Figure 1.
Using the color deconvolution plugin, an image of a tissue cores double stained with a counter staining dye (e.g. hematoxylin) and a second dye binding to cellular feature of interest (e.g. eosin or dye couple using IHC to a protein of interest) can be separated into two images highlighting amount and patterns of the feature of interest. The post deconvolution images are false-color coded in such a way that the RGB values of each pixels are associated to each dye absorption and according to the color properties of each original dye (i.e. pink for eosin, brown for DAB and blue for hematoxylin).
Figure 2.
Melan-A and nuclei segmentation dependence on the staining variability in the three patient cohorts available.
Using CellProfiler a Melan-A staining IHC positive mask is generated for TMAs sections from the discovery (A) and the validation (B) cohorts. In addition a second TISSUE mask was generated from the hematoxylin deconvolved channel (i.e. H-CHANNEL) highlighting all areas of the tissue were there was tissue present (C–D). Box and whiskers plot of average intensity of the DAB-CHANNEL and the H-CHANNEL in the discovery and the validation cohort shows similar values and no significant differences (E). On the segmented regions of interest (i.e. TISSUE and Melan-A) generated from the CellProfiler ruleset intensity of H-CHANNEL and DAB-CHANNEL was measured and distribution across the cohorts shows similar levels of total hematoxylin intensity (in the log10 scale), measured as sum of all intensity values within the segmented ROIs, but a significantly decrease of total intensity values of DAB in the Melan-A ROIs between the discovery and validation cohort with the hematoxylin counterstained cohort (F). Scatter plotting of total intensity values of the deconvolved H-CHANNEL for the same tissue samples images from the discovery cohort were same samples were once stained with counterstained with hematoxylin and then independently stained with Melan-A (G).
Figure 3.
Cell nuclei segmentation of Melan-A stained TMAs tissue spot images.
The CellProfiler ruleset take as input the original IHC stained tissue core image as well as the Melan-A binary mask (A). (B) In the next step the corresponding H and DAB channel images are deconvolved and only the H-CHANNEL image is further processed for cell nucleus segmentation. Initially H-CHANNEL image is pre-processed to sharpen the image and highlight object of interest (i.e. cell nucleus). Using the binary Melan-A mask in A, non-melanoma (C) and other types of cells (D) nucleus are segmented out (highlighted in red in panel C and D) using the robust background segmentation method in order to extract morphological and texture features.
Figure 4.
Cell nuclei segmentation on the discovery and the validation cohort.
Using the Melan-A binary masks, melanoma and the remaining cell nucleus are selectively segmented from all Melan-A stained tissue spot images in the discovery and the validation cohorts. (A) Linear regression analysis of average nuclear size, total number of nuclei and average granularity of size 1 features on the same tissue samples on the Melan-A_DISCOVERY and H_DISCOVERY cohorts after standartisation (divided by the maximum value and scaled at minimum value of 0 and maximum value of 1). (B) Average number of segmented melanoma (green bars) and other type of cell nuclei (blue bars) per tissue spot image in the MELAN-A_DISCOVERY and Melan-A_VALIDATION cohort. (C) Comparing the ratio of Melan-A positive tissue area ratio to the ratio of Melan-A positive cell nuclei ratio detected by the approach, the scatter plot shows how correlated the ratios are in both cohorts. (D) Ratio of Melan-A stained tumor area towards total amount of present tissue (green bars) and the ratio of melanoma cell nuclei (as defined by the Melan-A mask) to total number of detected nuclei (blue bars) in both cohorts. (E) Average intensity levels after stain deconvolution of IHC Melan-A intensity in Melan-A mask highlighted areas (blue bars) compared to the average hematoxylin intensity (green bars) in both cohorts.
Table 1.
Distribution of number images and identified cell nuclei in the training and validation data sets.
Figure 5.
Cell nuclei feature discrimination.
From the balanced pooled feature set from the Melan-A_DISCOVERY and Melan-A_VALIDATION dataset, selected features from the univariate analysis were tested using Receiver-Operator-Curve (ROC) analysis to show the performance of a binary classifier (i.e. melanoma or non-melanoma) system as its discrimination threshold is varied for each of the selected features (A). To illustrate the differences in granularity (top discriminative feature from the ROC analysis), a representative set of melanoma and non-melanoma nuclei were selected from the discovery TMA and the histogram of the granularity feature (i.e. six pixel radius) values from each set is shown in B.C–D) Top texture feature (i.e. granularity feature with structure element of six pixel radius) in melanoma cells (C) and non-melanoma cells (D) with cell nucleus masks shown using different colored segmentation labels to differentiate nucleus of touching cells. E) Box-plot showing distribution of top granularity features with variable structure element size (i.e. six and two pixel radius) that show significant difference (p-value <0.001) in expression values between melanoma and non-melanoma cells. F) Comparison of major and minor axis of bounding ellipse around the nuclei of melanoma and non-melanoma cells shows no significant difference, showing no correlation of cell nucleus segmentation size and granularity feature as well as no cell selection bias for the evaluation of the granularity feature.
Figure 6.
Classifier training and validation.
An overview of the cell nuclei in the balanced Melan-A_DISCOVERY dataset following the optimal feature set (i.e. 34 features selected in the previous step) (A). A SVM classifier is build using data from the training set (66.5%) and tested on the remaining cases (33.5%). To test the robustness of the classifier, during testing Weka varies the threshold (by default set to 0.5 and the label for the nuclei would be the class with probability higher then 50%) on the class probability estimates from LibSVM for melanoma and non-melanoma resulting in different classification labels for the cell nuclei in the test set (B). The SVM classifier derived from the balanced Melan-A_DISCOVERY is then validated in 270 randomly selected tissue samples from the validation cohort stained with Ki67 antibody (C). Linear regression analysis is used on the automated counting of ratio of melanoma cells expressing Ki67 by the SVM-based model (blue dotted line) and ImmuRatio (red dotted line) and the same ratio exactly evaluated by an experienced human annotator (D).
Table 2.
Performance evaluation on divers supervised classifiers upon the balanced Melan-A_DISCOVERY and Melan-A_VALIDATION cohorts.