GeoSapiens: Few-Shot Dental Landmark Detection

• GeoSapiens is a few-shot learning framework for dental landmark detection on anterior teeth CBCT slices, combining a human-centric foundation model with a novel geometric loss.
• It integrates Low-Rank Adapters into a Vision Transformer backbone, reducing trainable parameters from 330M to 24M while preserving robust localization performance.
• Empirical results show an 8.18% improvement in [email protected] mm and a 9.5% reduction in mean radial error, demonstrating significant gains over state-of-the-art methods.

Searching arXiv for the specified paper and closely related references to ground the article in current research. GeoSapiens is a few-shot learning framework for dental landmark detection on cone-beam computed tomography (CBCT) of anterior teeth. It was introduced in "Geometric-Guided Few-Shot Dental Landmark Detection with Human-Centric Foundation Model" (Wang et al., 7 Jul 2025) to address a setting in which manual landmark annotation by dentists is time-consuming, labor-intensive, and subject to inter-observer variability, while conventional deep learning is constrained by scarce training data and the high cost of expert annotations. The framework combines a baseline adapted from Sapiens, a human-centric foundation model, with a geometric loss designed to encode anatomical structure, and it is reported to surpass existing landmark detection methods on the authors' collected dataset of anterior teeth landmarks, including an 8.18% higher success detection rate at a strict 0.5 mm threshold than the leading approach (Wang et al., 7 Jul 2025).

1. Clinical task and problem formulation

Accurate detection of anatomic landmarks is described as essential for assessing alveolar bone and root conditions, thereby optimizing clinical outcomes in orthodontics, periodontics, and implant dentistry (Wang et al., 7 Jul 2025). Within this setting, GeoSapiens targets anterior-teeth CBCT slices and predicts 16 distinct anatomical landmarks per slice on upper and lower anterior teeth. The dataset used in the study, LDTeeth, contains 347 images of size $957 \times 555$ px and 5,552 total landmark points, with 16 landmarks for each image (Wang et al., 7 Jul 2025).

The paper frames the task as few-shot adaptation rather than episodic meta-learning. The training set consists of only 36 CBCT slices drawn from 3 patients, with validation on 149 images from 14 patients and testing on 162 images from 16 patients (Wang et al., 7 Jul 2025). This design emphasizes fine-tuning under severe annotation scarcity. A plausible implication is that the method is intended less as a generic meta-learning algorithm than as a parameter-efficient transfer pipeline for small clinical datasets.

The operating unit is a 2D single mid-sagittal slice. Each slice is first cropped to a dentist-annotated tooth bounding box corresponding to the full mid-sagittal cut, after which the model predicts landmark heat maps and coordinates (Wang et al., 7 Jul 2025). The paper explicitly notes that out-of-plane lesions are not addressed, which bounds the current formulation to 2D localization rather than full volumetric landmarking (Wang et al., 7 Jul 2025).

2. Architectural design

GeoSapiens uses Sapiens-0.3B as its backbone, a Vision Transformer foundation model pre-trained via Masked Autoencoding on 300 M in-the-wild human images (Wang et al., 7 Jul 2025). The motivation given in the paper is a transfer hypothesis: anterior teeth CBCT slices share symmetric, morphological patterns akin to human body parts, so Sapiens representations may transfer effectively (Wang et al., 7 Jul 2025). This is presented as a hypothesis rather than a formally proven property.

Instead of full fine-tuning, the framework inserts Low-Rank Adapters (LoRA) into all multi-head self-attention projections and multilayer perceptrons. The attention $Q/K/V$ projections use rank $r=4$ with $\alpha=4$, and the feed-forward projection layers use $r=8$ with $\alpha=8$ (Wang et al., 7 Jul 2025). Full fine-tuning would involve approximately 330 M trainable weights, whereas the LoRA configuration reduces trainable parameters to approximately 24 M, reported as approximately 0.3% of the original in the summary, while the implementation details also describe the reduction as $330\text{ M} \to 24\text{ M}$ (Wang et al., 7 Jul 2025).

A top-down heat-map head, described as following SimpleBaseline, transforms ViT features into $K=16$ per-landmark likelihood maps $H_k(x,y)$ (Wang et al., 7 Jul 2025). During training, the model uses differentiable soft-argmax to obtain continuous 2D coordinates:

$$\hat p_k = \sum_{x,y} (x,y)\;M_k(x,y)\,,\quad M_k(x,y)=\frac{\exp\bigl(H_k(x,y)/T\bigr)}{\sum_{u,v}\exp\bigl(H_k(u,v)/T\bigr)}\,,\;T=0.1\,.$$

At inference time, arg-max on the heat map recovers integer coordinates (Wang et al., 7 Jul 2025).

The complete pipeline is summarized in the paper as CBCT $Q/K/V$0 ViT+LoRA $Q/K/V$1 heat-map head $Q/K/V$2 soft-argmax $Q/K/V$3 coordinate and geometric loss (Wang et al., 7 Jul 2025). In functional terms, GeoSapiens is therefore a heat-map regression model with parameter-efficient foundation-model adaptation and explicit geometric regularization.

3. Geometric priors and loss construction

A defining feature of GeoSapiens is its geometric loss, which encodes two stated priors on dental landmark configurations. First, the tooth-axis line (AP–CP) is perpendicular to three level lines: root apex, apical-third, and mid-root. Second, these three level lines are mutually parallel (Wang et al., 7 Jul 2025). These priors are imposed after extracting the 16 landmark points.

The method fits each relevant line by least squares and computes unit direction vectors:

• $Q/K/V$4 for the AP–CP axis,
• $Q/K/V$5 for the three horizontal levels (Wang et al., 7 Jul 2025).

The geometric term is then defined as

$Q/K/V$6

The paper explains the construction as follows: $Q/K/V$7 if and only if the vectors are orthogonal, while $Q/K/V$8 if and only if the vectors are co-linear, so deviations from parallelism are penalized by $Q/K/V$9. The factor $r=4$0 normalizes the contributions equally (Wang et al., 7 Jul 2025).

The total loss is

$r=4$1

Here $r=4$2 is mean squared error over all heat-map pixels, and $r=4$3 is the geometric term above (Wang et al., 7 Jul 2025).

This design suggests a hybrid objective in which dense heat-map supervision determines local landmark evidence while the geometric term regularizes global configuration. The reported training plot for $r=4$4 shows steady convergence to near-zero perpendicular/parallelity error (Wang et al., 7 Jul 2025), which is consistent with the intended effect of enforcing line-level consistency.

4. Few-shot adaptation protocol and optimization

The few-shot learning strategy does not employ episodic meta-learning. Instead, the 36 training images constitute the support set for fine-tuning Sapiens+LoRA, with validation on held-out patients (Wang et al., 7 Jul 2025). This makes GeoSapiens an adaptation-centric framework in the transfer-learning sense.

Pre-processing is limited. Each image is cropped to a dentist-annotated tooth bounding box, and the paper states that it does not employ extensive synthetic augmentation. Only basic random flips and intensity scalings are used to avoid overfitting (Wang et al., 7 Jul 2025). A plausible implication is that the reported gains are not primarily attributable to aggressive augmentation policies.

Optimization uses AdamW with base learning rate $r=4$5 and weight decay $r=4$6 (Wang et al., 7 Jul 2025). The learning-rate schedule consists of a linear warm-up from $r=4$7 to $r=4$8 over the first 500 steps, followed by multi-step decay by a factor of $r=4$9 at epochs 170 and 200 (Wang et al., 7 Jul 2025). Training uses batch size 16 for approximately 230 epochs on a single NVIDIA RTX 3090 GPU, with total fine-tuning time of approximately 2 hours (Wang et al., 7 Jul 2025).

These implementation choices position GeoSapiens as a computationally modest fine-tuning procedure relative to the scale of the underlying foundation model. The paper’s emphasis on LoRA also indicates that parameter efficiency is treated as a practical requirement, not only as a regularization device.

5. Empirical performance

Evaluation uses two standard metrics: Mean Radial Error (MRE) in millimeters and Success Detection Rate (SDR), defined as the fraction of points within thresholds $\alpha=4$0 mm (Wang et al., 7 Jul 2025). On the few-shot LDTeeth test set, GeoSapiens is compared with GU2Net, FM-OSD, and NFDP.

Method	[email protected] mm	MRE (mm)
GU2Net	45.21%	1.312
FM-OSD	32.95%	1.520
NFDP	55.01%	0.825
GeoSapiens	63.19%	0.747

The full reported metrics are: GU2Net with SDR@1 mm $\alpha=4$1, SDR@2 mm $\alpha=4$2, and mean SDR $\alpha=4$3; FM-OSD with SDR@1 mm $\alpha=4$4, SDR@2 mm $\alpha=4$5, and mean SDR $\alpha=4$6; NFDP with SDR@1 mm $\alpha=4$7, SDR@2 mm $\alpha=4$8, and mean SDR $\alpha=4$9; and GeoSapiens with SDR@1 mm $r=8$0, SDR@2 mm $r=8$1, and mean SDR $r=8$2 (Wang et al., 7 Jul 2025).

At the clinically critical 0.5 mm threshold, GeoSapiens improves by $r=8$3 percentage points over the next best method, NFDP (Wang et al., 7 Jul 2025). The overall MRE is reduced from $r=8$4 mm to $r=8$5 mm, a reported $r=8$6 decrease (Wang et al., 7 Jul 2025). The abstract also characterizes the 0.5 mm threshold as a standard widely recognized in dental diagnostics (Wang et al., 7 Jul 2025).

Qualitatively, the paper reports that sample overlays of ground-truth landmarks in green and predicted landmarks in red show GeoSapiens points visually closer to the true anatomical locations, especially along the crown and apex regions, than GU2Net and FM-OSD (Wang et al., 7 Jul 2025). Since the evidence is figure-based, this qualitative assessment should be read as the authors’ comparative visual interpretation.

6. Ablations, significance, and stated limitations

The ablation study isolates the contributions of geometric loss and LoRA (Wang et al., 7 Jul 2025). Adding geometric loss alone to full fine-tuning raises [email protected] mm from $r=8$7 to $r=8$8. LoRA alone preserves approximately $r=8$9 at 0.5 mm with $\alpha=8$0 fewer weights. GeoSapiens, combining LoRA and geometric loss, yields $\alpha=8$1 at 0.5 mm, which is a net $\alpha=8$2 percentage points over LoRA-only (Wang et al., 7 Jul 2025).

These results indicate that the geometric constraint is beneficial in a fully fine-tuned setting and that LoRA retains comparable strict-threshold performance with a much smaller trainable parameter budget. They also suggest that the joint LoRA-plus-geometry configuration does not maximize [email protected] mm relative to all ablated variants, even though it defines the named method and delivers the best comparison result against prior baselines on the reported test set. This distinction is important for interpreting GeoSapiens as a trade-off among transfer efficiency, parameter count, and geometric regularization rather than as a single-axis optimization.

The paper identifies several limitations. The current model is 2D and therefore does not address out-of-plane lesions. It is trained exclusively on anterior teeth, leaving posterior and mixed dentition open. The few-shot regime uses only 3 patients, and the authors state that performance may further improve with even 10–20 cases (Wang et al., 7 Jul 2025). These are stated constraints rather than speculative criticisms.

7. Clinical relevance and future directions

The paper states that a 0.5 mm detection error is often the tolerance boundary in orthodontic and implant planning, and that GeoSapiens’ $\alpha=8$3, compared with $\alpha=8$4 for others, marks a significant advance (Wang et al., 7 Jul 2025). It further states that automating landmark annotation can save dentists approximately 5 minutes per tooth, reduce inter-observer variability, and enable large-scale epidemiological studies (Wang et al., 7 Jul 2025). These claims situate the model within workflow efficiency and standardization rather than only benchmark competition.

The future directions proposed in the paper are explicit. They include extending the framework to full 3D CBCT volumes by integrating volumetric transformers or spline-based heat-fields; jointly learning tooth segmentation and landmark detection to better encode context; exploring self-supervised pre-training on unlabeled dental CBCT images to further reduce annotation needs; and conducting clinical validation in a prospective study, including inter-operator comparisons and treatment outcome correlations (Wang et al., 7 Jul 2025). Code availability is also stated through a public repository: https://github.com/xmed-lab/GeoSapiens (Wang et al., 7 Jul 2025).

Taken together, GeoSapiens can be characterized as a foundation-model adaptation framework for few-shot dental landmark detection that couples human-centric pre-training, parameter-efficient fine-tuning, and explicit anatomical geometry (Wang et al., 7 Jul 2025). A plausible implication is that its broader significance lies in demonstrating that representations learned from large-scale human-centric vision data can be repurposed for narrowly annotated dental CBCT tasks when combined with domain-specific structural priors.