Robust Deep Learning-based Segmentation of Glioblastoma on Routine Clinical MRI Scans Using Sparsified Training.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuro-oncology

Background:

• Glioblastoma segmentation using deep learning is crucial for treatment planning.
• Clinical datasets often present challenges due to missing imaging sequences.
• Improving model robustness on heterogeneous and incomplete data is essential.

Purpose of the Study:

• To enhance the robustness of deep learning models for glioblastoma segmentation.
• To evaluate the impact of sparsified training on model performance with incomplete datasets.
• To assess the utility of site-specific data in improving segmentation accuracy.

Main Methods:

• Retrospective analysis of multimodal MRI from 117 (BraTS) and 634 (clinical) glioblastoma patients.
• Training a convolutional neural network (DeepMedic) on complete and incomplete datasets, with and without site-specific data.
• Introduction of sparsified training to simulate missing sequences during model training.

Main Results:

• A model trained solely on BraTS data performed poorly on clinical data (Dice score 0.49).
• Sparsified training significantly improved performance on incomplete data (median Dice score 0.67).
• Incorporating site-specific data during sparsified training achieved Dice scores >0.8, comparable to models using complete data.

Conclusions:

• Accurate, automatic glioblastoma segmentation on clinical scans is achievable with models trained on large, heterogeneous, incomplete datasets.
• Sparsified training can significantly boost the performance of models trained on limited public and site-specific data.
• Deep learning models can be robustly adapted for clinical glioblastoma segmentation despite data variability.