Brain tumour segmentation with incomplete imaging data.

Area of Science:

• Neuro-oncology
• Medical Imaging
• Machine Learning

Background:

• Brain tumor heterogeneity complicates treatment and outcome prediction.
• Large-scale, phenotyped data, including imaging, is needed for individual patient prediction.
• Real-world clinical data often has quality limitations, posing challenges for machine learning.

Purpose of the Study:

• To evaluate the performance of automated brain tumor segmentation models using incomplete and real-world MRI data.
• To determine if models trained on incomplete data can achieve comparable accuracy to those trained on complete datasets.
• To assess the ability of segmentation models to detect enhancing tumors without contrast agents.

Main Methods:

• Applied state-of-the-art brain tumor segmentation models to a large-scale, multi-site MRI dataset (1251 individuals).
• Replicated various levels of data completeness to simulate real-world clinical scenarios.
• Validated model performance on a heterogeneous, real-world 50-patient sample with diverse imaging conditions.

Main Results:

• Models trained on incomplete MRI data achieved high segmentation accuracy (Dice coefficients 0.907-0.945 for whole tumors, 0.701-0.891 for tissue types).
• Segmentation performance with incomplete data was often equivalent to models trained on complete image sets.
• Models accurately detected enhancing tumors without contrast agents (R² > 0.97), regardless of lesion morphology.

Conclusions:

• Automated segmentation models are robust to incomplete MRI data, facilitating the use of historical and real-world clinical data for predictive modeling.
• These models can quantify tumor enhancement without contrast administration, potentially revising current imaging protocols.
• The findings support the translational application of quantitative imaging analysis in diverse clinical settings, including those with limited data or post-operative imaging.