Direct image to subtype prediction for brain tumors using deep learning

Abstract

Background: Deep Learning (DL) can predict molecular alterations of solid tumors directly from routine histopathology slides. Since the 2021 update of the World Health Organization (WHO) diagnostic criteria, the classification of brain tumors integrates both histopathological and molecular information. We hypothesize that DL can predict molecular alterations as well as WHO subtyping of brain tumors from hematoxylin and eosin-stained histopathology slides.

Methods: We used weakly supervised DL and applied it to three large cohorts of brain tumor samples, comprising N = 2845 patients.

Results: We found that the key molecular alterations for subtyping, IDH and ATRX, as well as 1p19q codeletion, were predictable from histology with an area under the receiver operating characteristic curve (AUROC) of 0.95, 0.90, and 0.80 in the training cohort, respectively. These findings were upheld in external validation cohorts with AUROCs of 0.90, 0.79, and 0.87 for prediction of IDH, ATRX, and 1p19q codeletion, respectively.

Conclusions: In the future, such DL-based implementations could ease diagnostic workflows, particularly for situations in which advanced molecular testing is not readily available.

Keywords: IDH; adult-type diffuse gliomas; deep learning; molecular signatures; subtype.

Figure 1.

Overview of our experimental approach. The flow-chart Figure 1A and 1B outline the 2016 and 2021 WHO diagnostic algorithm for diffuse adult-type gliomas, for the targets that we included in our experiments. In the 2016 algorithm, gliomas were tested for IDH mutation status before the morphological features were assessed to determine grade. Both IDHmut and IDHwt tumors with high-grade (grade IV) features were designated glioblastoma. Tumors with lower grade (grade II or III) morphology were designated as astrocytoma, unless 1p19q codeletion was present. Lower grade tumors with IDHmut and 1p19q codeletion were designated oligodendrogliomas. In the 2021 system, molecular alteration status determines the subtype of glioma. Astrocytomas are IDHmut and ATRXmut. If CDKN2A or CDKN2B homozygous deletion are additionally present in an astrocytoma, this automatically upgrades the tumor to grade 4. Absence of CDKN2A/B homozygous deletion and absence of high-grade morphology indicates a grade 2 or 3 astrocytoma. Oligodendrogliomas are IDHmut, 1p19q codeleted (complete loss of both arms), TERTmut and ATRXwt. Glioblastomas are IDHwt with either classic morphology on histology (microvascular proliferation and/or necrosis) or at least one of TERTmut, EGFR amplification, +7/−10 genotype. Abbreviations: MUT, mutated form; WT, wildtype form; ND, not deleted. The doughnut graphs in Figure 1C and 1D shows the data split by both 2016 and 2021 subtype diagnosis for the UCL and TCGA cohorts, respectively. Subtype diagnosis was available for both 2016 and 2021 criteria for the UCL dataset. In the TCGA dataset, the molecular alteration status for IDH, 1p19q and/or ATRX was used to formulate the 2021 subtype. Figure 1E provides an overview of our Deep Learning pipeline. The first step is preprocessing where digital WSIs are tessellated into tiles and features extracted for each tile. These features are then given to our network and used for training, testing, or deployment, depending on the type of experiment being run. Further details of our Deep Learning methods can be found in Supplementary Methods.

Figure 2.

Results for subtype experiments. This figure shows Receiver Operating Characteristic (ROC) Curves and Confusion Matrices (CM) for both 2016 and 2021 subtype experiments. Subtype experiments were run as a single experiment. The AUROCs visualize results for each subtype individually, whereas the CMs and additional statistics in Figure 4 relate to overall model performance. CMs and additional statistics were calculated with a threshold of 0.5. In each ROC plot (A-C, E-G), the thin lines indicate ROC curves for internal validation experiments. Internal validation was performed as five-fold cross-validation. The line with shading indicates the external validation results; where the line is the external validation ROC curve and the shaded area around this line indicates the bootstrap CI. The AUC ± bootstrap CI is given in the bottom right of each plot. Please note, AUC refers to the area under the ROC curve and is thus the same as AUROC. D and H are heatmap confusion matrices for the 2016 and 2021 subtype experiments, respectively. The confusion matrices are constructed from the model prediction output for the external validation experiments, i.e., the class with the highest probability in external validation was selected as the predicted class.

Figure 3.

Results for molecular alteration experiments. (A)–(D) show Receiver Operating Characteristic (ROC) Curves for the molecular alteration experiments. In each ROC plot, the thin lines indicate ROC curves for internal validation experiments. Internal validation was performed as five-fold cross-validation. The line with shading indicates the external validation results; where the line is the external validation ROC curve and the shaded area around this line indicates the bootstrap CI. The AUC ± bootstrap CI is also given in the bottom right of each plot. Please note, AUC refers to the area under the ROC and is thus the same as AUROC. E–H are confusion matrices (CMs) for each molecular alteration experiment. The CMs are constructed from the model prediction output for the external validation experiments, i.e., the class with the highest probability in external validation was selected as the predicted class. CMs were calculated with a threshold of 0.5. Abbreviations: MUT, mutant; WT, wildtype; ALT, altered; UA, unaltered.

Figure 4.

Statistical heatmaps. This is a heatmap of the further statistical analysis for our results. (A) shows results for the direct prediction of the 2021 subtypes. (B) shows results for the final prediction following the sequential approach. Here, the 2021 subtype was calculated by stacking the final predictions for the IDH, ATRX, and 1p19q experiments. Results for IDH prediction were considered first, followed by ATRX and 1p19q which were assessed together. Statistics were calculated with a threshold of 0.5.