Abstracts

Rationale: The visual identification of subtle abnormalities in MR brain images that represent the origin of focal forms of epilepsy is a challenging and important problem in neuroimaging as localisation of such abnormalities can help guide curative neurosurgical procedures. Unfortunately, much of the time no abnormality can be identified visually in these images - the so called MR negative cases. In this study, we used machine learning techniques to uncover patterns of abnormality in multi-modal brain imaging from individuals with temporal lobe epilepsy (TLE) in (i) visually abnormal (MR positive) and (ii) visually normal (MR negative) cases.  Methods: The task was to correctly lateralise seizure onset to the right or left temporal lobe. The dataset comprised 108 subjects (82 MR positive, 26 MR negative). All the subjects had a known lateralisation of seizure origin (from clinical assessment, EEG, and sometimes imaging studies) and had image volumes from three different MR modalities available (T1, T2, FLAIR) along with Junction Maps (Huppertz HJ, et al. Epilepsy Research, 67 (2005) 35–50). In each subject group a random forest classifier was trained on image features and then feature importance measurements were carried out within the trained forest to quantify the significance of every image feature considered from the temporal lobes. We describe and present a visualisation approach for these feature importances that we call ‘Importance Maps’. The global predictive capacity of the selected image features was then assessed by measuring the lateralisation accuracy of a support vector machine classifier.  Results: Accuracies of 93.8% and 81.7% were obtained in the MR positive and MR negative groups respectively. Our results demonstrate that useful abnormalities exist in MR negative images reported to be normal by human readers, and that these abnormalities are found in a different spatial pattern than in individuals with visually apparent abnormalities. A particular contrasting significance of features arising from the hippocampus and amygdala in the MR positive and MR negative subjects respectively is demonstrated. The hippocampal volume and T2/FLAIR signal features typically looked for by radiologists in this context are found to be present in the MR positive subjects and absent in the MR negative subjects. By contrast the T1 and FLAIR signal variances within the amygdala prove particularly significant in the MR negative subjects. Conclusions: Our results demonstrate that abnormalities exist in MR images reported to be normal by a human reader, and that these abnormalities exist in a different spatial pattern to that seen in visually apparent cases. We gain some insight into why visual assessment may be uninformative in certain cases and provide suggestions on how to improve this situation. Funding: This work is supported by the EPSRC-funded UCL Centre for Doctoral Training in Medical Imaging (EP/L016478/1) and the NIHR-funded Biomedical Research Centre at University College London Hospitals.We are grateful to the Wolfson Foundation and the Epilepsy Society for supporting the Epilepsy Society MRI scanner. This work is also supported by the EPSRC, the MRC and the EU-FP7 project VPH-DARE@IT. GPW was supported by an MRC Clinician Scientist Fellowship (MR/M00841X/1).