Automated post-operative brain tumour segmentation: A deep learning model based on transfer learning from...

Area of Science:

• Medical imaging analysis
• Artificial intelligence in oncology
• Neurosurgery

Background:

• Automated brain tumour segmentation from post-operative images is crucial but challenging.
• Accurate segmentation of tumour subregions is vital for effective treatment planning and monitoring.

Purpose of the Study:

• To develop and evaluate an automated method for segmenting brain tumour subregions from multimodal post-operative MRI scans.
• To improve the accuracy and robustness of brain tumour segmentation models, particularly for small datasets.

Main Methods:

• A 3D densely-connected U-net model was developed for brain tumour segmentation.
• Transfer learning was employed, initially training on the BraTS dataset and then fine-tuning on a local dataset of post-operative scans.
• Ensemble learning was utilized to combine model outputs for enhanced accuracy and robustness.

Main Results:

• The model achieved high Dice Scores on the BraTS dataset (0.90 for whole tumour, 0.83 for tumour core, 0.78 for enhancing tumour).
• Despite a small local dataset, the final model demonstrated strong performance with Dice Scores of 0.83 (whole tumour), 0.77 (tumour core), and 0.60 (enhancing core).

Conclusions:

• The developed automated segmentation method shows significant potential for clinical application in post-operative brain tumour analysis.
• Transfer learning and ensemble techniques enable effective brain tumour segmentation even with limited post-operative data.