TLDR: This research introduces a transparent deep learning framework combining an Autoencoder (AE) and Vision Transformer (ViT) to improve and explain automated dental age estimation, specifically for second (tooth 37) and third (tooth 38) molars. The framework significantly boosts classification accuracy for both teeth. Crucially, it reveals that the lower performance for tooth 38 isn’t a model failure but a data-centric issue caused by high morphological variability in the tooth 38 dataset. This multi-faceted approach, using latent space analysis, image reconstructions, and attention maps, offers robust diagnostic insights beyond traditional “black box” explanations, making AI more trustworthy in high-stakes forensic applications.
Deep learning models are increasingly used in critical applications like forensic science, but their ‘black box’ nature often limits their adoption. A new study addresses this challenge in dental age estimation, a crucial process for legal proceedings involving juveniles and young adults. The research focuses on the automated staging of mandibular second (tooth 37) and third (tooth 38) molars, where a notable difference in model performance has been observed.
The paper, titled “An Autoencoder and Vision Transformer-based Interpretability Analysis of the Differences in Automated Staging of Second and Third Molars,” introduces a novel framework designed to enhance both the performance and transparency of deep learning models in this context. Authored by Barkin Buyukcakir, Jannick De Tobel, Patrick Thevissen, Dirk Vandermeulen, and Peter Claes, the study combines a convolutional autoencoder (AE) with a Vision Transformer (ViT) to achieve its goals.
The proposed AE + ViT framework significantly improves classification accuracy for both teeth. For tooth 37, accuracy increased from 0.712 to 0.815, and for tooth 38, it rose from 0.462 to 0.543, compared to a baseline ViT model. Beyond these performance gains, the framework offers multi-faceted diagnostic insights, moving beyond simple attention maps that can sometimes be misleading.
One of the key findings is that the remaining performance gap, particularly for tooth 38, is primarily data-centric. Analysis of the AE’s latent space metrics and image reconstructions suggests that high intra-class morphological variability within the tooth 38 dataset is a major limiting factor. This means that the third molars, due to their inherent anatomical variations, are harder for the model to consistently categorize across different individuals.
The framework’s interpretability features are crucial. The autoencoder acts as a preprocessing step, reducing image ‘noise’ and generating smoothed, prototypical reconstructions for each developmental stage. This simplified representation helps the Vision Transformer’s self-attention mechanism focus more effectively on diagnostically relevant anatomical structures. For tooth 37, the AE preprocessing resulted in clear prototypes with reduced noise, making dental structures more distinct. In contrast, tooth 38 reconstructions showed considerable blurring, especially in the root and crown regions, indicating difficulty in forming distinct stage prototypes due to high variability.
Furthermore, the study examined attention maps, which highlight areas of an image that are most influential in a model’s decision. While baseline ViT models sometimes focused on less relevant areas or failed to shift attention to critical root regions in later stages, the AE + ViT model displayed a more distributed and anatomically aligned attention pattern. For tooth 37, it incorporated root information more effectively, aligning with human diagnostic criteria. For tooth 38, even with AE preprocessing, the model still struggled with the root area, reinforcing the conclusion about its high morphological variability.
A quantitative analysis of the latent space, where images are encoded into a lower-dimensional representation, further supported these findings. For tooth 37, the latent space showed clear separation between stages and tight clustering within stages. However, for tooth 38, there was poor separation between several classes and higher variability within stages, indicating a less structured and harder-to-classify dataset.
Also Read:
- Geometry-Guided AI Enhances Multi-View Mammography Analysis
- Unveiling Data’s True Complexity: Introducing the Intrinsic Dimension Estimating Autoencoder (IDEA)
In conclusion, this research demonstrates that the performance disparity in automated dental staging is not an architectural flaw of deep learning models but rather a data-centric problem rooted in the intrinsic morphological variability of third molars. The proposed AE + ViT framework serves as a robust tool for diagnostic transparency, providing forensic odontologists with a ‘second opinion’ that includes reconstructions, latent space positions, and attention maps. This information can help experts understand why a model might be uncertain, adding a critical layer of context to their final assessments. For more details, you can read the full paper here.
