Applications of artificial intelligence in the analysis of histopathology images of gliomas: a review

Abstract

In recent years, the diagnosis of gliomas has become increasingly complex. Analysis of glioma histopathology images using artificial intelligence (AI) offers new opportunities to support diagnosis and outcome prediction. To give an overview of the current state of research, this review examines 83 publicly available research studies that have proposed AI-based methods for whole-slide histopathology images of human gliomas, covering the diagnostic tasks of subtyping (23/83), grading (27/83), molecular marker prediction (20/83), and survival prediction (29/83). All studies were reviewed with regard to methodological aspects as well as clinical applicability. It was found that the focus of current research is the assessment of hematoxylin and eosin-stained tissue sections of adult-type diffuse gliomas. The majority of studies (52/83) are based on the publicly available glioblastoma and low-grade glioma datasets from The Cancer Genome Atlas (TCGA) and only a few studies employed other datasets in isolation (16/83) or in addition to the TCGA datasets (15/83). Current approaches mostly rely on convolutional neural networks (63/83) for analyzing tissue at 20x magnification (35/83). A new field of research is the integration of clinical data, omics data, or magnetic resonance imaging (29/83). So far, AI-based methods have achieved promising results, but are not yet used in real clinical settings. Future work should focus on the independent validation of methods on larger, multi-site datasets with high-quality and up-to-date clinical and molecular pathology annotations to demonstrate routine applicability.

Fig. 1. Schematic diagnostic criteria for adult-type diffuse glioma as defined by the 2021 WHO Classification of Tumors of the Central Nervous System.

The 2021 WHO classification system reorganizes gliomas into adult-type diffuse gliomas, pediatric-type diffuse low-grade and high-grade gliomas, circumscribed astrocytic gliomas, and ependymal tumors. Adult-type diffuse gliomas comprise Glioblastoma, IDH-wildtype, CNS WHO grade 4, Astrocytoma, IDH-mutant, CNS WHO grade 2–4, and Oligodendroglioma, IDH-mutant and 1p/19q-codeleted, CNS WHO grade 2–3. Identified adult-type diffuse glioma are stratified from left to right according to molecular alterations, including mutations in isocitrate dehydrogenase 1/2 (IDH1/2) genes and whole-arm codeletion of chromosomes 1p and 19q, as well as histological features, including increased mitoses, necrosis and/or microvascular proliferation. “or” is understood to be non-exclusive.

Fig. 2. Typical workflow for artificial intelligence-based analysis of WSIs.

Most approaches partition WSIs into smaller image patches (“Tiling”). Using AI methods, the patches are processed individually and then aggregated to obtain a prediction for the entire WSI. AI methods are often based on weakly-supervised learning or attention-based multiple-instance learning, each of which are depicted in greater detail.

Fig. 3. Number of studies by year of publication and diagnostic task.

All 70 studies included in this review are shown. Studies published in 2024 are considered up until March 18, 2024.

Fig. 4. Patch sizes and magnifications employed by studies.

Studies were taken into account if at least one of the two pieces of information was specified. Other patch sizes and magnifications employed by single studies (e.g., 150 × 150 pixels, 4x magnification) are not shown.

Fig. 5. Convolutional neural network architectures employed by studies.

With only a few exceptions all convolutional neural networks were pre-trained using the ImageNet dataset. Except for ResNet architectures exact variants of stated architectures are not shown. “Custom” refers to custom (i.e., self-configured) architectures.

Fig. 6. Learning paradigms employed by studies.

Usage of regions of interest, weakly-supervised learning and multiple-instance learning by all 70 deep learning-based studies included in this review distributed by year of publication. The methodological differences of the three approaches are explained in Table 2.