Reducing noisy annotations for depression estimation from facial images.

Area of Science:

• Computer Science
• Artificial Intelligence
• Psychiatry

Background:

• Depression is a prevalent mental disorder requiring accurate severity assessment.
• Current deep learning models for depression estimation often struggle with noisy labels and fail to capture complex facial-BDI-II score relationships.

Purpose of the Study:

• To propose a novel automated deep architecture, the Self-Adaptation Network (SAN), for improved depression severity estimation.
• To address the challenge of uncertain labeling in automatic depression estimation (ADE) by incorporating label refinement.
• To enhance the learning of relationships between facial features and depression severity scores.

Main Methods:

• The proposed Self-Adaptation Network (SAN) integrates ResNet-18/50 for deep feature extraction, a self-attention module (SAM) for weight learning, a square ranking regularization module (SRRM) for data partitioning, and a re-label module (RM) for uncertain annotation refinement.
• Utilized AVEC2013 and AVEC2014 depression databases for extensive experimentation.
• Employed deep learning architectures for regression tasks, focusing on improving label quality.

Main Results:

• The SAN achieved performance comparable to existing state-of-the-art ADE methods on benchmark depression datasets.
• The method demonstrated an ability to learn valuable depression patterns directly from facial videos.
• The re-label module effectively mitigated issues arising from noisy or uncertain annotations.

Conclusions:

• The Self-Adaptation Network (SAN) offers a promising approach for more accurate and robust automated depression estimation.
• The SAN's ability to refine uncertain labels and learn intricate facial-depression relationships represents a significant advancement in the field.
• This work contributes to the development of reliable AI tools for mental health assessment.