Segmentation of Epileptic Focus from Multi-Channel Magnatic Resonance Images via Deep Learning

1999 

Abstract Submission 

Xiaodong Zhang^1,2, Dezhi Cao², Jinping Xu¹

¹Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, Shenzhen, Guangdong, ²Shenzhen Children's Hospital, Shenzhen, Guangdong

Xiaodong Zhang  
Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences|Shenzhen Children's Hospital
Shenzhen, Guangdong|Shenzhen, Guangdong

Dezhi Cao  
Shenzhen Children's Hospital
Shenzhen, Guangdong

Jinping Xu  
Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences
Shenzhen, Guangdong

Epilepsy is one of the most prevalent neurological disorders. About 50 million individuals worldwide suffer from epilepsy [1]. Focal cortical dysplasia (FCD), characterized by changes in cortical thickness, blurring of gray-white matter junctions, abnormalities in gyrus structure, is one primary cause of drug-resistant epilepsy (DRE) [2][3]. Resection of the epileptic focus (EF) through neurosurgery is the most effective treatment of DRE-FCD. However, it is still challenging to localize EF based on neuroimaging due to the subtle structural changes of brain cortex. Computer-Aided Diagnosis (CAD) has been applied to localize EF by analyzing magnatic resonance (MR) images of FCD patients. However, conventional methods usually employ low-level artificial features, which is time-consuming and lack of feature representation capacity. The artificial intelligence represented by deep learning has promoted the development of intelligent CAD technology [4][5]. However, the intelligent diagnosis of epilepsy is still under-explored. This article will investigate the use of deep models to segment EF from multi-channel MR images of FCD patients.

A deep learning based model is proposed in this paper to segment FCD EF. The proposed model receives a sub-volume sampled from multi-channel MR images and generates a EF probability map. The model architecture consists of a CNN encoder, a ViT encoder [6] and a CNN decoder. CNN encoder is a 4-level convolutional network responsible for extracting local detail features. ViT encoder is employed to extract the global semantic features. CNN decoder is composed of 4 Fusion blocks which are capable of fusing features from CNN encoder, ViT encoder and previous Fusion block. A ShapeMatch block is employed to reshape and upsample the ViT feature to match with CNN feature before fusion. Upsample block is inserted between two consecutive Fusion blocks to double the feature size. At last, a convolution operation with kernel size of 1 and a softmax operation compose a Output block to convert features into EF probability map.

We utilized a public open data set of FCD type II, provided by the Department of Epileptology at the University Hospital Bonn and approved by the ethics committee of the University of Bonn [8]. This data set includes information from 85 patients, each offering T1 and FLAIR sequences. Ground truth (GT) of EF was delineated on FLAIR by two experts of epilepsy imaging. We used FSL to perform intro-registration between T1 and FLAIR (along with GT) as well as inter-registration with MNI-152 brain altlas. The aligned T1 and FLAIR, as well as their x-axis flipped images, are concatenated into 4-channel MR image. The dataset are divided into training and testing set with a ratio of 8:2. Model was trained with a hybrid loss function composed of dice loss and binary cross-entropy loss using Python 3.8 and PyTorch 1.10. For fair performance comparison, we re-trained 3DResUNet [9], 3DAttentionUNet [10] and UNETR[7] on the same dataset. The evaluation metrics include subject-level sensitivity (Sens-sub), voxel-level sensitivity (Sens-vox) and Dice coefficient (DC). Evaluations were conducted on the testing set. The proposed model achieved the best performance with Sens-sub of 88.2%, Sens-vox of 0.564±0.313, DC of 0.416±0.257, better than 3DResUNet (64.7%, 0.380±0.316, 0.369±0.300), 3DAttentionUNet (82.4%, 0.370±0.272, 0.349±0.247), UNETR (52.9%, 0.279±0.296, 0.275±0.280).

This paper introduces a deep neural network to segment EF from MR images of FCD patients, which shows superior performance compared to the three other models. However, there is a performance discrepancy between the segmentation of EF and other structures. Some cases remain undetectable. These underscore the considerable challenges associated with EF segmentation. Future work will explore pre-training the encoders of our proposed model based on self-supervision models, expecting to further enhance model performance.

Neurodevelopmental/ Early Life (eg. ADHD, autism) ²

Methods Development

Segmentation and Parcellation ¹

Computational Neuroscience

Epilepsy

Machine Learning

MRI

Segmentation

STRUCTURAL MRI