Transferring Knowledge from High-Quality to Low-Quality MRI for Adult Glioma Diagnosis

Yanguang Zhao, Long Bai, Zhaoxi Zhang, Yanan Wu, Mobarakol Islam, Hongliang RenarXiv 2024 · DOI · arXiv

Introduction: The Global Imaging Divide

Glioma remains one of the most challenging brain tumors to treat, with mortality rates in Sub-Saharan Africa (SSA) rising by nearly 25% in recent years, contrasting with declines in the Global North. A significant factor in this disparity is the "imaging divide." While medical AI has made strides in tumor segmentation, most state-of-the-art models are trained on high-fidelity datasets from well-resourced institutions. In contrast, clinical settings in SSA often utilize older MRI hardware, resulting in images with significantly lower resolution and poorer contrast.

Visual comparison between high-quality BraTS 2021 scans and low-quality BraTS-Africa scans across axial, coronal, and sagittal planes.

When models optimized for "pristine" data are applied to these "noisy" scans, their performance typically degrades. This study investigates strategies to bridge this gap, specifically focusing on how to adapt high-performing AI architectures to the unique constraints of African MRI data.

Methodology: Data Utilization and Model Architecture

The researchers utilized an expanded version of the nnU-Net architecture, which won the 2021 BraTS Challenge. Key modifications included doubling the filter count in the encoder and implementing Group Normalization instead of Batch Normalization to ensure stable training with the small batch sizes common in 3D medical imaging.

To determine the most effective way to handle low-quality data, three training strategies were evaluated:

  1. Transfer Learning ($S_{GLI \to SSA}$): Pre-training the model on the large, high-quality BraTS-GLI 2021 dataset before fine-tuning on the smaller BraTS-Africa dataset.
  2. Native Training ($S_{SSA}$): Training the model from scratch using only the low-quality African data.
  3. Super-Resolution ($S_{srSSA}$): Artificially upscaling the African images by 2x using a 20-layer residual convolutional neural network prior to training.
Overview of the model training strategies and segmentation inference pipeline, highlighting the transfer learning and ensemble approach.

The final inference process employed an ensemble of an original U-Net and the larger U-Net variant to maximize segmentation accuracy.

Results and Discussion: The Efficacy of Transfer Learning

The transfer learning approach ($S_{GLI \to SSA}$) emerged as the superior strategy, securing second place in the BraTS-Africa Challenge. On the validation set, the model achieved impressive Dice Scores: 0.926 for the Whole Tumor (WT), 0.882 for the Enhancing Tumor (ET), and 0.840 for the Tumor Core (TC). Furthermore, the Hausdorff Distance (95%) for the whole tumor was reduced to 13.971, a significant improvement over the 24.742 achieved by the model trained solely on African data.

These results suggest that pre-training on high-quality data allows the model to learn fundamental anatomical and pathological features that generalize across different image qualities. Conversely, the super-resolution strategy performed the worst. This "super-resolution paradox" likely stems from the fact that low-quality source images contain an inherent loss of information that artificial upscaling cannot recover; instead, the process may introduce artifacts that mislead the segmentation model.

Conclusion and Future Directions

This research demonstrates that transfer learning from high-fidelity datasets is a robust and effective method for improving medical AI performance in resource-limited settings. While advanced techniques like super-resolution are promising, they currently struggle with the information loss present in low-quality clinical scans.

Future work should explore more nuanced domain adaptation techniques that can better align the feature spaces of high- and low-quality scans. Bridging the imaging divide is not just a technical challenge but a critical step toward global health equity in neuro-oncology.

Summary updated

BibTeX
@misc{zhao2024transferring,
  author = {Yanguang Zhao and Long Bai and Zhaoxi Zhang and Yanan Wu and Mobarakol Islam and Hongliang Ren},
  title = {{Transferring Knowledge from High-Quality to Low-Quality MRI for Adult Glioma Diagnosis}},
  year = {2024},
  eprint = {2410.18698},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2410.18698},
}
Figure