MIL normalization -- prerequisites for accurate MRI radiomics analysis.

Area of Science:

• Medical Imaging
• Radiomics
• Artificial Intelligence

Background:

• Magnetic resonance (MR) image quality varies due to different instruments and acquisition parameters, impacting radiomics analysis.
• Key variations include modality mismatch (M), intensity distribution variance (I), and layer-spacing differences (L), collectively termed MIL differences.
• Standardized, high-quality MR images are crucial for reliable radiomics insights.

Purpose of the Study:

• To develop and validate an MIL normalization system for reconstructing heterogeneous MR images into high-quality, standardized data.
• To assess the impact of MIL normalization on radiomics tasks, specifically tumor segmentation, pathological grading, and genetic diagnosis of glioma.
• To improve the accuracy and consistency of radiomics analysis by addressing acquisition variations.

Main Methods:

• Proposed an MIL normalization system comprising multimodal synthesis (encoder-decoder), intensity normalization (CycleGAN), and layer-spacing unification (SPM).
• Utilized three retrospective glioma datasets (BraTs, TCGA, HuaShan) for validation.
• Evaluated performance using Dice similarity coefficient, AUC, and other metrics for segmentation, grading, and genetic diagnosis.

Main Results:

• MIL normalization significantly improved radiomics task performance compared to non-normalized data.
• Achieved a 9% increase in Dice coefficient for tumor segmentation (P < .001).
• Enhanced AUC for pathological grading by 32% (P < .001) and IDH1 status prediction by 25% (P < .001).

Conclusions:

• The proposed MIL normalization system effectively standardizes heterogeneous MR images, producing high-quality data essential for accurate radiomics.
• MIL normalization demonstrably improves the performance of critical radiomics tasks in glioma analysis.
• This system offers a robust solution for overcoming acquisition variations in MR imaging for radiomics applications.