Deep Learning Glioma Grading with the Tumor Microenvironment Analysis Protocol for Comprehensive Learning, Discovering, and Quantifying Microenvironmental Features

Abstract

Gliomas are primary brain tumors that arise from neural stem cells, or glial precursors. Diagnosis of glioma is based on histological evaluation of pathological cell features and molecular markers. Gliomas are infiltrated by myeloid cells that accumulate preferentially in malignant tumors, and their abundance inversely correlates with survival, which is of interest for cancer immunotherapies. To avoid time-consuming and laborious manual examination of images, a deep learning approach for automatic multiclass classification of tumor grades was proposed. As an alternative way of investigating characteristics of brain tumor grades, we implemented a protocol for learning, discovering, and quantifying tumor microenvironment elements on our glioma dataset. Using only single-stained biopsies we derived characteristic differentiating tumor microenvironment phenotypic neighborhoods. The study was complicated by the small size of the available human leukocyte antigen stained on glioma tissue microarray dataset - 206 images of 5 classes - as well as imbalanced data distribution. This challenge was addressed by image augmentation for underrepresented classes. In practice, we considered two scenarios, a whole slide supervised learning classification, and an unsupervised cell-to-cell analysis looking for patterns of the microenvironment. In the supervised learning investigation, we evaluated 6 distinct model architectures. Experiments revealed that a DenseNet121 architecture surpasses the baseline's accuracy by a significant margin of 9% for the test set, achieving a score of 69%, increasing accuracy in discerning challenging WHO grade 2 and 3 cases. All experiments have been carried out in a cross-validation manner. The tumor microenvironment analysis suggested an important role for myeloid cells and their accumulation in the context of characterizing glioma grades. Those promising approaches can be used as an additional diagnostic tool to improve assessment during intraoperative examination or subtyping tissues for treatment selection, potentially easing the workflow of pathologists and oncologists.

Keywords: Automated glioma grading; Deep learning; Human leukocyte antigen; Quantification of tumor microenvironment elements; Tissue microarrays.

Fig. 1

Examples of Human Leukocyte Antigen (HLA)-stained tissue microarray’s cores. The chosen examples have clear histological differences. However, in some cases, differences in the highest grade are difficult to grasp with the naked eye

Fig. 2

Performance averaged across 10 folds of each model. The “HSV” suffix indicates the model tested on input images in HSV color space. Scores not reaching 50% were marked in gray, between 50 and 60% in light red, and those exceeding 60% in light blue

Fig. 3

Confusion matrices. On the left: ResNet18 pretrained evaluation on the test set. On the right: DenseNet121 pretrained with HSV color space input evaluation on the test set

Fig. 4

View of clusters (neighborhoods). This figure presents 2 frames made of patches of the tumor images that were assigned together within 2 groups during patch contrastive learning protocol and resulted in the 2 most distinctive clusters of cell neighborhoods

Fig. 5

Cell type abundance. TME composition of fields — a heatmap showing the sum-to-1 normalized distribution of neighborhoods in 206 images/patients, colored by label (tumor grade). Both images/patients and neighborhoods are ordered by hierarchical clustering. The most abundant phenotypes are P4 and P3. The most abundant neighborhood is N2 — mostly in grade 4 and 2

Fig. 6

Violin plots showing a relative abundance of clusters. The plots represent how significant are particular differences between cluster abundance between tumor grades. The distribution of each plot illustrates the variation and density of the cluster abundance, emphasizing the statistical significance of group differences. Marking of *** indicates p-value of<0.001, marking of ** the p-value of<0.01, and marking of * the p-value of<0.05

Fig. 7

Graph of clusters of tumor microenvironment features. On the left: the undirected graph with connections linking tumors of various grades where occurred significant differences between cluster abundances. The bold line indicates two grades with the most different characteristics. On the right: a weighted sum of common clusters occurring with significant differences in the abundance between tumors of different grades. The diagram on the right is based on the same results as the graph on the left, but it more strongly represents the number and significance of the differences between grades (when the p-value is<0.01 the number of shared clusters is multiplied by 2, and when the p-value is<0.001 — the number of shared clusters is multiplied by 3)

Fig. 8

Left: UMAPs of tumor grades. UMAP plots with color-coded labels for tumor grades and neighborhoods. Starting from the left: 0, represents tissue “grade 0”; 1, tumor WHO grade 1; 2, tumor WHO grade 2; 3, tumor WHO grade 3; and 4, tumor WHO grade 4. Right: UMAPs of neighborhoods: 0, denotes N1 neighborhood; 1, N2 neighborhood; 2, N3 neighborhood; 3, N5 neighborhood; 4, N6 neighborhood; 5, N7 neighborhood; 6, N8 neighborhood; and 7, N9 neighborhood