Purpose: To evaluate the performance of discrete compactness and tortuosity in the hippocampi (right and left) as possible magnetic resonance neuroimaging biomarkers to discriminate between populations of controls and subjects with mild cognitive impairment. Methods: 98 subjects were analyzed, including 49 healthy and 49 with mild cognitive impairment. Magnetic resonance images and other patient data were obtained from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. The structure of the hippocampi (right and left) was segmented using image processing techniques and volume, normalized volume, discrete compactness and tortuosity metrics were calculated. Statistical analysis of the data was performed in order to differentiate the classes. The metrics were incorporated into an automatic random forest classifier to discriminate between study populations. Results: The statistical analysis showed statistically significant differences for the calculated metrics with a p-value < 0.01 between the two study classes. The automatic random forest classification algorithm, based on the indicators of volume, normalized volume, discrete compactness and tortuosity, achieved an accuracy in the training stage of 89.83 % ± 0.052 %. For the test stage with reserved data, the final accuracy was 85 %. Conclusions: Discrete compactness and tortuosity are sensitive to morphological changes in the right and left hippocampus, and characterize the stage of mild cognitive impairment; therefore, they can be considered imaging biomarkers, useful in the detection of mild cognitive impairment when the first symptoms of Alzheimer's disease begin to appear. Keywords: discrete compactness; tortuosity; mild cognitive impairment; random forest algorithm.

Alzheimer's disease is a multifactorial neurodegenerative disorder preceded by a prodromal stage called mild cognitive impairment (MCI). Early diagnosis of MCI is crucial for delaying the progression and optimizing the treatment. In this study we propose a random forest (RF) classifier to distinguish between MCI and healthy control subjects (HC), identifying the most relevant features computed from structural T1-weighted and diffusion-weighted magnetic resonance images (sMRI and DWI), combined with neuro-psychological scores. To train the RF we used a set of 60 subjects (HC=30, MCI=30) drawn from the ADNI database, while testing with unseen data was carried out on a 23-subjects Mexican cohort (HC=12, MCI=11). Features from hippocampus, thalamus and amygdala, for left and right hemispheres were fed to the RF, with the most relevant being previously selected by applying extra trees classifier and the mean decrease in impurity index. All the analyzed brain structures presented changes in sMRI and DWI features for MCI, but those computed from sMRI contribute the most to distinguish from HC. However, sMRI+DWI improves classification performance in training (AUROC=93.5+/-8%, accuracy=88.8+/-9%) and testing with unseen data (AUROC=93.79%, accuracy=91.3%), having a better performance when neuro-psychological scores were included. Compared to other classifiers the proposed RF provide the best performance for HC/MCI discrimination and the application of a feature selection step improves its performance. These findings imply that multimodal analysis gives better results than unimodal analysis and hence may be a useful tool to assist in early MCI diagnosis.

This work presents an evaluation of two machine learning schemes for lid segmentation on meibography images. Both schemes use same input features based on pixel gray levels, laplacian and entropy filters, and also distances to anatomical landmarks like pupil and eye lashes. The methods evaluated were support vector machines (SVM) with 4th degree polynomial kernel, a Neural Network (NN) with 60 neurons distributed on three layers and the intersection of both. Performance was evaluated with AUC on a bootstrapped cross validation (CV) tests of 20 folds. Dataset is conformed by 465 images for each fold entire dataset was split on 70% training and remainder for testing. Results of CV: SVM 0.851 ± 0.103, NN 0.713 ± 0.144 and SVM & NN 0.835 ± 0.118, suggest that SVM is a suitable model to be used for this task.