Shuffle-ResNet: Deep learning for predicting LGG IDH1 mutation from multicenter anatomical MRI sequences

Area of Science:

• Neuro-oncology
• Medical Imaging
• Computational Biology

Background:

• The World Health Organization recommends integrating gene information, like isocitrate dehydrogenase 1 (IDH1) mutation status, for improved central nervous system tumor management.
• Low-grade glioma (LGG) diagnosis and treatment can be enhanced by predicting IDH1 mutation status.

Purpose of the Study:

• To develop and evaluate a Shuffle Residual Network (Shuffle-ResNet) for predicting IDH1 gene mutation status in LGG.
• To utilize multicenter anatomical magnetic resonance imaging (MRI) sequences (T2-w, T2-FLAIR, T1-w, T1-Gd) for this prediction.

Main Methods:

• A dataset from The Cancer Genome Atlas LGG project (105 patients) was used, split into training and testing sets.
• A random image patch extractor and RGB dataset creation from image concatenation were employed to leverage tumor heterogeneity.
• A 3-fold cross-validation and random channel-shuffle layer in the ResNet architecture were used to improve generalization.

Main Results:

• The Shuffle-ResNet achieved an accuracy of 85.7% and an AUC of 0.943 on the test dataset.
• Validation accuracy reached 81.29% with an AUC of 0.96 when using a combined T2-FLAIR, T1-Gd, and T2-w RGB dataset.
• The early stopping algorithm optimized training duration across cross-validation folds.

Conclusions:

• The developed Shuffle-ResNet demonstrates the potential for predicting IDH1 gene mutation status from multicenter MRI data in LGG.
• Further investigation is necessary to establish the clinical applicability of this AI-driven approach.