💻 computer science

ProtoPointNet: Prototype-Based Interpretable Classification of 3D Dental Point Clouds with Verifiable Spatial Activations

The paper introduces ProtoPointNet, an interpretable prototype-based model that classifies 3D dental occlusion by leveraging multi-task learning on registered intraoral arch pairs and achieving high performance on the Bits2Bites dataset while providing spatially grounded, anatomically plausible visual explanations for its predictions.

Original authors: George V. Jose, Thao Liang Chiam, Toby Hughes, Dilan Patel, Alan Brook, Lyle J. Palmer, Nikhil Cherian Kurian

Published 2026-07-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: George V. Jose, Thao Liang Chiam, Toby Hughes, Dilan Patel, Alan Brook, Lyle J. Palmer, Nikhil Cherian Kurian

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you're trying to figure out how a person's top and bottom teeth fit together. Usually, a computer might look at a 3D scan of those teeth and just spit out a label like "Good Bite" or "Bad Bite," but it keeps its reasoning a secret. It's like a magic 8-ball that says "Yes" without telling you why.

The researchers behind ProtoPointNet decided to build a different kind of computer brain—one that acts more like a detective with a magnifying glass. Instead of guessing, this system looks for specific "clues" (called prototypes) that it has learned from past cases. It's like teaching a student to identify a "cross-bite" not by memorizing a whole picture, but by showing them a specific patch of teeth that always looks like a cross-bite. When the computer sees that same patch again, it can point right at it and say, "Aha! I know this! This is why I think it's a cross-bite."

The Secret Sauce: A 14-Dimensional Super-Description
To make this work, the team didn't just feed the computer the shape of the teeth. They gave every single point on the 3D scan a special "ID card" with 14 different pieces of information. Think of it like describing a person not just by their height, but by how far they are standing from their neighbor, the angle of their nose, and the curve of their cheek.

This ID card includes the usual shape details, but the magic part is the "relational" info: it tells the computer exactly how far a point on the top teeth is from the closest point on the bottom teeth, and in which direction. This lets the computer see the gap or the overlap between the two arches, which is the whole point of checking a bite.

The "This Looks Like That" Game
The system works by comparing the new scan to a library of "exemplars" (prototypes) it learned during training.

If the computer sees a patch of teeth that looks very similar to a "Deep Bite" prototype it learned earlier, it lights up that clue.
It then adds up all the matching clues to make a final decision.
Crucially, because it knows exactly which patch matched which clue, it can project that match back onto the 3D model. You can literally see a glowing highlight on the computer screen showing exactly which teeth (like the back molars or front incisors) convinced the computer to make its choice.

The Challenge: Learning from a Tiny Library
Here's the tricky part: The team only had about 140 patient scans to train on. That's like trying to learn a new language with only a few dozen flashcards. Most AI models would get confused or just memorize the cards (overfitting).

To solve this, the researchers used a clever training trick. They first let the computer learn general shapes using a helper system, and then they "froze" that part so it wouldn't forget. Then, they focused the computer's energy only on learning the specific clues (prototypes) for the bite problems. This allowed the system to learn effectively even with such a small dataset.

The Results: Good at Some Things, Not Others
When they tested this system on 30 new patients (a total of 200 scans in the dataset), here is what happened:

The Big Wins: The system was very good at spotting vertical problems (like deep or open bites) and left-side alignment issues. It got a score of 0.828 for vertical bites and 0.807 for left-side alignment (on a scale where higher is better).
The Mixed Bag: It was decent at right-side alignment and side-to-side (transverse) issues, but it struggled with the "midline" (whether the center of the top teeth lines up with the bottom). For the midline, it scored only 0.457, which is basically guessing.
The Comparison: When they compared this "detective" style to a standard "black box" AI, the detective style was just as good at the hard tasks (like left-side alignment) but had the huge advantage of showing its work.

What They Ruled Out
The paper makes it clear that simply looking at the shape of one set of teeth isn't enough. You must look at how the top and bottom teeth relate to each other. Also, they found that while standard "attention" maps (which just highlight a blurry area) can show where the computer looked, they don't explain why like this prototype system does. The prototype system is the only one that can say, "I matched this specific patch to a known bad-bite example."

How Sure Are They?
The authors are confident that this method works for explaining why a decision was made, and the numbers suggest it's a solid alternative to black-box models for dental scans. However, they are careful to note that because the test group was small (only 30 patients), the scores are just estimates. They suggest the midline results might be weak because the scans sometimes picked up too much of the gum or roof of the mouth, confusing the system. They don't claim this is a perfect, solved problem for every dental case yet, but they do show that this "transparent" way of thinking is a viable and promising path forward for analyzing 3D dental surfaces.

Technical Summary: ProtoPointNet

Problem Statement
Deep learning models for 3D point clouds, such as PointNet and Point Transformer, have achieved strong performance in shape classification and segmentation. However, in clinical settings like dentistry, these models often function as "black boxes," failing to expose the specific anatomical regions or spatial relationships driving their predictions. While prototype-based networks (e.g., ProtoPNet) offer inherent interpretability by linking predictions to learned exemplars ("this looks like that"), their application has been limited to single-object classification or 2D images. There is a lack of methods capable of handling relational 3D tasks, where the classification label depends on the spatial interaction between two surfaces (e.g., the upper and lower dental arches) rather than a single object. Specifically, dental occlusion classification requires analyzing the contact, gap, and alignment between paired intraoral arches, a task not well-addressed by existing 3D classifiers or 2D projection-based methods.

Methodology
The authors propose ProtoPointNet, a prototype-based architecture designed for inherently interpretable classification of 3D dental occlusion from registered upper–lower intraoral arch pairs.

Relational Point Representation: Unlike standard point cloud descriptors, ProtoPointNet encodes each point with a 14-dimensional descriptor. This includes local surface geometry (coordinates, normals, shape index, Gaussian curvature) and explicit relational channels: the 3D displacement vector ( $\Delta_n$ ) and scalar clearance ( $\delta_n$ ) to the nearest point on the opposing arch. This explicitly exposes inter-arch relationships (contact, separation, misalignment) to the model.
Architecture: The model utilizes a shared PointNet++-style hierarchical encoder to process the concatenated point cloud of both arches. An adapter module re-injects positional coordinates and normalizes the output into a prototype space.
Multi-Task Prototype Heads: The network employs a shared backbone with five task-specific heads for sagittal-left, sagittal-right, vertical, transverse, and midline classification. Each head contains a set of class-specific prototypes.
Prototype Matching: Predictions are generated by computing the similarity between local point embeddings and learned prototypes. Activations are calculated as the mean of the top- $k$ similarities. Crucially, the classifier weights are parameterized to ensure prototypes contribute non-negative evidence to their assigned class and non-positive evidence to others, allowing for decomposable logits.
Training Strategy for Small Data: Given the limited dataset (approx. 140 training pairs), the authors employ a small-data training recipe:
- Auxiliary Supervision: An auxiliary classifier (using max-pooling over embeddings) is trained alongside the prototype head to stabilize learning.
- Encoder-Freeze Hand-off: The auxiliary weight is annealed to zero while the encoder is frozen, allowing the adapter, prototypes, and classifiers to adapt solely to the prototype path.
- Regularization: The loss function includes cluster, separation, diversity, and coverage terms to ensure prototypes are distinct, cover the data manifold, and do not collapse.
Interpretability: During inference, the model projects the best-matching input regions of activated prototypes back onto the original dental meshes. This provides a forward-pass explanation (identifying specific anatomical patches supporting a decision) rather than a post-hoc saliency map.

Key Contributions

Interpretable Paired-Surface Reasoning: Introduction of ProtoPointNet, the first prototype-based point-cloud model capable of mapping prototype evidence directly back to the original meshes for dental occlusion classification.
Relational Inter-Arch Descriptors: Development of a 14-dimensional descriptor that combines local geometry with explicit displacement and clearance to the opposing arch, enabling prototypes to capture contact and alignment patterns.
Small-Data Multi-Task Validation: Demonstration of a shared encoder with axis-specific prototype heads trained via auxiliary supervision and encoder-freeze hand-off, validating that prototype reasoning can be learned from limited clinical 3D data while retaining spatially meaningful explanations.

Results
The model was evaluated on the Bits2Bites benchmark, which contains 200 registered upper–lower STL pairs with five occlusion labels.

Performance: ProtoPointNet achieved a mean test macro-F1 of 0.724 and an AUROC of 0.825.
Task-Specific Performance: The model showed strongest performance on vertical classification (F1 0.828, AUROC 0.970) and sagittal-left classification (F1 0.807, AUROC 0.900). Performance was moderate for sagittal-right and transverse tasks, while the midline task remained weak (F1 0.457), likely due to the subtlety of the deviations and sampling noise.
Ablation Studies:
- Descriptors: The full 14-dimensional relational descriptor outperformed raw geometry (6-d) and intrinsic geometry (9-d) descriptors, with the largest gains observed in sagittal tasks.
- Classifier Heads: While direct max-pooling and attention heads achieved slightly higher mean macro-F1 or AUROC in some metrics, the prototype head provided the best performance on sagittal tasks (where inter-arch reasoning is critical) and uniquely offered readable, case-based decompositions.
Spatial Coherence: Visualizations confirmed that prototype activations localized to anatomically plausible regions (e.g., posterior molars for cross-bite evidence, anterior incisors for bite-depth evidence).

Significance and Claims
The paper claims that ProtoPointNet establishes prototype-based reasoning as a viable, transparent alternative to black-box 3D classifiers for dental surface-pair analysis. By linking predictions to learned exemplars and projecting them onto specific anatomical regions, the model supports "clinically inspectable decisions." The authors emphasize that while the model does not strictly reproduce all clinical criteria (e.g., it struggles with midline deviations), it successfully exploits discriminative features to predict labels while providing spatially grounded, forward-pass explanations. The work demonstrates that inherently interpretable models can be effectively trained on small, limited clinical datasets without sacrificing the ability to explain where and why a classification was made.

Technical Summary: ProtoPointNet

More like this