Hakim Achterberg MSc, Presenter: Nothing to Disclose

Marleen de Bruijne PhD, Abstract Co-Author: Research grant, AstraZeneca PLC

Fedde Van Der Lijn PhD, Abstract Co-Author: Nothing to Disclose

Tom den Heijer PHD, Abstract Co-Author: Nothing to Disclose

Meike Willemijn Vernooij MD, Abstract Co-Author: Nothing to Disclose

Mohammad Arfan Ikram, Abstract Co-Author: Nothing to Disclose

Wiro Niessen PhD, Abstract Co-Author: Nothing to Disclose

To develop and evaluate a method to predict incidence of dementia based on hippocampal shape descriptors derived from brain MRI scans.

We used MRI scans from 511 participants in the Rotterdam Scan Study (mean age 73.49 (7.90 SD), 256 male) who were non-demented at the time of scan. Participants were followed intensively for incident dementia using cognitive screening and medical records. Over a 10-year follow-up period, 52 subjects developed dementia with a median time between scan and diagnosis of 4.0 (4.9 IQR) years. The left and right hippocampus were segmented using an automatic method followed by visual inspection and manual correction in case of gross errors. In total, 1024 corresponding landmark points were automatically placed on each hippocampal surface. The resulting shapes were scaled to identical volume so that only shape information, not volume, was contained in the model. A support vector machine (SVM) classifier was trained to discriminate between shapes of subjects who stayed cognitively intact during the follow-up and of those who developed dementia. A subset of 50 subjects who developed dementia and 150 age- and gender matched controls was used for training. Evaluation was performed in a leave-one-out cross-validation manner on all subjects such that there was no overlap between train and test data used for each of the classifiers. To evaluate the added value of shape information, performance was compared to SVMs based on volume alone and to SVMs combining shape and volume information. Two logistic regression models were created: one with hippocampal volume and a second with both volume and shape as covariates. Both models were corrected for age and gender. 

The area under the ROC curve (AUC) was 0.75 when using only shape information, against 0.70 when using only volume information. When combining volume and shape the AUC increased slightly to 0.76. In the regression model with only volume, the volume term was significant (P < 0.0001). When adding shape, volume was no longer significant (P = 0.1109) and shape was significant (P < 0.0001).

Hippocampal shape can predict dementia incidence in the general population with a median of 4 years before clinical diagnosis. 

We show that -in addition to volume- hippocampal shape can improve predictive accuracy. Hippocampal shape may be included in future predictive models for risk assessment of dementia.

Achterberg, H, de Bruijne, M, Van Der Lijn, F, den Heijer, T, Vernooij, M, Ikram, M, Niessen, W, Hippocampal Shape Predicts Development of Dementia in the General Population.  Radiological Society of North America 2011 Scientific Assembly and Annual Meeting, November 26 - December 2, 2011 ,Chicago IL. http://archive.rsna.org/2011/11004703.html