Machine learning based analysis of stroke lesions on mouse tissue sections

Abstract

An unbiased, automated and reliable method for analysis of brain lesions in tissue after ischemic stroke is missing. Manual infarct volumetry or by threshold-based semi-automated approaches is laborious, and biased to human error or biased by many false -positive and -negative data, respectively. Thereby, we developed a novel machine learning, atlas-based method for fully automated stroke analysis in mouse brain slices stained with 2% Triphenyltetrazolium-chloride (2% TTC), named "StrokeAnalyst", which runs on a user-friendly graphical interface. StrokeAnalyst registers subject images on a common spatial domain (a novel mouse TTC- brain atlas of 80 average mathematical images), calculates pixel-based, tissue-intensity statistics (z-scores), applies outlier-detection and machine learning (Random-Forest) models to increase accuracy of lesion detection, and produces volumetry data and detailed neuroanatomical information per lesion. We validated StrokeAnalyst in two separate experimental sets using the filament stroke model. StrokeAnalyst detects stroke lesions in a rater-independent and reproducible way, correctly detects hemispheric volumes even in presence of post-stroke edema and significantly minimizes false-positive errors compared to threshold-based approaches (false-positive rate 1.2-2.3%, p < 0.05). It can process scanner-acquired, and even smartphone-captured or pdf-retrieved images. Overall, StrokeAnalyst surpasses all previous TTC-volumetry approaches and increases quality, reproducibility and reliability of stroke detection in relevant preclinical models.

Keywords: Mouse stroke; TTC brain atlas; automated infarct volumetry; lesion analysis; machine learning; neuroanatomical mapping.

Figure 1.

Pipeline of software development. Left column (a) shows the steps for constructing the TTC-atlas. Slices from normal mice stained with TTC were scanned ( $I_a$ ), digitally pre-processed (for details see also Suppl. Figure 1) for production of segmented slices $I_s$ (i.e. extraction, rotation and background segmentation), indexed according to their anteroposterior (AP) distance from bregma (“indexing”), co-registered and summed to average (Aj) and mathematical standard deviation (SDj) images for each index. After software-assisted manual registration of the Allen anatomical masks to each corresponding Aj, the TTC-atlas with incorporated neuroanatomical data was created. Right column (b) shows the flowchart of lesion detection by StrokeAnalyst. The scanned images $I_a\mathrm{ }$ were preprocessed as those for the atlas (pre-processing) to segment single ones $I_s$ . Then, $I_s$ were indexed, registered and "normalized" to the corresponding Aj of the atlas using the Dramms software, which produced two transformation-maps (T1, T2). Utilization of SDj images, z-score masks, T1/T2 transformation maps, detection of each hemisphere surface and the middle-line of each TTC-slice within the machine learning algorithm returned the final mask of lesion L_F for volumetric analysis. Lesion masks provided also neuroanatomical information derived from the TTC-atlas.

Figure 2.

Construction of TTC-atlas. A corresponding plate of each coronal slice in TTC-atlas. An index is the anteroposterior distance of the slice from bregma, Aj is the average slice for each index and SDj the corresponding mathematical standard deviation image for Aj. The plate includes the corresponding Allen's anatomical mask and its overlay over the Aj. For each plate, the number of used single TTC slices is indicated (here, for index j = +0.14 mm we used n = 9 slices).

Figure 3.

Results of SA on the initial experimental set (Exp1) compared to manual volumetry. The infarct volumes measured by SA in mm³ (a) or as % percentage of left hemisphere (b) are similar to that of manual volumetry (h₁ and h₂), do not differ to CRS (as "ground truth") and are larger but not statistical different to those measured by InfarctSizer (IS). Bland-Altman plot for infarct volumes per single slices (c) shows an almost zero bias by SA compared to CRS. Infarct volumes (as % Vinf) correlate strongly to fESS values per animal (d), indicating that focal neurological deficits in fESS reflect the underlying infarct size. Hemispheric edemas as % percentage are also similar to human and CRS (e), despite that hemispheric volumes (right and left hemispheric volumes, RHV and LHV respectively, in mm³) are proportionally overestimated by SA compared to human or CRS measurements (f and g), as also shown in Bland-Altman plots (i) and (j). Image analyses of SA- and IS- versus CRS-masks through Dice-scores (k) and false positive rates (FPR, l) show that SA detects infarct areas better compared to IS; this applies in both cases where TTC-staining has a poor and good infarct-healthy differentiation. Most important, SA detects infarcts more reliably (specific), as indicated by a very low FPR rate in all cases (poor and good “differentiated” slices). A representative overlay projection (white) of SA (pink) and CRS (green) mask is shown in (h).

Figure 4.

SA can detect stroke lesions on images from variable sources and initial dpi. SA detects infarcts on scanned images with 600, 1200 and 2400 dpi (a to c) or even from the image captured with a common mobile phone at 72 dpi (d). Note that the white or black background does not affect the detection or segmentation process. Smartphone acquisition brings shadows (d, at white background) and its low resolution at 72 dpi brings relative inaccuracies. SA can surprisingly detect strokes even in images acquired online from previously published manuscripts^, (e and f); here resolution is again the only determining factor: the low resolution of 72 dpi in (e) provides moderate accuracy compared to the 300 dpi in (f).

Figure 5.

Important neuroanatomical mapping data derived from SA. Each blue bar shows (as % percentage) the corresponding prevalence of the damaged area in the cohort of Exp1 animals. Of note, agranular insular area, amygdalar nuclei, caudoputamen, globus pallidus, primary motor area, primary somatosensory area and hypothalamus were almost constant damaged by the fMCAo model in our cohort. Interestingly, our model affected also -among others- visual areas, visceral areas and auditory areas, with unknown significance and effects to preclinical stroke studies.