A Deep Learning Framework for Pediatric TLE Detection Using 18F-FDG-PET Imaging

Purpose:The key to successful epilepsy control involves locating epileptogenic focus before treatment. 18F-FDG PET has been considered as a powerful neuroimaging technology used by physicians to assess patients for epilepsy. However, imaging quality, viewing angles, and experiences may easily degrade the consistency in epilepsy diagnosis. In this work, the investigators develop a framework that combines radiomics analysis and deep learning techniques to a computer-assisted diagnosis (CAD) method to detect epileptic foci of pediatric patients with temporal lobe epilepsy (TLE) using PET images.

Methods:Ten PET radiomics features related to pediatric temporal bole epilepsy are first extracted and modelled. Then a neural network called Siamese network is trained to quanti-fy the asymmetricity and automatically locate epileptic focus for diagnosis.The performance of the proposed framework was tested and compared with both the state-of-art clinician software tool and human physicians with different levels of experiences to validate the accuracy and consistency.

Inclusion criteria

• Clinical diagnosis of temporal lobe epilepsy.
• Age range from six to eighteen years old.
• Underwent PET, EEG, computed tomography (CT) and MRI.

Exclusion criteria

• Image quality is unsatisfactory (e.g. severe image artifacts due to head movement).
• 18F-FDG PEG examination is negative.
• Clinical data is incomplete.
• EEG or MRI report is missing.