Nuquantus: Machine learning software for the characterization and quantification of cell nuclei in complex immunofluorescent tissue images

Determination of fundamental mechanisms of disease often hinges on histopathology visualization and quantitative image analysis. Currently, the analysis of multi-channel fluorescence tissue images is primarily achieved by manual measurements of tissue cellular content and sub-cellular compartments. Since the current manual methodology for image analysis is a tedious and subjective approach, there is clearly a need for an automated analytical technique to process large-scale image datasets. Here, we introduce Nuquantus (Nuclei quantification utility software) - a novel machine learning-based analytical method, which identifies, quantifies and classifies nuclei based on cells of interest in composite fluorescent tissue images, in which cell borders are not visible. Nuquantus is an adaptive framework that learns the morphological attributes of intact tissue in the presence of anatomical variability and pathological processes. Nuquantus allowed us to robustly perform quantitative image analysis on remodeling cardiac tissue after myocardial infarction. Nuquantus reliably classifies cardiomyocyte versus non-cardiomyocyte nuclei and detects cell proliferation, as well as cell death in different cell classes. Broadly, Nuquantus provides innovative computerized methodology to analyze complex tissue images that significantly facilitates image analysis and minimizes human bias.

4. The cell membranes were extracted from the red channel by applying the steerable filters 1 at scales [1.0,1.25,1.5,1.75,2.0,5.0,7.5,10.0,15.0]. The output of these filters were combined into a unique membrane map following 2 . Tensor Voting method 3 for 10, 8, and 6 scales was finally used for improving the output of the detector. 5. Extracellular pixels were defined as those with intensity less than 40. The binary map of extracellular pixels was dilated by 1, 2 and 3 pixels. Connected components were extracted and their size was reported into each pixel part of the component. Measuring these sizes allows discriminating the interstitial tissue from the large extracellular compartments, which greatly helps discarding extra tissue nuclei. Example: Extracellular component size (ECS) is equal to 0 when the pixel is in cardiac tissue. Otherwise, ECS is greater than 0.
Other state-of-the art features were tested, such as Gabor filters, the size of the nuclei, different color representations and different image scales. They appeared to be redundant with our current set of features and were removed for the sake of simplicity.

FIGURE LEGENDS
Supplemental Figure S1. Fluorescence Image preprocessing for intensity correction. a. Original confocal microscopy image of cardiac tissue that was fluorescently stained for α-sarcomeric actin and DAPI. b. The original image is processed using Gaussian blur filter (scale of 10 pixels to the red channel representing αactin/tropomyosin staining) to compute the average brightness around each pixel. This computation illustrates heterogeneous color and illumination artifacts that may occur due to variation in staining and confocal laser scanning. c. Illumination artifacts are corrected by normalizing the average brightness such that the maximum across the image is 1. Each pixel intensity value was divided by the normalized average intensity of its neighborhood. d. The corrected image is visualized via Gaussian blur filter to visualize the pixels that were bright enough before the correction remain unchanged, while the dim pixels are enhanced. Scale 20 µm.
Supplemental Figure S2. Histogram of probability scores given to nuclei. Distribution of CM and non-CM nuclei based on probabilities scores for a. swine prediction model and. b. mouse predication model. Each nucleus was screened by the Nuquantus prediction model and assign with a probability score. Higher probability score (>0.2) provides prediction of prospective non-CM nucleus. Lower probability score (<0.2) classifies the nucleus as a CM nucleus.

Supplemental Figure S3. Nuquantus validation for EdU and DAPI positive nuclei count.
The nuclei in 48 images obtained from post MI mice (N=4) and control shams (N=4) were counted twice: manually and then with Nuquantus software. The MI cardiac sections were analyzed for IA, BZ and VZ (3 images per sub-region per animal). Shams cardiac sections were analyzed using 3 images per animal similarly to VZ. The total nuclei count and CM nuclei count undergoing DNA synthesis (EdU + DAPI) that was measured by Nuquantus was compared to the manual counts. Trends showed matching counts without any statistical significant difference. (Non-parametric paired Wilcoxon test).
Supplemental Figure S4. Note that cases in which few segmented EdU labels co-localized with one nucleus identified by DAPI, were considered as a single EdU + DAPI nucleus. Scale 20µm.
Supplemental Figure S5. Nuquantus validation for TUNEL and DAPI positive nuclei count. 30 images obtained from N=4 healthy mice were analyzed for nuclei total count, total TUNEL positive nuclei count, CM nuclei count and TUNEL positive CM nuclei count. The counts were compared between negative and positive TUNEL controls using manual approach and Nuquantus software with applied user correction step. No statistical significant difference was detected by a non-parametric paired Wilcoxon test.

Original Image Image After Intensity Correction
Average Intensity Before Correction Average Intensity After Correction