Abstracts

In patients with temporal lobe epilepsy (TLE), hippocampal or Ammon[apos]s horn sclerosis (AHS) is the major neuropathological substrate. Magnetic resonance imaging (MRI) is currently regarded as the gold standard for the in vivo diagnosis of AHS. The visual inspection of MRI images for diagnostic purposes, however, requires expert knowledge, special sequences as well as a proper angulation of slices. Radiologists outside epilepsy centers using standard MRI protocols often fail to detect AHS. In order to avoid these shortcoming a number of computer-aided diagnosis methods have been proposed that aim at an automated classification of MRI images. We here introduce an application that allows the detection of AHS using a special class of neural network, called Cellular Neural Networks (CNN)., 3D MRI data from 144 subjects were collected with a T1-weighted MRI protocol on a 1.5-T scanner (Siemens) using the MPRAGE sequence. MRI images of 46 subjects were diagnosed as without pathological findings. In 98 subjects AHS was confirmed by a board-certified radiologist. Within this group 45 patients had a left-sided AHS, 40 had a rightsided AHS, and 13 patients had bilateral AHS. For network optimization we compiled a training set that consisted of 20 left-hemispheric non-adjacent coronal slices from 10 subjects with AHS and 20 left-hemispheric nonadjacent coronal slices from 10 subjects without AHS. All slices were normalized prior to analysis., An in-sample optimization of the CNN led to a correct classification of 82.5% of slices. To reduce the risk of an overestimation of our methods performance due to this in-sample optimization, we performed an out-of-sample validation using 1178 slices and obtained a correct classification in 68% of slices presented to the network (p [lt] 0.001)., We have presented a method for an automated classification of MR images with AHS. Our method differs from previous analytical or statistical approaches for brain tissue classification since it does not require a parametrization of the image data. At present, the achieved performance can not be regarded as sufficient to allow broader clinical application. Nevertheless, it is conceivable that an increased classification performance can be achieved using specifically defined training sets that also take into account demographic as well as disease-related factors. With further improvements and due to the high computational performance of CNNs one might envisage an implementation of our approach into medical imaging systems for a pre-screening of images.,