Brain Lesion Synthesis via Progressive Adversarial Variational Auto-Encoder.

Area of Science:

• Medical imaging and artificial intelligence
• Neurosurgery and epilepsy treatment

Background:

• Laser interstitial thermal therapy (LITT) is a minimally invasive treatment for mesial temporal lobe epilepsy (MTLE).
• Accurate segmentation of lesions before and after LITT is crucial for assessing treatment efficacy.
• Deep learning models, like CNNs, require extensive annotated data for training, which is challenging to obtain for emerging treatments like LITT.

Purpose of the Study:

• To develop a progressive brain lesion synthesis framework (PAVAE) to expand training datasets for LITT-related lesion segmentation.
• To improve the quantity and diversity of available training data for deep learning models.

Main Methods:

• Proposed a two-stage framework: a mask synthesis network and a mask-guided lesion synthesis network.
• Introduced condition embedding blocks (CEB) and mask embedding blocks (MEB) to incorporate extrinsic information during training.
• Trained a segmentation network using both real and synthetically generated lesion images.

Main Results:

• The proposed PAVA E framework generated realistic synthetic brain lesion images.
• The synthetic data significantly boosted the performance of downstream lesion segmentation tasks.
• PAVAE outperformed traditional data augmentation techniques in enhancing segmentation accuracy.

Conclusions:

• The PAVA E framework effectively addresses the data scarcity issue in training deep learning models for LITT lesion segmentation.
• This method offers a viable solution for objective assessment of LITT treatment efficacy in MTLE.
• The developed framework has the potential to advance automated lesion quantification in neurosurgical applications.