Deep learning-based postoperative glioblastoma segmentation and extent of resection evaluation: Development, external...

Area of Science:

• Neuro-oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Assessing glioblastoma resection extent is crucial but challenging, often relying on preoperative scans.
• Postoperative assessment requires precise measurement of residual tumor volume, a gap in current automated methods.
• Developing AI for postoperative tumor segmentation is essential for accurate evaluation.

Purpose of the Study:

• To develop a deep learning model for segmenting glioblastomas on postoperative MRI.
• To compare the performance of the developed model against existing segmentation algorithms.
• To evaluate the model's accuracy in classifying the extent of resection (EOR).

Main Methods:

• Trained deep learning models using multiparametric MRI scans from multiple institutions and public databases.
• Utilized MONAI and nnU-Net frameworks for model development and training.
• Compared model performance with existing algorithms on external and independent validation cohorts, assessing segmentation and EOR classification accuracy.

Main Results:

• The nnU-Net framework yielded the best segmentation model, achieving median Dice scores of 0.81 for enhancing tumor (ET), 0.77 for edema, and 0.81 for surgical cavities.
• The best-trained model achieved 96% accuracy in classifying maximal versus submaximal resection in the comparison dataset and 84% in the independent validation cohort.
• The model demonstrated superior performance in both segmentation and EOR classification compared to other algorithms.

Conclusions:

• A novel nnU-Net-based deep learning model effectively performs postoperative glioblastoma segmentation and EOR classification.
• The developed model shows superior performance over existing methods, offering a valuable tool for neurosurgical oncology.
• This freely accessible tool has significant potential for clinical application in improving glioblastoma patient management.