Figure 1.
A wide morphologic spectrum of nuclei in GBMs with variable combinations of oligodendroglioma and astrocytoma features.
A subset of GBMs, defined as grade IV astrocytic neoplasms, exhibits a variable degree of oligodendroglioma morphology. The shaded interval consists of a continuum of morphologies across the oligodendrolgioma to astrocytoma spectrum. Variable combinations of oligodendroglioma and astrocytoma cells, as well as morphologically ambiguous forms, make it challenging to reproducibly and accurately subclassify GBMs based on oligodendroglioma component.
Figure 2.
Overall schema of image analysis pipeline for GBM nuclei analysis.
(A) Parallel computation mechanism. Whole-slide pathology images for analysis were partitioned into smaller image tiles for parallel processing with a computing cluster infrastructure. (B) Schematic view of the system involving data collection, annotation, analysis and integration. All data elements, including microscopy imaging features, molecular data, clinical outcomes, and expert pathology review results are stored in a centralized Pathology Analytical Imaging Standards (PAIS) database, allowing researchers to query. The module for analysis and query of histologic features consists of image analysis, parallel high performance computation, analytical result and provenance data representation, creation of interfaces for human markup and annotation acquisition, data management, query support and data sharing. (C) Nuclei segmentation method. All nuclei were segmented by an efficient segmentation method where image morphological reconstruction and the watershed algorithm were used to normalize background and to segregate clumped nuclei, respectively. (D) Visualization of nuclei with distinct Nuclear Scores (NS). An NS was calculated with a set of most discriminating features derived from the associated segmented nucleus. We use color blue, green, and red to indicate nuclei with low, median, and high NS.
Figure 3.
Nuclear features and discriminating feature selection.
(A) Nuclear features can be divided into four categories: morphometry, intensity, texture, and gradient statistics. (B) To have the optimal representation of nuclear morphology, we plotted average absolute nuclear score difference associated with increasing numbers of selected features for the regression validation. We obtained a subset of discriminative features for Nuclear Score (NS) estimation for each given feature number, ranging from 2 to 23. The average absolute nuclear score difference reached a minimum with 12 selected features. (C) We also studied the histogram of average cross-validation errors associated with 12 individual features from the original feature set. Morphometry features had better performances than those of other categories. Of the morphometry features, eccentricity and circularity had the lowest NS estimation error (red dashed line). When the combined 12 features were used, the NS estimation error was lower than any single feature (green dashed line).
Figure 4.
Validation of Nuclear Scores (NS) by reviewing nuclear appearances.
(A) Spectrum of analyzed nuclei with distinct NS. Nuclei from feature-driven clusters (columns) were reviewed and the correlation between OC and NS was confirmed. (B) Arrays of GBM nuclei grouped by machine-classified NS. To validate results further, we aggregated 84 nuclei with (left panel) NS 1 and 84 nuclei with (right panel) NS 10, scored by machine-based regression analysis. Segmented nuclear boundaries (in green) produced by machine algorithms are overlaid. Visual and quantitative assessments verify that nuclei with NS 1 are typical of oligodendroglioma nuclei and those with NS 10 are typical of astrocytoma nuclei.
Figure 5.
Comparisons of Oligodendroglioma Component Percentages (OC%) in Human-annotated (HOC) and Machine-derived Oligodendroglioma Component (MOC) groups.
Estimated Gaussian distributions of OC%s for patients reviewed as OC 0, OC 1, and OC 2 groups by (A) TCGA neuropathologists; and (B) machine clustering approach, with NS intervals [1,2,6-10] for oligodendroglioma and astrocytoma nuclei.
Figure 6.
Estimated probability density functions associated with means of eccentricity and circularity.
Estimated probability density functions associated with means of eccentricity and circularity demonstrate significant differences between (A-B) HOC 0+1 and HOC 2; and (C-D) MOC 0+1 and MOC 2 groups. Note that the probability density function of eccentricity average presents lower population mean for OC 2 group and higher mean for OC 0+1 group. Similarly, the probability density function of circularity average presents higher population mean in the OC 2 group and lower mean for OC 0+1 group. The findings agree with the domain knowledge of oligodendroglioma and astrocytoma nuclear morphology.
Figure 7.
Analysis of survival with Human-annotated (HOC) and Machine-derived Oligodendroglioma Component (MOC) patient groups.
Kaplan-Meier plots of the survival of TCGA GBM patients classified as OC 0 vs. those in OC 1+2 group by (A) TCGA neuropatholgists (P = 0.49669); (B) machine algorithms (P = 0.42865). Kaplan-Meier plots of the survival of TCGA GBM patients classified as OC 2 vs. those in OC 0+1 group by (C) TCGA neuropatholgists (P = 0.44479); (D) machine algorithms (P = 0.30348).
Figure 8.
Analysis of treatment response with Human-annotated (HOC) and Machine-derived Oligodendroglioma Component (MOC) patient groups.
Treatment responses of patients receiving standard and aggressive therapies are shown for patients in the OC 1+2 group defined by (A) TCGA neuropatholgists (P = 0.48943); (B) machine algorithms (P = 0.07256); and within the OC 0+1 group defined by (C) TCGA neuropatholgists (P = 8.71e-3); (D) machine algorithms (P = 7.58e-3).
Figure 9.
Associations of Oligodendroglioma Component (OC) groups with molecular data.
(A) Genetic alteration profiles of GBMs in the three (left panel) HOC and (right panel) MOC groups. Mutation is depicted in red (upper row). Homozygous deletion (-2), hemizygous deletion (-1), no change (0), gain (1), and high-level amplification (2) conditions are represented in light green, dark green, black, dark red, and light red (lower row). (B-D) Analyses of MOC groups identify oligodendrocyte signature genes. (B) Heat map of gene expression (high expression in red) for four genes with significant overexpression in MOC 2 group compared to MOC 0+1 groups. (C) Heat map of expression profiles (high expression in red) of significant genes positively (upper half) and negatively (lower half) correlated with OC% by SAM. Patients are sorted on the x-axis by ascending OC%; (D) Smoothed expression profiles of the oligodendrocyte signature genes MBP, PLP1, HOXD1, MOBP, and PDGFRA are plotted with samples of increasing OC%.
Figure 10.
Correlation of feature means with gene expression of oligodendrocyte signature genes.
Distributions associated with eccentricity and circularity means are shown for groups of low and high expression of (A) HOXD1, (B) MBP, and (C) PLP1, highlighting the close association between the expressions of oligodendrocyte-specific genes and nuclear features typical of oligodendroglioma.