A self-supervised learning strategy for postoperative brain cavity segmentation simulating resections.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neurosurgery

Background:

• Accurate segmentation of brain resection cavities (RCs) is crucial for postoperative analysis and treatment planning.
• Convolutional neural networks (CNNs) are effective for image segmentation but require large annotated datasets.
• Manual annotation of 3D medical images is labor-intensive, time-consuming, and prone to inter-rater variability.

Purpose of the Study:

• To develop a self-supervised learning strategy for training 3D CNNs to segment RCs in postoperative MRI.
• To leverage simulated resection data to overcome the need for extensive manual annotations.
• To evaluate the generalizability of the proposed method across different datasets and institutions.

Main Methods:

• Developed an algorithm to simulate resections from preoperative MRI scans.
• Employed self-supervised learning for training a 3D CNN on simulated resection data.
• Curated the EPISURG dataset (430 postoperative, 268 preoperative MRIs) from refractory epilepsy patients.
• Fine-tuned the model on small annotated datasets from multiple institutions and the EPISURG dataset.

Main Results:

• The self-supervised model trained on simulated data achieved median Dice score coefficients (DSCs) ranging from 74.9 to 82.4.
• After fine-tuning, DSCs improved to a range of 80.2 to 89.2 across datasets.
• The model's performance, particularly after fine-tuning, approached or exceeded the inter-rater agreement of human annotators (84.0).

Conclusions:

• A novel self-supervised learning strategy effectively segments real RCs in postoperative MRI using simulated data.
• The method demonstrates strong generalization capabilities across diverse institutional data, pathologies, and imaging modalities.
• The study provides open-source code, models, and the EPISURG dataset to facilitate further research.