof the same diameter. Whereas gray matter itself possesses arelatively homogenous composition, the neurovasculature thatcourses through it varies considerably in terms of thicknessand strength.

In order to assess the contribution of tip geometry andinsertion speed to the damage caused during device insertion,

we have developed an ex vivo insertion procedure that allowsus to track the movement of labeled vascular and cellularelements using real-time video microscopy. Analysis ofthe resulting images enabled us to observe general insertioneffects including surface compression, tissue distortion, aswell as specific damage to the neurovasculature that mayhave profound repercussions on the surrounding neural tissue.Qualitative analysis revealed that in general, faster insertionsresulted in less damage to the neurovasculature. In order touse quantitative analytical methods, we developed algorithmsfor automated image analysis to permit tracking of up to100 interest points within each image sequence during a singleinsertion. These measures permitted calculations of tissuedeformation and maximum and effective strains resulting frominsertion. Within the range of speeds and device geometriesstudied, faster insertion of sharp devices resulted in lower meaneffective strain in superficial and middle regions of cortex. Theresults of this study will help guide future design of neuralprosthetic devices and enable the establishment of optimalinsertion conditions.

2. Materials and methods

2.1. Probe fabrication

Model silicon prosthetic devices were fabricated from "1 0 0#oriented, p-type 4$$ silicon wafer polished on both the frontand back. The overall fabrication process combined dry andwet etching techniques for silicon micromachining. Briefly,a photoresist was deposited and patterned to permit deepreactive-ion silicon etching to a depth of 60 µm. The etchdepth determines the final shank thickness of the devices.To protect the sidewall of the shaft from anisotropic KOHetchant, silicon nitride film was deposited onto the patternedside. In addition to the thin film, a Teflon R% shield was usedto protect one side from the KOH etching. This Teflon R%

shield containing the wafer was immersed in a 30% wt KOHsolution at 65 &C. The etch rate was approximately 0.55 µmmin'1. The silicon on the backside of the wafer was etcheduntil the structure was shown. Finally, the shield was peeledoff and then the silicon nitride film used as a hard mask for thefront side was removed by reactive-ion etching. All devicesused for this experiment had a single 2 mm long shaft, and a60 ( 100 µm cross-section. Three different tip shapes wereused for this study (figure 1). Sharp tips had a 5& interior angle,medium tips had a 90& interior angle, and blunt tips had a 150&

interior angle.

2.2. Ex vivo insertions

2.2.1. Animals. P21-30 male Sprague–Dawley rats(SAMTACO Bio Korea, Inc.) were housed at standardtemperature in a light-controlled environment, with ad libitumaccess to food and water. All protocols were approved by theInstitutional Animal Care and Use Committee (IACUC).

2.2.2. Tissue slice preparation. Animals were anesthetizedwith urethane at a dosage of 1 g kg'1 body weight.

197

C S Bjornsson et al

(A)

(B)

Figure 1. Experimental design. (A) Slices were adhered to a culture dish using fibrin glue and held in place by a nylon mesh suspendedfrom a crescent of sliver wire (inset). (B) Microfabricated silicon devices were 2 mm long with a 60 ( 100 µm cross-section. Sharp,medium and blunt tips had interior angles of 5&, 90& or 150&, respectively. Devices were inserted at fast (2000 µm s'1), medium(500 µm s'1), or slow (125 µm s'1) speeds. Scale bar is 1 mm (B; top row) or 200 µm (B; bottom row).(This figure is in colour only in the electronic version)

Transcardial perfusion of cold (5–8 &C) 50 ml fresh externalsolution (125.0 mM NaCl, 2.50 mM KCl, 1.25 mM Na2HPO4,2.00 mM CaCl2, 1.00 mM MgCl2, 25.0 mM NaHCO3,25.0 mM dextrose) preceded injection of vascular fluorescentmarkers consisting of 3 ml external solution containing 100 mgFITC-labeled 2 ( 106 MW dextran (Sigma-Aldrich, St Louis,MO) and 30 µl of 1 µm diameter Fluobeads with 488 nmexcitation and 560 nm emission (Molecular Probes, Eugene,OR). Brains were harvested following decapitation, blocked,and cut into 500 µm-thick coronal sections using a Vibratome1500 (Vibratome, St Louis, MO). Sectioning was performedin cold slicing buffer (85.0 mM NaCl, 2.50 mM KCl, 1.25 mMNaH2PO4, 0.50 mM CaCl2, 4.00 mM MgCl2, 25.0 mMNaHCO3, 25.0 mM dextrose, 75.0 mM sucrose) equilibratedto 95% O2/5% CO2. Slices were kept in a custom-builtholding chamber for 30 min to 6 h prior to insertion, at roomtemperature, in a 1:1 mixture of external and slicing solutionssaturated with 95% O2/5% CO2.

2.2.3. Device insertion and image acquisition. Brain sliceswere placed in a coverslip-bottomed 60 mm diameter Petridish and held in place by a fine nylon mesh supported by aU-shaped flattened 10-gauge silver wire placed on top ofthe slice (figure 1). Slices were immersed in patch solutionsaturated with 95% O2/5% CO2 under steady flow conditions.Slices were firmly attached to the coverslip using Tisseel fibringlue (Baxter, Inc., Vienna, Austria).

At the beginning of each day, devices were glued to theinsertion arm using cyanoacrylate adhesive (Elmer’s ProductsInc., Columbus, OH). The device was oriented parallel to thelong axis of the insertion arm using a stereomicroscope withalignment guides fitted in the ocular lens. The device was

positioned 80–100 µm above the dish floor, and leveled byensuring the tip and base of the shaft were both in focus.

The custom Petri dish was placed on an Axiovert 100Mfluorescence microscope (Carl Zeiss Inc., Germany) fittedwith a 10( objective and an AxioCam HRm camera (CarlZeiss Inc., Germany). The focal plane for all experimentswas typically 30 µm above the slice bottom. Images werecaptured with frame rates of approximately 3.3 frames s'1.For each insertion, separate videos of device insertion andsubsequent retraction were recorded. The size of each imagewas 1040 pixels high by 1388 pixels wide, corresponding to1061 ( 1416 µm. For qualitative analysis, insertion videoswere scored on the basis of evidence of vascular displacement,rupture, severing, or dragging across the entire field of view.

2.3. Image analysis: measuring Pia deformation

Real-time insertion images were opened using ImageJsoftware [27] and a feature on the pial surface where the devicewas inserted was tracked through successive frames using theManual Tracking plugin. Tracking stopped when no furtherpositive displacement (compression) was observed. The totaldisplacement along the axis of insertion was then calculatedfor each insertion.

2.4. Image analysis: measuring tissue deformation

The image analysis approach was based on identifying a set ofdistinctive points in each video frame, termed ‘interest points’.These points were tracked from one frame to the next togenerate trajectories. These points and tracks were filteredto eliminate outliers, and the deformation field was estimated.

198

Effects of insertion conditions on tissue strain and vascular damage

For quantitative analysis, an area of size 1000 pixels highby 1100 pixels wide was used, from row 20 to row 1020 andfrom column 15 to column 1115.

2.4.1. Extracting interest points: Harris detector. Pointsin an image, denoted I, at which the intensity value changesabruptly in more than one direction are useful as landmarksfor image matching [28]. These points are called ‘interestpoints’ [29]. The Harris corner detector [30] is widely usedfor detecting interest points. It is based on calculating theautocorrelation matrix M for each pixel using a small windowof surrounding pixels W:

M =

!

"#

W

$!I!x

%2 ) w#

W

$!I!x

• !I!y

%) w

#W

$!I!x

• !I!y

%) w

#W

$!I!y

%2 ) w

&

'

where the sum is over W,w is a Gaussian kernel used forsmoothing the image to avoid corners due to image noise, and) stands for the convolution operation. If at a certain point thetwo eigenvalues of the matrix are large, then a small motion inany direction will cause a significant change in intensity. Thisindicates that the point is a corner. For a given pixel (i, j), thecorner response proposed by Harris is given by

Ri,j = det(Mi,j ) ' k(trace(Mi,j ))2

where k is a parameter set to 0.04–0.06. R is the responsematrix and its size equals the image size, which means thatevery point (pixel) in the image has its corner response. Interestpoints (corners) were defined as local maxima of the cornerresponses matrix R. For this purpose, the image was dividedinto 100 ‘parent’ windows (100 pixels high by 110 pixels wide)that were further subdivided into four sub-windows (50 pixelshigh by 55 pixels wide), for a total of 400 sub-windows. Thelocal maximum values of Harris responses were identifiedwithin each sub-window to generate 400 ‘children’ interestpoints scattered across the image.

2.4.2. Tracking interest points. Interest points were assumedto move independently of one another, even though theyrepresent different regions of the same neurovascular network.This assumption was important to permit identification andremoval of outliers by comparing tracking results to resultsfrom neighboring interest points. Groups of four interestpoints were filtered as described in section 2.4.5, and outlierswere rejected on the basis of significant deviation from themedian deformation of the four interest points. In this way,specific damage events such as dragging of blood vessels bythe device could be identified and separated from our analysisof overall tissue deformation, since these vessels would exhibitgreater displacement than the surrounding, unaffected bloodvessels.

In order to test similarity measures (matching algorithms),we used the gray-level information around each interest pointin the first image, and searched for the best matches. Thismethod was used for two reasons: (1) since gray-levelinformation around each interest point is assumed to containhigh information and be geometrically stable, it is easier tofind a match for such a window, and (2) our approach yielded

accurate results after using a Harris detector only once, forthe first frame, instead of using features matching techniquesthat require extracting interest points at each frame which iscomputationally expensive.

2.4.3. Template matching. Template matching techniquesare relatively routine. The first step was to take a windowaround an interest point in a given frame as a template(figure 2). The size of this template was not fixed and couldbe used as a tuning factor that depended on the animation.The second step was to search for this template in the nextframe using a template-matching algorithm. The search areawas centered on the interest point from the previous frame.The size of this search area depended on the device insertionspeed, and was always greater than the template area. Thetemplate window started from the upper left corner of thesearch window, and scanned pixel by pixel from left to rightand from top to bottom. At each pixel the normalized cross-correlation between the template window and its projection onthe search window was calculated. The point that providedthe best match was then established as the new location of theinterest point in the next frame.

2.4.4. Normalized cross-correlation (NCC). Given an imageI and a template window T , centered at the point (u, v) inthe search window, the normalized cross-correlation (NCC)between the template and its projection on the image (a windowcentered at (u, v) and with the same size as T) is given by

NCC I,T (u, v)

= "x,y[(I (x, y)' Iu,v)(T (x ' u, y ' v)' T )]{"x,y[I (x, y)' Iu,v]2"x,y[T (x ' u, y ' v)' T ]2}

.

This operation is performed over all the points inside thesearch window, and the maximum value of NCC is at thepoint that gives the best match. The normalized cross-correlation overcomes some disadvantages of the commoncross-correlation operator: (1) cross-correlation is invariant toa linear (affine) transformation of image intensities (a constantaddition and common scaling of intensity), and (2) the rangeof CC depends on the size of the template window. NCCproduces a peak with a value of 1.0 at the perfect matchbetween the template window and its corresponding windowin the search area. The center of this matching window will bethe new interest point location. NCC is relatively robust in thepresence of noise, changes in scale and gray level, and imagedeformation [31].

2.4.5. Rejecting outliers. As outlined above, ‘children’interest points were treated as independent of one other, andthen regrouped to reject the outliers by choosing a single‘parent’. The portion of the image containing the tissue wasinitially divided into 100 ‘parent’ windows (100 pixels high by110 pixels wide), with the goal of identifying one interest pointin each block that best represents that block’s deformation(figure 2(B)). In some cases interest points may not reflectoverall tissue deformation; for example, when the interestpoint and its vascular counterpart were dragged through the

199

C S Bjornsson et al

(A)

(D) (E)

(F)

(B) (C)

Figure 2. The concept of template matching. (A) A template (white box) centered on an interest point (IP). (B) The next image in the timeseries is shown with a large window representing the search area. The original template is shown by a dashed box. A box the size of theoriginal template slides within the search area in a rasterized pattern to find an area similar to the original. (C) The same frame as in (B)shown with the template that best matches the original template (white box). The center of this box is the new IP location. (D) A magnifiedview of the search area, showing the original (dashed box) and new (solid box) templates. ((E) and (F)) Magnified views of the original (E)and new (F) templates.

tissue by the device during insertion. The ‘children’ strategyreduced the risk of tracking these outliers. Each parentwindow was divided into four sub-windows (50 pixels high by55 pixels wide), and a child interest point with the maximumHarris response was chosen for each sub-window, to yield400 children interest points. The parent interest point isthe child interest point with the maximum Harris response.The deformation occurring on the parent interest point wascomputed after using the following two-step method forfiltering the results:

1. Direction constraint. Deformation in the directionof device insertion should be greater or equal to thedeformation perpendicular to this axis. Any point thatdoes not pass this test was rejected. Rejection wasaccomplished by making the deformation on this pointequal to the deformation on the nearest-neighbor pointthat passed this test.

2. Local median filtering. Median filtering of every fourchild interest points was performed by taking the medianof their x and y deformations, and considering the resultantvalues as the deformation of their parent interest point.

In summary, the above image analysis steps provide anobjective and fully automated approach to measuring tissuedeformation at 100 discrete points in the tissue slice across theentire field of view.

2.5. Biomechanical strain analysis

Our aim is to first obtain a continuous representation of thedeformation field from the discrete displacement data providedby the point traces. Such a representation allows predictionof displacements at any point within the image domain. Wedenote the displacement at any point (x, y) of the image byits x- and y-components ux and uy, respectively. To get ridof outliers in the image data, the moving least squares (MLS)technique [32] is used to first rectify the data in a least-squaressense. For example, the u-displacement is rectified as

urectified =N(

I=1

hIuI

where N in the number of interest points tracked in the imageand hI is a function defined in [32] which is nonzero only ona finite disk around each interest point I. In order to generatea smooth and continuous deformation field we project therectified displacements onto a regular 21 ( 21 grid coveringthe original 1000 pixel by 1100 pixel area used to generateinterest points.

Once the displacement field is generated, the strain fieldsmay be computed in a straightforward manner using thefollowing expressions [33],

#xx = !ux

!x, #yy = !uy

!y, and $xy = !ux

!y+

!uy

!x

200

Effects of insertion conditions on tissue strain and vascular damage

Figure 3. Selected frames from two slow (125 µm s'1) insertionsdemonstrating minimal (left column) and severe (right column)tissue deformation. Interest points are marked in red or yellow, andtheir motion is tracked by green lines in subsequent images.

M A movie of this figure is available. See movie 2.

where #xx and #yy are the normal strains and $xy is the shearstrain. Instead of looking at three different strains, whichdepend on the coordinate system in which they are measured,our goal was to compute the maximum strains (shear andnormal) in the tissue, which are significantly better indicatorsof damage. Specifically, we have calculated (1) principalstrains, (2) maximum shear strain and (3) effective strain fromthese nominal and shear strain data.

In a body under plane strain, the strain state can becompletely described by two mutually orthogonal strains anda direction of one of the strains. In this coordinate system,known as the principal axes, all of the shear strains disappearand the remaining strains, known as the principal strains(denoted as #1 and #2), are purely tensile or compressive. Themathematical expressions for the principal strains are

#1,2 = #xx + #yy

2±

)*#xx ' #yy

2

+2

+,$xy

2

-2.

For any given strain situation, max(#1, #2) represents themaximum normal strain in the tissue.

The set of axes inclined at 45& to the principal axes of strainis the set for which shear strain component is a maximum,$xyMax

$xyMax =

)*#xx ' #yy

2

+2

+,#xy

2

-2.

The effective strain is computed from the principal strains #1

and #2 using the formula

# =.

23

$#2

1 + #22

%.

This measure is used extensively in the applied mechanicsliterature to assess damage to materials.

Thus, we have taken deformation information from 100interest points derived from our vascular markers locatedthroughout each tissue slice and transformed them into aregular grid of 21 ( 21 effective strain measurements acrossthe tissue area.

2.6. Statistical analysis

2.6.1. Criteria for data inclusion. Insertion data were firstscreened to meet several criteria. A few of the first insertionsexhibited considerable drift or vibration of the insertion arm,evidenced by visual inspection of the insertion movies aswell as significant global lateral deformations. Insertionsdemonstrating these properties were excluded. In some cases,the fibrin glue adhering tissue slices to the culture dish wasinsufficient, and the slice buckled. Each insertion moviewas followed by a movie scanning different brain regionsbeneath the insertion site, and buckled tissue slices were easilyidentified by out-of-focus vasculature around deeper corticallayers, usually the corpus callosum and/or hippocampus.Insertions resulting in tissue buckling were excluded fromsubsequent analysis. Slices that were incubated longer than6 h demonstrated less deformation and strain, suggesting initialdegradation of the pia. This effect was not observed priorto 6 h, nor was there a gradual decline in strain over timeduring this period. Finally, insertion attempts through pialblood vessels resulted in severe compression (see figure 5)regardless of condition and were consequently excluded fromaverage strain measurements.

201

C S Bjornsson et al

Figure 4. Several forms of vascular damage accompany neural prosthetic device insertion, including fluid displacement, vessel rupture, andsevering, tugging, and dragging of microvasculature. (A) This slow insertion of a sharp-tipped device results in fluid displacement withintranscerebral arteries approximately 100 µm away. (B) The artery to the left of the insertion track ruptures as fluid reaches downstreambranch points. Fluid first accumulates as a bolus within the tissue (B4, arrow) before it reaches the edge of the slice and diffuses intosolution (B5, arrow). (C) Along the insertion track, microvascular networks are tugged or dragged along by the device as it travels inward.In this case a smaller (precortical) pial artery is also carried into the brain along with the device (D, arrow).

M A movie of this figure is available. See movie 3.

2.6.2. Statistical analysis. For each insertion, meaneffective strain values were calculated for each of nine regionscomprising a 3 ( 3 grid of strain measurements (figure 6);49 individual strain measurements from the original 21 ( 21grid (section 2.5) were averaged in each sector. Regionalaverages for each insertion condition were compared using aone-way ANOVA followed by Hsu’s Multiple Comparison tothe Best (MCB) post hoc analysis using JMP software (SASInstitute, Cary, NC). Differences were considered significant atp ! 0.05.

To compare the slope and intersects of regression lines fordata correlating pial deformation versus mean effective strain,we used the linear regression function of GraphPad Prism(GraphPad Software, Inc., San Diego, CA). Differences wereconsidered significant at p ! 0.05.

3. Results

In order to directly visualize brain tissue deformation andvascular damage during device insertion, the vasculatureof intact animals was labeled with fluorescent dextranand microbeads prior to tissue slice preparation. The

fluorescent dextran filled the blood vessels, providing cleardescriptions of vascular size, distortion, compression andrupture. The microbeads provided discrete sites that improvedour automated tracking. Real-time video imaging of thesepreparations provided a clear description of specific vasculardamage and tissue deformation during device insertion(see movie 1). These real-time images were analyzed usingcustom software to track several hundred interest points ineach image as they moved in response to insertion (figure 3;movie 2).

Two types of information derived from the real-timeimages will be discussed: qualitative descriptions of damageto the neurovasculature and quantitative analyses of tissuedeformation provided by automated tracking.

3.1. Qualitative analysis of vascular damage

Four general types of vascular damage were observed: (1) fluiddisplacement, (2) vessel rupture, (3) vessel severing and (4)dragging of blood vessels by the device. Fluid displacementwas frequently observed within the transcerebral arteries thattraverse the cortex perpendicular to the surface (figure 4(A)),often as far as 300 µm away from the insertion site. Fluid

202

Effects of insertion conditions on tissue strain and vascular damage

Figure 5. Frequency of specific vascular damage relative to insertion conditions. Fluid displacement along transcerebral arteries, rupture atthe downstream ends of these vessels, severing and dragging of microvasculature were recorded. Overall, faster insertions resulted in lessdamage to the neurovasculature; tip geometry did not contribute to vascular damage in a predictable way.

contained within the vessels was usually pushed away fromthe pial surface. In many instances, the downstream branchesof arteries where fluid was displaced were unable to withstandthe accumulating pressure, and consequently ruptured. Thesecases were clearly not due to fluid leaking from vessels at theedge of the slice, since the fluorescent material escaping fromruptured vessels first formed a compact, confined bolus thatexpanded beside the ruptured vessel before reaching the edgeof the tissue slice and diffusing into the saline (figure 4(B);movie 3).

Severing, tugging, and dragging of the microvasculatureoccurred along the insertion track (figure 4(C); movie 3). Inmany cases a substantial network of microvasculature wasinvolved, presumably resulting in considerable damage tothe surrounding neural parenchyma. In addition, smallerprecortical and cortical pial arteries were sometimes carriedinto the brain along with the device (figure 4(D); movie 3).Surprisingly, these events did not correlate with tip geometry(figure 5). Generally speaking, all of these events occurredmore frequently at slower (500 and 125 µm s'1) insertionspeeds (figure 5). Fast insertion (2000 µm s'1) of sharpdevices was the only condition where no specific vasculardamage was observed.

3.2. Importance of cortical surface features

Attempts to insert devices of any shape at any speed wereseverely impaired when devices were inserted through thickerregions of the pia, specifically where large distributing arteries

(!200 µm diameter) or similarly sized veins were presentalong the cortical surface. In the majority of insertionswhere severe compression was observed, examination of thedevice retraction videos revealed that a pial vessel interferedwith the insertion process and was adhered to the device(figure 6). Insertions through smaller precortical and corticalarteries (!30–60 µm diameter) often did not result in severecompression, although these vessels were sometimes carriedinto the neural parenchyma by the device (movie 3). Devicesinserted through larger surface vessels occasionally managedto partially penetrate the slice; this may reflect differencesin vessel wall thickness between arteries and veins. Underno circumstances did devices interacting with surface vesselsinsert cleanly.

3.3. Automated image analysis and manual validation

In order to make quantitative measurements of insertion-generated tissue deformation, an automated method fortracking discrete interest points throughout each insertionwas developed (see section 2). Briefly, this procedure firstidentified 400 candidate interest points (each a box of 10 (10 pixels with a unique arrangement of intensities) in the firstimage frame, then tracked the 100 points with the highestHarris scores through subsequent image frames.

The dextran- and microbead-labeled vasculature providedideal markers for tracking tissue deformation during deviceinsertion. The method for tracking interest points betweensuccessive frames (figure 2) worked equally well for insertions

203

C S Bjornsson et al

(A) (B)

(C) (D)

(E)

Figure 6. Severe compression results when a blunt device isinserted at slow speed through a large pial blood vessel. In this case,the tissue slice was perfused with Hoechst 33342 to label cell nucleiin addition to the vascular labels. (A) A large blood vessel(!100 µm) can be seen at the cortical surface prior to insertion(arrow). (B) In this case, the device completely failed to penetratethe tissue slice. (C) As the device was retracted it carried the bloodvessel with it, clearly demonstrating the interference afforded bycortical surface features. (D) Devices inserted through smallervessels sometimes penetrated the pia after causing severecompression. Here, a large accumulation of Hoechst 33342-labeledcells can be seen around the device tip. (E) Upon retraction, it isevident that these cells are associated with a pial blood vessel. Thissuggests that devices can carry leptomeningeal cells with them intothe brain. These cells may participate in sheath formation and mayestablish a glia limitans that could repel growing neurites, furtherlimiting device performance.

M A� movie� of� this� figure� is� available.� See� movie� 3.

at all speeds. Although some of the vascular distortion resultedin a change in position relative to the surrounding tissue(e.g. displacement, rupture, dragging, severing) the methodof excluding interest points that moved differently than theirneighbors ensured that interest points reflected overall tissuedeformation (see section 2).

In order to ensure that automated tracking accuratelydescribed insertions, 3000 interest points were manuallytracked through the insertion phase of all fast and somemedium and slow insertions. For insertions at all speeds,interest point tracking was exceedingly accurate over thecentral portion of the image. Once interest points reachedthe edges of the image, because a significant fraction of

pixels from the previous template window were no longerpresent, the tracking algorithm either stalled at the previouslocation or adopted a new interest point for tracking. Incases where severe compression was observed, interest pointsderived from the upper cortical layers in the direct path ofthe device sometimes became difficult to track because largetemplate changes occurred from one frame to the next as thevasculature underwent rapid changes during deformation. Inthese cases, the algorithm typically adopted a new interestpoint that underestimated the true deformation.

Because automated tracking of interest points was mostaccurate over the central portion of the image, the strain datawere grouped into nine regions comprising a 3 ( 3 grid acrossthe tissue (figure 7) to ensure that erroneous tracking near theedges of the tissue did not detract from measurements madealong the insertion track. This allowed us to use the automatedtracking data without resorting to manual editing of individualdata sets.

3.4. Quantitative analysis of the effects of tip geometry andinsertion speed

For superficial and middle cortical layers along the insertiontrack, lower mean effective strains were observed for sharptips with faster insertions (figure 7). Using Hsu’s MultipleComparisons to the Best (MCB) post hoc test to assess thesignificance of differences in mean effective strains in eachof the nine regions, we found that fast insertion of sharp tipsresulted in significantly lower strains than (1) slow insertionsof medium-shaped tip geometries ( p = 0.033) and medium-speed insertions of medium-shaped tips ( p = 0.021) in thesuperficial region of cortex along the insertion track near thesurface (figure 7(A)), and (2) slow insertions of sharp ( p =0.035) and medium-shaped ( p = 0.046) tip geometries in themid-region of cortex along the insertion track (figure 7(B)).

Our quantitative analysis indicated that additionalbiological variables might contribute to the variation in theobserved mean effective strains for each insertion. To addressthis possibility, we chose to consider one such variable—the amount of surface deformation along the insertion track.We chose this variable because it reflects the ease withwhich devices penetrate the pial surface. We compared thecalculated mean effective strain with the maximum distancethe pia was displaced for each insertion (a measure of surfacecompression), grouped according to device tip geometry(figure 8(A)) or insertion speed (figure 8(B)). These twoparameters are strongly correlated ( p ! 0.001 for all cases),and this relationship was independent of tip geometry, butdependent on insertion speed, with insertion at medium speed(500 µm s'1) causing significantly less mean effective strainfor a given amount of pial compression ( p ! 0.01).

4. Discussion

Labeled brain slices provide a powerful approach to assessthe consequences of different insertion parameters in terms ofvascular and cellular damage and overall tissue deformationand strain. This study clearly demonstrates that there is much

204

Effects of insertion conditions on tissue strain and vascular damage

(A) (B)

Figure 7. Optimal insertion conditions as indicated by comparing the average maximum effective strain under different insertion conditions.Data are presented for the superficial and middle cortical layers along the insertion track, as shown in the diagram (left). Fast insertion ofsharp devices resulted in lower tissue strain overall. Note that the scale bars are different for each graph. Mean effective strain for insertionconditions marked with an asterisk were significantly lower than those marked with a dagger ( p ! 0.05). The tables below each graphprovide mean ± SEM for each condition.

(A) (B)

Figure 8. Mean effective strain is dependent on pial deformation. The same data are presented in (A) and (B), categorized according to tipshape (A) or insertion speed (B). The trend line for the entire data set is defined by the equation y = 0.0539x – 0.0204 (R2 = 0.8264).Regression lines and their corresponding equations accompany each graph. (A) Different tip shapes exhibit similar trends to the overalltrend, suggesting they do not affect this relationship. (B) Insertion at medium speed (500 µm s'1) resulted in a significantly differentrelationship compared to the overall trend ( p < 0.001), demonstrating that insertion speed affects this relationship.

more happening than simple deformation and compressionof a homogenous tissue, rather the brain is a complexheterogeneous material that requires greater understanding andcharacterization before accurate simulations can be developed.The vascular damage described here could not have beenobserved using computer modeling or imaging the corticalsurface during insertions into whole brains or agar tissuephantoms. This technique will serve as the cornerstone offuture studies addressing the importance of additional insertionparameters for optimal device insertion.

We anticipated that sharp tips might tend to severmicrovessels whereas blunt tips might drag microvascularnetworks. Surprisingly, tip geometry had little effect on thefrequency or type of vascular damage (figure 5). Consideringthe relatively small cross-sectional area of our devices(60 ( 100 µm), it is possible that tip geometry is notas critical as it may be for larger devices (e.g. deep-brainstimulating electrodes, !1500 µm in diameter). However,insertion speed played an important role, with faster insertionsgenerally resulting in fewer cases of observed vascular

damage. Regarding fluid displacement and vessel rupture,faster insertions (2000 µm s'1) appeared to cause the leastamount of fluid displacement along transcerebral arteries,whereas slower insertions (500 and 125 µm s'1) often resultedin displacement and downstream rupture of these vessels.Tissue distortion during slower insertions may have displacedfluid along transcerebral arteries without permitting time torecover, resulting in a greater tendency to rupture. If so, usinginsertion speeds slower than 125 µm s'1 may provide thesearteries time to re-equilibrate and recover without subsequentrupture. Future studies will address this possibility.

Review of the videos demonstrated that vascular damagecan occur a considerable distance away from the insertionsite. Compression and rupture of transcerebral arteriesoccurred as far as 300 µm away, and dragging of smallervessels could distort extensive networks of neurovasculature,presumably disrupting the surrounding cells. This raisesthe possibility that the immediate damage triggering reactiveresponses may extend further from the insertion site thanpreviously thought. We have hypothesized that cell-to-cell

205

C S Bjornsson et al

signaling resulting from release of pro-inflammatory factorswas responsible for the extent of the observed reactiveresponses [15]; however, these observations suggest that directcell damage due to tissue distortion and vascular damagemay also contribute. Minimizing the initial insertion damagemay therefore help reduce subsequent reactive responses.Combining conditions that produce minimally damaginginsertions with pharmacological intervention strategies tocontrol reactive responses [34] may provide adequate controlof the reactive responses and device encapsulation to ensurelong-term, consistent device performance, enabling stablesignal recording.

One of our primary goals was to determine whether thedesign and insertion parameters we selected would reveal aclear set of optimal insertion conditions. Quantitative analysisof fast (2000 µm s'1), medium (500 µm s'1) and slow(125 µm s'1) insertions indicated that fast insertions ofsharp tips resulted in the lowest mean effective tissue strains(figure 7). However, the differences between insertionconditions were only significant between the best (fastinsertions of sharp-tipped devices) and worst conditions (slowinsertions of medium- and sharp-tipped devices), and onlyalong the insertion track. Our interpretation of this resultis two-fold. First, while our range of insertion speedsspanned the full range available to our insertion tool, theyare not representative of the full range of speeds used inresearch using neuroprosthetic devices, including both slower(!100 µm min'1) [35, 36] and faster (>8.3 m s'1) [37]insertions. Similarly, our device geometries were of consistentthickness and were not as thin as Michigan electrodes(16 µm thick) or tapered three-dimensionally like electrodesin the Utah array. Future improvements in our experimentalapparatus will permit exploration of a wider range ofinsertion speeds and different device geometries; differencesbetween these more extreme insertion conditions may be moresignificant. Second, there are likely a number of as yetunrecognized biological factors affecting these measurements,which may contribute to large variances that obscure thedifferences between our experimental conditions. Clearly wemust understand and address sources of biological variability,since these may be of key importance to determiningconditions that will reliably produce minimally damaginginsertions.

One such variable appears to be the location ofthe insertion site relative to pial structures. Forinstance, approximately 20% of insertions resulted in severecompression (e.g. figure 6). Movies recording the withdrawalof these devices demonstrated that most of the devices hadblood vessels adhering to the tip, suggesting these insertionswere attempted through pial blood vessels. It is clearthat pial structure varies across the surface of the brain.Correlating maximum pia deformation and mean effectivestrain provided data consistent with this hypothesis. Thesecorrelations demonstrate no significant effect of tip geometryon insertion; however, insertion speed did have a significanteffect. Specifically, the medium insertion speed (500 µm s'1)resulted in lower mean effective strain for a given amountof pia deformation. However, the effect of tip geometry

requires additional investigation, since examination of thesedata reveals that the sharper tips did not produce the largerstrains associated with blunt and medium tip geometries(figure 8(A)). These observations indicate that effectivelypenetrating the pia is a critical step in reducing tissuedeformation and damage, and that optimal insertion conditionsfor penetration of the pia and travel through the brainparenchyma may be different. We hypothesize thatslower insertions (125 µm s'1) failed to penetrate the piathus generating more tissue strain, while faster insertions(2000 µm s'1) were able to penetrate the pia but did notallow as much tissue relaxation as the medium insertion speed(500 µm s'1). Future studies including measurement ofinsertion forces will test these hypotheses.

The contribution of cortical surface features cannotbe overemphasized. Devices successfully inserted throughsmaller pial vessels can carry leptomeningeal cells into thebrain (figures 6 ((D) and (E)). These cells may then interactwith local astrocytes to form a thick sheath similar to the pialglia limitans, which could suppress neurite outgrowth muchlike the pia [38, 39]. Furthermore, inserting devices throughpial vessels likely impairs circulation over an extensive tissuevolume. The smallest ‘precortical’ arteries deliver blood toa column of cortical tissue beneath an area of the surfaceapproximately 1 mm2, while upstream ‘distributing’ arteriessupply a territory of approximately 7.5 mm2 in humans [25].Both classes of arteries were disrupted in our experiments,with severe consequences at least in terms of insertion success.From a functional standpoint, very few (sometimes single)precortical arteries often supply an entire functional unit inbarrel cortex [40, 41]; the distributing area of these arteries alsocorresponds to the width of human ocular dominant columns ofthe visual cortex [25]. Clearly, disruption of even the smallestsurface vessels can have profound neurological consequences.Increased damage caused by insertion through cortical surfacevessels may help explain why the electrodes on some shanksbut not others within multishank arrays simply fail to work, andwhy amplified reactive responses are often observed aroundsome shanks but not others [42]. While it is possible toorient single shank devices so they avoid pial vessels, the rigidconformation of multishank arrays precludes this approach,especially when facing the density and inter-animal variation inthe distribution of surface vessels. Thus, minimally damaginginsertions will require the use of carefully regulated insertionspeeds, devices with a relatively small number of shanks, andcareful positioning of these devices at the time of insertion.

Acknowledgments

This work was supported in part by grants from the NIH,NIBIB, EB-000359 and NINDS, NS-044287, NSF, EEC-9986821, and the International Collaboration Program ofNBS-ERC/KOSEF.

References

[1] Csicsvari J et al 2003 Massively parallel recording of unit andlocal field potentials with silicon-based electrodesJ. Neurophysiol. 90 1314–23

206

Effects of insertion conditions on tissue strain and vascular damage

[2] Kipke D R et al 2003 Silicon-substrate intracorticalmicroelectrode arrays for long-term recording of neuronalspike activity in cerebral cortex IEEE Trans. Neural Syst.Rehabil. Eng. 11 151–5

[3] Moran D W and Schwartz A B 1999 Motor corticalrepresentation of speed and direction during reachingJ. Neurophysiol. 82 2676–92

[4] Moxon K A et al 2004 Ceramic-based multisite electrodearrays for chronic single-neuron recording IEEE Trans.Biomed. Eng. 51 647–56

[5] Serruya M D et al 2002 Brain–machine interface: instantneural control of a movement signal Nature 416 141–2

[6] Vetter R J et al 2004 Chronic neural recording usingsilicon-substrate microelectrode arrays implanted incerebral cortex IEEE Trans. Biomed. Eng. 51 896–904

[7] Breit S, Schulz J B and Benabid A L 2004 Deep brainstimulation Cell Tissue Res. 318 275–88

[8] Goodman J H 2004 Brain stimulation as a therapy for epilepsyAdv. Exp. Med. Biol. 548 239–47

[9] Wallace B A, Ashkan K and Benabid A L 2004 Deep brainstimulation for the treatment of chronic, intractable painNeurosurg. Clin. N. Am. 15 343–57

[10] Hetling J R and Baig-Silva M S 2004 Neural prostheses forvision: designing a functional interface with retinal neuronsNeurol. Res. 26 21–34

[11] Loeb G E 1990 Cochlear prosthetics Annu. Rev. Neurosci.13 357–71

[12] Edell D J et al 1992 Factors influencing the biocompatibilityof insertable silicon microshafts in cerebral cortex IEEETrans. Biomed. Eng. 39 635–43

[13] Polikov V S, Tresco P A and Reichert W M 2005 Response ofbrain tissue to chronically implanted neural electrodesJ. Neurosci. Methods 148 1–18

[14] Turner J N et al 1999 Cerebral astrocyte response tomicromachined silicon implants Exp. Neurol. 156 33–49

[15] Szarowski D H et al 2003 Brain responses to micro-machinedsilicon devices Brain Res. 983 23–35

[16] Sipple B A, Spence A J W, Retterer S, Craighead H G,Isaacson M, Turner J N and Shain W 2006 Vascular damageand repair following insertion of model neural prostheticdevices Brain Res. submitted

[17] Welsh J P and Schwarz C 1999 Multielectrode recording fromthe cerebellum Methods for Neural Ensemble Recordings(CRC Methods in Neuroscience Series) ed M A L Nicolelis(Boca Raton, FL: CRC) pp 79–100

[18] Lee H et al 2005 Biomechanical analysis of siliconmicroelectrode-induced strain in the brain J. NeuralEng. 2 81–89

[19] Subbaroyan J, Martin D C and Kipke D R 2005 Afinite-element model of the mechanical effects ofimplantable microelectrodes in the cerebral cortex J. NeuralEng. 2 103–113

[20] Gefen A et al 2003 Age-dependent changes in materialproperties of the brain and braincase of the ratJ. Neurotrauma 20 1163–77

[21] DiMaio S P 2003 Needle insertion modeling and simulationIEEE Trans. Robot. Autom. 19 864–75

[22] Pease D C and Schultz R L 1958 Electron microscopy of ratcranial meninges Am. J. Anat. 102 301–21

[23] Rhodin J A G 1974 Histology: A Text and Atlas (New York:Oxford University Press) pp 322–8

[24] Morse D E and Low F N 1972 The fine structure of the piamater of the rat Am. J. Anat. 133 349–68

[25] Nonaka H et al 2003 Microvasculature of the human cerebralmeninges Neuropathology 23 129–35

[26] Bar T 1980 The vascular system of the cerebral cortexAdvances in Anatomy, Embryology and Cell Biology vol 59,ed A Brodal (New York: Springer)

[27] Rasband W S et al 1997–2005 ImageJ (US National Institutesof Health, Bethesda, MD) http://rsb.info.nih.gov/ij/

[28] Deriche R and Giraudon G 1993 A computational approachfor corner and vertex detection Int. J. Comput. Vis.10 101–24

[29] Schmid C, Mohr R and Bauckhage C 1998 Comparing andevaluating interest points Proc. 6th Int. Conf. on ComputerVision (Bombay, India)

[30] Harris C and Stephens M 1988 A combined corner and edgedetector Proc. 4th Alvey Vision Conf. pp 147–51

[31] Martin J and Crowley L 1995 Experimental comparison ofcorrelation techniques IAS-4 Int. Conf. on IntelligentAutonomous Systems (Karlsruhe)

[32] De S and Bathe K J 2001 Towards an efficient meshlesscomputational technique: the method of finite spheres Eng.Comput. 18 170–192

[33] Timoshenko S P and Goodyear J N 1970 Theory of Elasticity(New York: McGraw-Hill)

[34] Retterer S et al 2004 Model neural prostheses withintegrated microchannels: treating the reactive responsesto implantable devices IEEE Trans. Biomed. Eng.51 2063–73

[35] Nicolelis M A et al 2003 Chronic, multisite, multielectroderecordings in macaque monkeys Proc. Natl Acad. Sci.USA 100 11041–6

[36] Nicolelis M A L et al 1999 Methods for simultaneous multisiteneural ensemble recordings in behaving primates Methodsfor Neural Ensemble Recordings (CRC Methods inNeuroscience Series) ed M A L Nicolelis (Boca Raton, FL:CRC) pp 121–56

[37] Rousche P J and Normann R A 1992 A method forpneumatically inserting an array of penetrating electrodesinto cortical tissue Ann. Biomed. Eng. 20 413–22

[38] Ness R and David S 1997 Leptomeningeal cells modulate theneurite growth promoting properties of astrocytes in vitroGlia 19 47–57

[39] Shearer M C and Fawcett J W 2001 The astrocyte/meningealinterface—a barrier to successful nerve regeneration? CellTissue Res. 305 267–73

[40] Cox S B, Woolsey T A and Rovainen C M 1993 Localizeddynamic changes in cortical blood flow with whiskerstimulation corresponds to matched vascular and neuronalarchitecture of rat barrels J. Cereb. Blood Flow Metab. 13899–913

[41] Woolsey T A et al 1996 Neuronal units linked tomicrovascular modules in cerebral cortex: responseelements for imaging the brain Cereb. Cortex 6 647–60

[42] Rousche P J and Normann R A 1998 Chronic recordingcapability of the Utah Intracortical Electrode Array in catsensory cortex J. Neurosci. Methods 82 1–15

207