Evaluating Input Modalities for Non-Contrast MRI Brain Tumor Segmentation Via 3D AIR-UNet
Abstract
Accurate brain tumor segmentation traditionally relies on contrast-enhanced MRI. However, gadolinium-based contrast agents pose patient health risks and environmental toxicity. To develop a safer alternative, this study utilizes non-contrast multi-modality fusion images as input for a 3D AIR-UNet segmentation model. Findings reveal that while adding a wavelet fusion channel (4-channel model) slightly improves whole tumor (WT) segmentation, it introduces noise that degrades tumor core (TC) and enhancing tumor (ET) accuracy, inflating the Hausdorff distance (HD95). Thus, an optimized 3-channel non-contrast pipeline (T1, T2, T2-FLAIR) coupled with static thresholding post-processing is advocated to suppress false positives. The proposed model achieves a WT Dice score of 82.29%, comparable to the contrast-enhanced baseline (83.22%), and a competitive Dice score of 43.81% for the challenging ET sub-region. Ultimately, this framework demonstrates significant potential as a computationally efficient, ecologically responsible, and clinically safer alternative.