An autoencoder and vision transformer based interpretability analysis on the performance differences in automated...

Area of Science:

• Forensic Science
• Artificial Intelligence
• Biomedical Imaging

Background:

• Deep learning models in forensic applications like dental age estimation are often limited by their 'black box' nature, hindering practical adoption.
• Performance disparities exist in automated dental age estimation, particularly between different tooth types.

Purpose of the Study:

• To introduce a novel framework enhancing both performance and transparency in deep learning for forensic dental age estimation.
• To address the 'black box' problem and provide multi-faceted diagnostic insights for improved decision-making.

Main Methods:

• A framework combining a convolutional autoencoder (AE) with a Vision Transformer (ViT) was developed.
• The framework was evaluated using a case study on the performance disparity in staging mandibular second (tooth 37) and third (tooth 38) molars.
• Latent space metrics and image reconstructions from the AE were analyzed to understand model uncertainty.

Main Results:

• The proposed AE-ViT framework improved classification accuracy over a baseline ViT: from 0.712 to 0.815 for tooth 37 and from 0.462 to 0.543 for tooth 38.
• Analysis indicated that the remaining performance gap is data-centric, with high intra-class morphological variability in the tooth 38 dataset being a key limitation.
• The study demonstrated that relying solely on interpretability methods like attention maps is insufficient for identifying underlying data issues.

Conclusions:

• The developed framework offers a more robust tool for forensic age estimation by enhancing accuracy and providing evidence for model uncertainty.
• This approach supports expert decision-making by offering transparency and identifying data-centric limitations.
• Multi-faceted interpretability is crucial, moving beyond single methods like attention maps to uncover data-related challenges.