Computer Aided Diagnosis System for Detection of Focal Cortical Dysplasia Lesions on T1 and T2 Weighted MRI

Abstract

Focal cortical dysplasia (FCD) is the most frequent malformation for patients with pharmacoresistant epilepsy that require surgical treatment. Providing automated procedures to detect FCD lesions is greatly desirable because visual diagnosis is often challenging, time consuming, and relies highly on the individual's expertise. In this thesis, we propose two Computer Aided Diagnosis (CAD) approaches for Focal Cortical Dysplasia (FCD) lesion detection and segmentation on T1 and T2 weighted MRI. For the rst CAD system, an automatic detection algorithm for FCD lesions on T1 weighted MRI is proposed. Instead of using the traditional voxel-based analysis, we introduce a set of volume-based statistical features with Naive Bayes Classier. Subsequently, a set of cluster-based differential features with a Support Vector Machine (SVM) classier is used to eliminate the false positives (FPs) resulting from the rst processing stage. The advantage of our system lies on the use of volume-based analysis to allow the study of feature distributions in a spatial neighborhood. The second CAD system automatically segments FCD lesions on T2 weighted MRI. We present a Markov Random Field (MRF) model for the segmentation task with a particular emphasis on the incorporation of T1 information with a location prior. By integrating such location prior, we take the advantage of T1 weighted MRI in producing better differentiation of soft tissues into the T2 lesion segmentation task. The proposed algorithms are validated on a dataset that consists a total of 51 subjects with FCD lesions provided by the Radiology Department of Mayo Clinic. The experimental results show a 87% FCD lesion detection rate for T1-weighted MRI and a 100% FCD lesion detection rate for T2 weight MRI. The experimental results also show that proposed methods outperform previous methods in the literature .