6533b86ffe1ef96bd12cda22

RESEARCH PRODUCT

Détection automatique des repères visuels associés à la dépression

subject

description

Depression is the most prevalent mood disorder worldwide having a significant impact on well-being and functionality, and important personal, family and societal effects. The early and accurate detection of signs related to depression could have many benefits for both clinicians and affected individuals. The present work aimed at developing and clinically testing a methodology able to detect visual signs of depression and support clinician decisions.Several analysis pipelines were implemented, focusing on motion representation algorithms, including Local Curvelet Binary Patterns-Three Orthogonal Planes (LCBP-TOP), Local Curvelet Binary Patterns- Pairwise Orthogonal Planes (LCBP-POP), Landmark Motion History Images (LMHI), and Gabor Motion History Image (GMHI). These motion representation methods were combined with different appearance-based feature extraction algorithms, namely Local Binary Patterns (LBP), Histogram of Oriented Gradients (HOG), Local Phase Quantization (LPQ), as well as Visual Graphic Geometry (VGG) features based on transfer learning from deep learning networks. The proposed methods were tested on two benchmark datasets, the AVEC and the Distress Analysis Interview Corpus - Wizard of Oz (DAICWOZ), which were recorded from non-diagnosed individuals and annotated based on self-report depression assessment instruments. A novel dataset was also developed to include patients with a clinical diagnosis of depression (n=20) as well as healthy volunteers (n=45).Two different types of depression assessment were tested on the available datasets, categorical (classification) and continuous (regression). The MHI with VGG for the AVEC’14 benchmark dataset outperformed the state-of-the-art with 87.4% F1-Score for binary categorical assessment. For continuous assessment of self-reported depression symptoms, MHI combined with HOG and VGG performed at state-of-the-art levels on both the AVEC’14 dataset and our dataset, with Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE) of 10.59/7.46 and 10.15/8.48, respectively. The best performance of the proposed methodology was achieved in predicting self-reported anxiety symptoms in our dataset, with RMSE/MAE of 9.94/7.88.Results are discussed in relation to clinical and technical limitations and potential improvements in future work.