Slope-Aware Implant Localization Network

• The paper introduces a slope-aware deep learning framework that explicitly models the 3D implant tilt, improving CBCT-based dental implant localization.
• It employs multi-slice 2D and 3D volumetric architectures with modules like TVP and NDP to integrate anatomical context and textural cues.
• Quantitative evaluations demonstrate that slope-aware loss strategies significantly boost localization precision in clinically challenging scenarios.

A slope-aware implant position prediction network is a deep learning framework specifically designed to localize dental implants in cone-beam computed tomography (CBCT) or similar volumetric imaging modalities, with explicit modeling and supervision of the 3D inclination ("slope") of the implant. This approach addresses clinical challenges in implant planning where the variation in implant tilt across slices, as well as anatomical context and textural cues, are critical yet under-exploited by traditional localization networks. Slope-aware methods such as TCSloT (Yang et al., 2023) and RegFreeNet (Yang et al., 21 Jan 2026) integrate explicit slope prediction, leverage cross-slice or 3D context, and introduce loss terms and network branches responsive to the varying difficulty of accurately localizing implants with high clinical inclination.

1. Fundamentals of Slope-Aware Implant Position Localization

The core problem is to infer the spatial location of dental implants within high-resolution scans, which is complicated by both variability in anatomical context and the often considerable slope or tilt of implanted devices. Unlike generic keypoint or bounding box detection, accurate dental implant planning necessitates high geometric precision due to the risk of injury to surrounding tissues and the requirement for prosthetic fit.

Slope-aware networks explicitly estimate not only the implant center or trajectory but also the 3D angle of the implant as it traverses the patient volume. This slope is clinically meaningful: implants with greater tilt produce rapidly varying projections across adjacent slices and yield greater localization uncertainty unless modeled directly. The slope can be formalized as:

$$k^1 = \frac{N\sum_i x_i z_i - \sum_i x_i \sum_i z_i}{N\sum_i z_i^2 - (\sum_i z_i)^2}, \quad k^2 = \frac{N\sum_i y_i z_i - \sum_i y_i \sum_i z_i}{N\sum_i z_i^2 - (\sum_i z_i)^2}$$

given $N$ sampled points $(x_i, y_i, z_i)$ along the trajectory (Yang et al., 21 Jan 2026).

2. Network Architectures: 2D Slice-Based and 3D Volumetric Approaches

Slope-aware implant localization has been instantiated in two main architectural variants:

• Multi-Slice 2D Networks (e.g., TCSloT):
  • Input: three axial 2D slices from CBCT at $(t-k)$, $t$, $(t+k)$, with optimal $k=7$.
  • Feature extraction via weight-shared ResNet-50 encoders for each slice.
  • Fusion through the Texture Variation Perception (TVP) module employing cross-attention between adjacent and central slices, producing fused features sensitive to contextual and slope-based texture differences.
  • Decoder upsamples features and combines them with Conditional Text Guidance (CLIP-based), allowing for integration of high-level priors such as "left," "middle," or "right" site indications.
  • Regression head outputs Gaussian heatmap (center) and sub-pixel offsets (Yang et al., 2023).
• 3D Volumetric Networks (e.g., RegFreeNet):
  • Input: masked CBCT volume (implant digitally removed).
  • Encoder: Neighboring Distance Perception (NDP) module combines multi-scale dilated convolutional features with anatomical graph convolution outputs to rebuild structural context lost by masking.
  • Four-stage 3D U-Net backbone with skip connections.
  • Dual decoder branches:
  • Implant Position Prediction Branch (IPPB): predicts volumetric segmentation of the implant path.
  • Slope Prediction Branch (SPB): estimates 2D slope vector via fully connected layers.
  • NDP and SPB enhance anatomical awareness and regularization (Yang et al., 21 Jan 2026).

3. Slope-Aware Supervision and Loss Functions

Both TCSloT and RegFreeNet introduce supervision directly responsive to the clinical tilt (slope) of the implant:

• TCSloT Slope-Aware Loss (SAL):
  • For each implant trajectory, compute the total slope magnitude $\tau = |s_1| + |s_2|$ for annotated slices.
  • Normalize $\tau$ to preserve total regression loss scale.
  • Regression loss for implant center heatmap is up-weighted by $\hat\tau$, emphasizing large-slope (challenging) cases:

  $$\mathcal{L}_{\mathrm{SAL}} = \sum_{i=1}^{N^p_c} \hat\tau\,\mathcal{L}_h^i$$ - Final total loss is composite across the slope-aware term, heatmap offset, and knowledge alignment (Yang et al., 2023).

• RegFreeNet Slope-Prediction Branch (SPB):

  • Explicitly outputs and supervises the clinical inclination vector $\hat{k}$.
  • Slope loss:

  $$\mathcal{L}_{\mathrm{Slope}} = \sum_j \|k_j - \hat{k}_j\|_1$$ - Overall loss combines standard volumetric segmentation (cross-entropy + Dice) and slope supervision:

  $$\mathcal{L} = \mathcal{L}_{\mathrm{Seg}} + \lambda \mathcal{L}_{\mathrm{Slope}}$$

  with $\lambda=1$ as empirically optimal (Yang et al., 21 Jan 2026).

Up-weighting or explicitly regressing difficult examples via slope ensures the network’s feature space encodes variability arising from clinical conditions that are most critical in practice.

4. Contextual and Textural Feature Modules

Slope-aware approaches enhance localization by embedding anatomical and contextual cues:

• Texture Variation Perception (TVP) Module (TCSloT):

  • Utilizes cross-attention between central and adjacent slices, capturing evolution of textures such as emerging edentulous zones or implant artifacts across slices.
  • Cross-attention alignment allows perception of 3D transitions without full volumetric computation.
  • Residual fusion further preserves central slice semantics (Yang et al., 2023).
• Neighboring Distance Perception (NDP) Module (RegFreeNet):
  • Multi-scale dilation encodes both local tooth details and larger-scale structural gaps.
  • Keypoint pooling and graph convolution networks impose explicit adjacency priors, encoding relationships such as adjacent teeth or the mandibular canal.
  • Output is residually fused and upsampled, then passed to the 3D encoder (Yang et al., 21 Jan 2026).

A plausible implication is that these modules directly address the loss or distortion of reference context (either through projectional slice limits or masked data), which is a key obstacle in high-precision implant position prediction.

5. Training Protocols and Datasets

Both frameworks employ extensive clinical datasets from Shenzhen University General Hospital:

• TCSloT: 154 patients, 3045 annotated slices, five-fold cross-validation. Data augmentation includes random crop, scaling, and horizontal flips. Adam optimizer with tiered learning rate decay (Yang et al., 2023).
• RegFreeNet: 1,622 full CBCT scans (ImplantFairy), isotropic voxel size 0.2 mm. Masked CBT volumes (implant erased) and random mask perturbations ensure registration-free, robust learning. Subvolumes (128×128×128) randomly sampled; model optimized with AdamW, weight decay, and test-time sliding-window augmentation (Yang et al., 21 Jan 2026).

RegFreeNet enables dramatic scaling by removing the prerequisite for pre/post-operative scan registration, thus facilitating the aggregation of large, multi-center datasets.

6. Quantitative Performance and Ablation

Both slope-aware approaches achieve state-of-the-art results on internal and external datasets.

TCSloT (ResNet-50 Encoder, $AP_{75}$ metric):

Model	$AP_{75}$
CenterNet	10.9%
ATSS	12.1%
RepPoints	11.2%
TCSloT (full)	20.4%
TCSloT (ViT)	21.6–21.9%

Ablations demonstrate that TVP, CTG, and SAL each contribute $\sim$+3–4% on the base; full combination yields the maximal gain. With SAL, over 70% of predictions are within 10 pixels of ground truth, compared to $\sim$60% without SAL (Yang et al., 2023).

RegFreeNet (Dice/IoU):

Model	SUGH Dice	SUGH IoU	Ext. Dice	Ext. IoU
3D U-Net	45.02%	0.3325	22.10%	0.1404
RegFreeNet (full)	47.22%	0.3555	31.87%	0.2058

Ablating NDP or SPB results in an external Dice drop of $\sim$10%. Combined, these modules yield the best accuracy and clinically plausible placements (Yang et al., 21 Jan 2026).

7. Implications, Limitations, and Outlook

Explicit slope modeling—whether by loss weighting (TCSloT) or direct regression supervision (RegFreeNet)—enables networks to address the most error-prone and clinically significant implant localizations. Incorporation of anatomical adjacency and context (via TVP, NDP) further improves robustness, particularly in edentulous or multi-region cases.

A plausible implication is that further integration of large-scale text guidance, as in TCSloT, and harmonization with registration-free paradigms, as in RegFreeNet, may yield even more generalizable and clinically adaptable solutions for implant position planning. Limitations include potential domain shift when applied to other anatomical regions and sensitivity to slice interval (TCSloT) or subvolume selection (RegFreeNet), suggesting directions for future refinement (Yang et al., 2023, Yang et al., 21 Jan 2026).