Clinical Applications of Artificial Intelligence, Machine Learning, and Deep Learning in the Imaging of Gliomas: A Systematic Review

Abstract

In neuro-oncology, magnetic resonance imaging (MRI) is a critically important, non-invasive radiologic assessment technique for brain tumor diagnosis, especially glioma. Deep learning improves MRI image characterization and interpretation through the utilization of raw imaging data and provides unprecedented enhancement of images and representation for detection and classification through deep neural networks. This systematic review and quality appraisal method aim to summarize deep learning approaches used in neuro-oncology imaging to aid healthcare professionals. Following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines, a total of 20 low-risk studies on the established use of deep learning models to identify glioma genetic mutations and grading were selected, based on a Quality Assessment of Diagnostic Accuracy Studies 2 score of ≥9. The included studies provided the deep learning models used alongside their outcome measures, the number of patients, and the molecular markers for brain glioma classification. In 19 studies, the researchers determined that the deep learning model improved the clinical outcome and treatment protocol in patients with a brain tumor. In five studies, the authors determined the sensitivity of the deep learning model used, and in four studies, the authors determined the specificity of the models. Convolutional neural network models were used in 16 studies. In eight studies, the researchers examined glioma grading by using different deep learning models compared with other models. In this review, we found that deep learning models significantly improve the diagnostic and classification accuracy of brain tumors, particularly gliomas without the need for invasive methods. Most studies have presented validated results and can be used in clinical practice to improve patient care and prognosis.

Texture Analysis of Enhanced MRI and Pathological Slides Predicts EGFR Mutation Status in Breast Cancer

“…This method has made progress not only in the identification of EGFR status or breast cancer but also in gastrointestinal tumors, gliomas, reproductive system tumors, respiratory system tumors, etc. [ 56 – 58 ]. In both qualitative and quantitative medical problems, this approach has undisputed obvious advantages [ 59 , 60 ].…”

Recapitulating the Key Advances in the Diagnosis and Prognosis of High-Grade Gliomas: Second Half of 2021 Update

“…Under optimal organization conditions, those images allow a neuro-oncological disease management team (DMT) to perform initial classification of the tumor and plan its possible treatment [ 68 , 69 ]. These plans are often complicated by uncertainties relating to the natural history of the tumor (e.g., whether or not it is due to the evolution of lower-grade lesions) and by the presence of outcomes of any previous treatments (e.g., necrosis or pseudoprogression) [ 70 , 71 ]. In the last decade, some molecular characteristics have been identified, making the diagnosis and prognosis of HGG more precise [ 72 , 73 ].…”

Recognizing schizophrenia using facial expressions based on convolutional neural network

“…This brings new dimensions to clinical research and brings convenience to clinicians. Among many deep learning algorithms, the convolutional neural network (CNN) highlights the advantages of high accuracy in image recognition and classification (Alhasan, 2021). The advantages of CNN are as follows:…”

“…This brings new dimensions to clinical research and brings convenience to clinicians. Among many deep learning algorithms, the convolutional neural network (CNN) highlights the advantages of high accuracy in image recognition and classification (Alhasan, 2021). The advantages of CNN are as follows: (1) it is able to preserve spatial properties of images due to their highly parameterized and sparsely connected kernels; (2) it learns through labeled images and identifies important features without explicitly specifying them; and (3) it learns a representation of input data as the information flow ascends through multiple layers.…”