Multiproject-multicenter evaluation of automatic brain tumor classification by magnetic resonance spectroscopy

Area of Science:

• Neuroimaging
• Medical Diagnostics
• Machine Learning

Background:

• Automatic brain tumor classification using magnetic resonance spectroscopy (MRS) has been explored for over a decade.
• Previous evaluations of predictive models have lacked testing on unseen cases from different centers.
• The multicenter eTUMOUR project provided a unique opportunity to evaluate predictive models on diverse, unseen data.

Purpose of the Study:

• To evaluate the performance of predictive models for brain tumor classification using MRS data acquired across different institutions.
• To assess the generalizability of classifiers developed from the INTERPRET project data when applied to the eTUMOUR dataset.
• To determine the feasibility of using previously developed classifiers for assisting in new brain tumor diagnoses.

Main Methods:

• 253 pairwise classifiers were developed for glioblastoma, meningioma, metastasis, and low-grade glial tumors.
• Classifiers were trained on 211 short TE INTERPRET magnetic resonance spectra (MRS) acquired at 1.5 T.
• The trained classifiers were tested on 97 subsequently acquired spectra from the eTUMOUR project.

Main Results:

• Accuracies around 90% were achieved for most pairwise tumor classifications using unseen spectra.
• Discrimination between glioblastoma and metastasis yielded lower accuracies (<78%).
• Alternative methods like MRSI + MRI might offer clearer metastasis definition.

Conclusions:

• In-vivo MRS tumor type prediction is possible with classifiers trained on data from different hospitals and instrumentation, provided consistent acquisition protocols.
• This methodology can assist in diagnosing new brain tumor cases.
• The approach is valuable for quality control of multicenter MRS databases.