Machine Learning-assisted Quantitative Mapping of Intracortical Axonal Plasticity Following a Focal Cortical Stroke in Rodents

Abstract

Stroke destroys neurons and their connections leading to focal neurological deficits. Although limited, many patients exhibit a certain degree of spontaneous functional recovery. Structural remodeling of the intracortical axonal connections is implicated in the reorganization of cortical motor representation maps, which is considered to be an underlying mechanism of the improvement in motor function. Therefore, an accurate assessment of intracortical axonal plasticity would be necessary to develop strategies to facilitate functional recovery following a stroke. The present study developed a machine learning-assisted image analysis tool based on multi-voxel pattern analysis in fMRI imaging. Intracortical axons originating from the rostral forelimb area (RFA) were anterogradely traced using biotinylated dextran amine (BDA) following a photothrombotic stroke in the mouse motor cortex. BDA-traced axons were visualized in tangentially sectioned cortical tissues, digitally marked, and converted to pixelated axon density maps. Application of the machine learning algorithm enabled sensitive comparison of the quantitative differences and the precise spatial mapping of the post-stroke axonal reorganization even in the regions with dense axonal projections. Using this method, we observed a substantial extent of the axonal sprouting from the RFA to the premotor cortex and the peri-infarct region caudal to the RFA. Therefore, the machine learningassisted quantitative axonal mapping developed in this study can be utilized to discover intracortical axonal plasticity that may mediate functional restoration following stroke.

Keywords: Ischemic stroke; Machine learning; Motor cortex; Neuroanatomical tract-tracing techniques; Neuronal plasticity; Support vector machine.

Fig. 1

A schematic workflow of the multi-pixel pattern analysis (MPPA) of the cortical axonal plasticity. (A) Tissue preparation process of BDA injected cortices flattening using customized plexiglass. (B, C) Fluorescence images were subjected to object detection and transposed into Cartesian coordinates. (D) Pixelated coordinates were extracted as a pattern data, (E) and followed machine learning classification to discriminate each pattern into sham or stroke groups. (F) Individual pixel classification accuracy, and accuracy-based statistical significance plotted on 3D brain images.

Fig. 2

Image acquisition of BDA-traced axons and extraction of the Cartesian coordinates from digitized axonal signals. (A) A representative picture of cortical biotinylated dextran amine (BDA) injection. Small burr hole (orange dotted line) was made on premotor cortex, and BDA was delivered through pulled-capillary glass. (B) Custom plexiglass setting for tissue flattening (left). Cortices were placed in between two slide glasses with 2 mm thick of plexiglass to ensure the flattened tissue thickness. Weight pressure was place on top of it to give constant pressure. Flattened cortex after flattening process (right). (C) Representative flattened tissue section image following labeling of vesicular-glutamate transporter 2 (vGLUT2) and BDA. BDA injection site was marked by asterisk. Whole tissue was outline by white dotted line. Scale bar=1000 μm. Axis of tissue section is indicated on bottom-right of image. M: medial, R: rostral, L : lateral, C: caudal. (D) Tissue alignment of vGLUT2 labeled barrel cortex. Manually contoured barrel cortex (red solid line), another tissue’s barrel cortex (orange solid line) was aligned to fit together. Scale bar=500 μm. (E) BDA positive signals (top) were marked with blue round shape markers (bottom) by mark object function embedded in StereoInvestigator software. Each marker location was transposed into Cartesian coordinates.

Fig. 3

Machine learning algorithm for analysis of axon density distribution pattern. (A) Schematic explanation of pattern extraction. From pixelated density map, 5 pixels were extracted as pattern data (red dotted line) from the randomly visited reference pixel (asterisk). (B) This process was repeated through whole pixelated density map to obtain all pattern data sets from each group. (C) Linear support vector machine (SVM) classifier was trained using extracted pattern data and classified test data based on hyperplane (solid line) set in training session with designated margin (dotted line), distinguishing each pattern into sham or stroke group. (D) Classifier accuracy map was plotted after calculation of individual pixel accuracy.

Fig. 4

Conversion of the histological BDA axon signals to the pixelated axon density map. (A) Biotinylated dextran amine (BDA) labeled tangential tissue images of individual subjects of sham groups. Scale bar=1500 μm. (B) Pixelated BDA density map of individual sham group subjects. (C) BDA labeled tangential tissue images of stroke groups. Scale bar=1500 μm. (D) Pixelated BDA density map of individual stroke subjects. Whole tissues (white dotted line) or infarction area (green dotted line) was outlined. Tissue axis is indicated at the left top corner (R: rostral, L: lateral).

Fig. 5

Comparison of BDA labeled axon sprouting using machine learning-based pattern analysis. (A, B) Mean axon density map of sham, and stroke group. Pattern data set extracted from mean axon density map were classified by support vector machine (SVM) classifier. (C) Calculated individual pixel accuracy was plotted. Color-coded accuracy map showed high accuracy area that has distinguishable pattern difference. (D) Biotinylated dextran amine (BDA) labeled axon images in high accuracy region of interest (ROI I) and low accuracy ROI II of sham group. Right panel is enlarged images of white dotted square. Scale bar=500 μm (left), 200 μm (right). (E) BDA labeled axon images in high accuracy (ROI I) and low accuracy (ROI II) in stroke group. Scale bar=500 μm (left), 200 μm (right). White dotted square was enlarged on the right. In high accuracy ROI I showed increased BDA labeled axons, while ROI II showed no difference in sham and stroke groups with relatively low density of BDA.

Fig. 6

Classifier accuracy-based axonal plasticity mapping and statistical analysis. (A, B) Accuracy map and p value map plotted on 3d brain images. Individual pixel’s p value was statistically analyzed by binomial test. p value map was plotted after Benjamini-Hochberg false discovery rate (FDR) correction. (C) BDA density of FDR<0.05 pixels following binomial tests, BDA density was extracted from FDR<0.05 pixels of each subject and analyzed by two-tailed unpaired t-tests. Stroke group showed significantly higher density of BDA compared to sham-operated group (p=0.0014, n=5). Data were plotted as mean±SEM. **p<0.01 for sham vs stroke group.