Affiliations: [a] State Key Laboratory of Reliability and Intelligence of Electrical Equipment, Tianjin, China | [b] School of Artificial Intelligence and Data Science, Hebei University of Technology, Tianjin, China | [c] Tianjin Universal Medical Imaging Diagnostic Center, Tianjin, China | [d] Department of Molecular Imaging and Nuclear Medicine, Tianjin Medical University Cancer Institute and Hospital, Tianjin, China | [e] Traditional Chinese Medicine Hospital of Beichen District, Tianjin, China | [f] Key Laboratory of Big Data Computing, Hebei University of Technology, Tianjin, China | [g] School of Electrical Engineering, Hebei University of Technology, Tianjin, China
Correspondence:
[*]
Corresponding author: Xu Guizhi, School of Electrical Engineering, Hebei University of Technology, Tianjin 300130, China. E-mail: [email protected]
Abstract: Epilepsy is one of the most serious neurological diseases in the world. The mesial temporal lobe epilepsy (MTLE), especially hippocampal sclerosis (HS) is the most common pathological causes of epilepsy. With the development of computer visualization technology, there are many auxiliary diagnostic approaches based on deep learning algorithms. However, the traditional 2D-CNN framework can only accept single layer inputs. In such case, the associations between the brain planes are ignored, which may lead to misdiagnosis or missed diagnosis. 3D-CNN framework can accept cubes as the input of the neural network, so that network parameters will carry more structural and logical information of the brain in the spatial domain. Therefore, this study designed a 3D-CNN framework for MTLE diagnosis in T2-FLAIR MRI images. We retrospectively collected 15 patients with the MTLE and 15 age-matched controls who underwent T2-FLAIR studies. Then, we proposed three 3D-CNN based on ResNet to identify symmetrical differences in the corresponding areas of the brain in both sides. The symmetrical cubes were combined as the inputs for the 3D-CNN framework. Performances of the proposed framework were compared with radiomics algorithms and visual assessment. The proposed 3D-CNN based on ResNet-34 performs the best among all the algorithms. Moreover, due to the non-inferiority testing for paired data, the proposed 3D-CNN frameworks based on the ResNet were not inferior to that of visual assessment which was unblinded to the clinical information. The proposed 3D-CNN framework could diagnosis MTLE in MRI images accurately and efficiently, which might be applied as a computer-assisted approach for the future diagnosis of epilepsy patients.
