An integrated MRI-based diagnostic framework for glioma with incomplete imaging sequences and imperfect annotations

Area of Science:

• Artificial Intelligence in Medicine
• Neuro-oncology
• Medical Imaging Analysis

Background:

• Glioma diagnosis relies on integrating histopathology with molecular features (IDH, 1p/19q), but multi-center studies face data heterogeneity due to varied imaging protocols and molecular testing.
• Inconsistent data, including incomplete imaging and missing annotations, impede the development of robust AI diagnostic tools.

Purpose of the Study:

• To propose SSL-MISS-Net, a unified framework designed to simultaneously address input-side modality incompleteness and output-side annotation deficiencies in glioma data.
• To reduce the reliance on complete, perfectly annotated datasets for training AI models in neuro-oncology.

Main Methods:

• Developed SSL-MISS-Net (Self-Supervised Learning with MIssing-label encoding and Semantic Synthesis), a framework employing cross-modal self-supervised learning for modality incompleteness.
• Integrated a missing-label synergistic strategy to handle annotation deficiencies, enabling the use of imperfect clinical data.

Main Results:

• SSL-MISS-Net achieved significant accuracy gains of 4% (validation) and 10% (test) for integrated glioma diagnosis compared to sub-optimal methods.
• The framework expanded the amount of clinically usable data by 256% and outperformed state-of-the-art methods trained on complete data.
• Demonstrated exceptional resilience to data imperfections in multi-center cohorts (N=2238) across six in-house and three public datasets.

Conclusions:

• SSL-MISS-Net is the first framework to jointly address modality incompleteness and missing labels in glioma data, unlocking the potential of imperfect clinical datasets.
• The proposed method shows strong clinical translatability and robust performance in neuro-oncology AI diagnostics, overcoming common data challenges.