Deep Learning Approaches for Imaging-Based Automated Segmentation of Tuberous Sclerosis Complex.

Area of Science:

• Neurology
• Medical Imaging
• Artificial Intelligence

Background:

• Tuberous Sclerosis Complex (TSC) is a genetic disorder causing benign tumors (tubers) in multiple organs.
• Identifying epileptogenic tubers is crucial for effective treatment and surgical planning in TSC patients.
• Current methods for tuber segmentation and identification can be labor-intensive and subjective.

Purpose of the Study:

• To develop and validate a novel approach for identifying epileptogenic tubers in TSC.
• To automate tuber segmentation using a three-dimensional convolutional neural network (3D CNN).
• To investigate neuroimaging characteristics differentiating epileptogenic from non-epileptogenic tubers.

Main Methods:

• Retrospective analysis of neuroimaging data from 31 TSC patients.
• Manual annotation of lesions and determination of epileptogenicity via presurgical evaluation and stereoelectroencephalography.
• Extraction and comparison of neuroimaging metrics between epileptogenic and non-epileptogenic tubers.
• Development and training of 3D CNNs on five datasets with varying preprocessing strategies for automated segmentation.

Main Results:

• Epileptogenic tubers showed significantly lower normalized PET metabolic values (p = 0.001).
• 3D CNNs achieved high localization accuracy (0.992-0.994) across datasets.
• Automated segmentations demonstrated high correlation with clinician-based features and excellent agreement with reference segmentations.

Conclusions:

• Neuroimaging characteristics can effectively differentiate epileptogenic tubers in TSC.
• The developed 3D CNN approach provides accurate and automated tuber segmentation.
• This deep learning tool enhances surgical confidence and offers clinical value in managing TSC.