Keywords:
Staging, Computer Applications-Detection, diagnosis, Neural networks, MR, CT, Neuroradiology brain, CNS, Artificial Intelligence, Segmentation, Cancer, Neoplasia, Tissue characterisation
Authors:
A. Nanapragasam1, P. de Souza2, H. Harvey2; 1Newcastle/UK, 2London/UK
DOI:
10.26044/ecr2019/C-3246
Conclusion
We hope to have demonstrated the utility of machine learning and deep learning in the creation of a unified workflow for the non-invasive prognostication of glioma.
From automated tumour segmentation,
to the generation of clinically useful quantitative radiomic biomarkers,
and finally to the creation of prognostic models,
machine learning has demonstrated its role in the provision of personalised medicine.
The early work in this field is promising,
but the work is still in its infancy.
The numbers of patients used for machine learning training is relatively low,
and often from a single centre.
For greater validity and reduction of bias,
future work will have to explore means of analysing much larger volumes of data.
We also note that the means of data collection varies from study to study,
this precludes a statistical cross-centre validation.
Further research groups may want to consider taking a standardised approach such that comparison can be made across papers.
While the majority of papers reviewed considered imaging as their sole input to the machine learning algorithms,
there is benefit to incorporating clinical,
histopathological and biochemical factors.
This should further increase the accuracy of these models.
The existing body of work supports the principle of machine learning as a tool for glioma prognostication.
However,
further work on a larger scale is necessitated before widespread uptake of this tool can be advocated.