Data Harmonization with StyleTransfer-GANs: Enhancing Non-Invasive IDH Classification in Brain Tumors.

Area of Science:

• Neuro-oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Isocitrate dehydrogenase (IDH) mutation status is a key prognostic marker in brain gliomas.
• Accurate, non-invasive IDH status determination is vital for patient management.
• Variability in MRI acquisition protocols across institutions limits the reliability of deep learning (DL) models for IDH classification.

Purpose of the Study:

• To develop and evaluate a StyleTransfer-GAN (ST-net) for harmonizing multi-site MRI data.
• To assess the impact of data harmonization on the performance of a DL network (MC-net) for IDH classification.

Main Methods:

• Development of ST-net, a StyleTransfer-GAN, to harmonize multi-site MRI datasets.
• Application of ST-net to diverse MRI databases while preserving essential imaging features.
• Evaluation of IDH classification accuracy using MC-net before and after ST-net harmonization.

Main Results:

• ST-net effectively harmonized style features across different MRI datasets, reducing protocol-induced discrepancies.
• Post-harmonization, IDH classification accuracy significantly improved.
• Enhancements in sensitivity and specificity were observed, varying with the style transfer reference.

Conclusions:

• Data harmonization using GANs, like ST-net, can significantly improve the generalizability and clinical utility of DL models in neuro-oncology.
• This approach offers a scalable solution for enhancing neuroimaging tasks, particularly for IDH classification in brain gliomas.