Trends in Development of Novel Machine Learning Methods for the Identification of Gliomas in Datasets That Include...

Area of Science:

• Neuroimaging
• Artificial Intelligence
• Oncology

Background:

• Machine learning (ML) is increasingly used in diagnostic imaging for gliomas, aiding classification, prognostication, segmentation, and treatment planning.
• This systematic review focuses on ML applications for identifying gliomas within mixed datasets, simulating real-world clinical scenarios.

Approach:

• A comprehensive literature search was conducted across four major databases (Embase, MEDLINE, CENTRAL, Web of Science) up to February 1, 2021.
• The search combined terms for artificial intelligence, ML, deep learning, radiomics, magnetic resonance imaging, and glioma.
• Included studies were screened, full texts reviewed, data extracted, and reporting quality assessed using TRIPOD criteria.

Key Points:

• Twelve articles were included from 11,727 initial candidates, analyzing ML for glioma detection in normal/abnormal and glioma/non-glioma image differentiation.
• Neural networks were the predominant algorithm type (10 studies), achieving a median accuracy of 0.96.
• Reporting quality was moderate, with a mean TRIPOD ratio of 0.50, indicating room for improvement.

Conclusions:

• Current ML applications for glioma identification in mixed datasets face significant limitations, including small, non-diverse datasets and inadequate algorithm training/testing strategies.
• Poor reporting quality further impedes clinical translation.
• Future research requires larger, more heterogeneous datasets and external validation for ML algorithms in neuro-oncology.