Assessing Impact of Data Quality in Early Post-Operative Glioblastoma Segmentation.

Area of Science:

• Neuroimaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Accurate quantification of residual glioblastoma tumor post-surgery is vital for patient management.
• Early post-operative magnetic resonance imaging (MRI) segmentation is challenging due to small, fragmented lesions and noise.

Purpose of the Study:

• To assess the impact of image and annotation quality on deep learning model performance for early post-operative glioblastoma segmentation.
• To evaluate how different data quality levels influence model training and generalization.

Main Methods:

• An Attention U-Net model was trained using a dataset of 423 early post-operative MRI scans from two hospitals.
• Five-fold cross-validation was employed on subsets stratified by image and annotation quality.
• Expert neurosurgeons evaluated image and annotation quality.

Main Results:

• Models trained on exclusively high-quality annotations achieved performance comparable to models trained on the full dataset.
• Excluding low-quality images from training did not harm performance on high-quality images.
• Models trained solely on high-quality images failed to generalize to low-quality data.

Conclusions:

• Both image and annotation quality significantly impact segmentation model performance.
• Dataset curation should include images representative of real-world quality variations.
• Prioritizing high-quality ground truth annotations is essential for robust model development.