Evaluation of the HD-GLIO Deep Learning Algorithm for Brain Tumour Segmentation on Postoperative MRI

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Deep learning for 3D brain tumor segmentation shows promise for routine radiological workflows.
• Accurate segmentation is crucial for assessing treatment response in brain tumors.

Purpose of the Study:

• To externally evaluate the HD-GLIO deep learning algorithm's performance on an independent cohort of post-operative patients.
• To assess the algorithm's accuracy in segmenting contrast-enhancing (CE) and non-enhancing (NE) brain tumor lesions.

Main Methods:

• Compared HD-GLIO segmentations with radiologist delineations on 66 MRI scans.
• Utilized Dice similarity coefficients (Dice) for segmentation accuracy and concordance correlation coefficients (CCCs) for volume agreement.

Main Results:

• HD-GLIO achieved high performance for NE volumes (median Dice = 0.79) and large CE tumor volumes (>1.0 cm³; median Dice = 0.86).
• Performance significantly decreased for all CE tumor lesions (median Dice = 0.40).
• Excellent volume agreement was observed (CCCs of 0.997 for CE, 0.922 for NE).

Conclusions:

• The HD-GLIO algorithm shows potential but requires further validation, especially for small contrast-enhancing tumors in routine clinical use.
• Independent validation on diverse clinical datasets is essential to confirm the robustness of deep learning segmentation algorithms.