BREA-Depth: Airway Depth Estimation

• The paper introduces BREA-Depth, an anatomically constrained monocular depth estimator for bronchoscopy that leverages synthetic airway geometry and foundation-model pseudo-depth for domain adaptation.
• It employs a depth-aware CycleGAN architecture with dual U-Net transformer branches and an airway structure awareness loss to enforce lumen-centered depth ordering and smooth transitions.
• Empirical evaluation on ex vivo and phantom datasets demonstrates improved 3D airway reconstruction and real-time navigation, despite challenges with conventional depth metrics and phantom ground-truth quality.

Searching arXiv for BREA-Depth and closely related bronchoscopy depth-estimation papers. BREA-Depth, introduced as "BREA-Depth: Bronchoscopy Realistic Airway-geometric Depth Estimation" (Zhang et al., 15 Sep 2025), is a monocular depth-estimation framework for bronchoscopy that incorporates airway-specific geometric priors into foundation-model adaptation. It is designed for real-time navigation in complex branching airways, 3D airway reconstruction, and depth-aware support for interventions such as targeted biopsy or localized drug delivery. The framework is motivated by the observation that generic depth foundation models, including Depth Anything, can generalize to endoscopic imagery yet often remain anatomically underconstrained in bronchoscopy, emphasizing local texture rather than airway structure under ambiguous depth cues and poor lighting (Zhang et al., 15 Sep 2025).

1. Clinical and methodological setting

The problem addressed by BREA-Depth is monocular depth estimation from a single bronchoscopic RGB frame. In bronchoscopy, reliable ground-truth depth is difficult to obtain, and earlier approaches often depend on synthetic depth derived from computed tomography. The paper argues that such depth maps are frequently oversimplified and may lack realistic variation, which limits transfer to real bronchoscopic video. It further argues that airway scenes are structurally constrained but visually challenging: the lumen often exhibits weak texture, illumination is poor, and global branching topology is more informative than local appearance alone (Zhang et al., 15 Sep 2025).

A central premise of the method is that bronchoscopic depth should encode airway anatomy rather than only photometric cues. The specific prior emphasized by the paper is that the airway lumen is generally the deepest region in a bronchoscopic view. From this, the method aims to enforce global structural consistency, lumen-centered depth ordering, and smooth transitions around airway walls and bifurcations. This suggests that BREA-Depth is intended less as a generic monocular depth predictor than as an anatomically regularized estimator specialized to respiratory endoscopy.

2. Anatomical priors and synthetic airway generation

BREA-Depth grounds its anatomical prior in a geometrically accurate synthetic airway model built in Blender. The synthetic bronchial tree incorporates airway branching, bifurcation structure, and carinal rounding derived from anatomical and airflow-based models. Airways are modeled as cylindrical branches with binary branching, and the paper parameterizes daughter-to-parent diameter ratios as

$h_a = D_a / D_p, \qquad h_b = D_b / D_p,$

with $h_a, h_b \in (0,1)$. Branch curvature is constrained through

$$R_a^* = \frac{D_a}{2 \sin \Phi_a}, \qquad R_b^* = \frac{D_b}{2 \sin \Phi_b},$$

while additional geometric variability is introduced with

$$\Theta \sim \mathcal{U}(0^\circ, 360^\circ), \qquad L \sim \mathcal{N}(L_{\text{mean}}, 0.3L_{\text{mean}}).$$

The resulting synthetic corpus contains 9,500 image-depth pairs (Zhang et al., 15 Sep 2025).

These synthetic data provide perfect depth supervision and anatomically constrained geometry, but they do not match real bronchoscopic appearance. The framework therefore treats them as a geometric source domain rather than a complete training solution. A plausible implication is that the synthetic model serves two roles simultaneously: it provides paired depth supervision unavailable in vivo, and it anchors the learned representation to explicit airway structure before domain adaptation to real imagery.

3. Depth-aware CycleGAN architecture

The core network is a depth-aware CycleGAN with two U-Net Transformer-like encoder-decoder branches and a PatchGAN discriminator. One branch maps synthetic RGB-depth input to realistic-style outputs; the other maps real bronchoscopic images to synthetic-style imagery while predicting depth. The paper describes the two directional mappings as

$$X_s = \{I_s, D_s\}, \qquad Z_s = E_s(D_s), \qquad \hat{Y}_s = D_r(Z_s), \qquad \hat{Y}_s = \{\hat{I}_r, \hat{D}_r\},$$

for synthetic-to-real translation, and

$$X_r = \{I_r\}, \qquad Z_r = E_r(X_r), \qquad \hat{Y}_r = D_s(Z_r), \qquad \hat{Y}_r = \{\hat{I}_s, \hat{D}_s\},$$

for real-to-synthetic translation. A second cycle reconstructs $\{I'_s, D'_s\}$ and $\{I'_r, D'_r\}$ for consistency losses (Zhang et al., 15 Sep 2025).

The framework is termed “depth-aware” because depth is integrated directly into the translation process rather than added as a detached auxiliary prediction. Real-image supervision is provided through pseudo-depth generated by Depth Anything on 55,000 bronchoscopic images. The total objective is

$$\mathcal{L}_{\text{total}} = \lambda_{\text{adv}} \mathcal{L}_{\text{adv}} + \lambda_{\text{cycle}} \mathcal{L}_{\text{cycle}} + \lambda_{\text{identity}} \mathcal{L}_{\text{identity}} + \lambda_{\text{airway}} \mathcal{L}_{\text{airway}},$$

with

$$\lambda_{\text{adv}} = 5,\qquad \lambda_{\text{cycle}} = 1,\qquad \lambda_{\text{identity}} = 1,\qquad \lambda_{\text{airway}} = 0.5.$$

The paper states that $h_a, h_b \in (0,1)$0 and $h_a, h_b \in (0,1)$1 are computed with $h_a, h_b \in (0,1)$2 loss, but it does not provide explicit closed forms for $h_a, h_b \in (0,1)$3 or a separate pseudo-depth regression term (Zhang et al., 15 Sep 2025).

Component	Function	Supervision source
Geometric airway model	Generates synthetic RGB and perfect depth	Blender airway synthesis
Syn-to-Real branch	Translates synthetic geometry into realistic style	Synthetic RGB-depth pairs
Real-to-Syn branch	Maps real bronchoscopy to synthetic-style image and depth	Real RGB with pseudo-depth
PatchGAN discriminator	Enforces target-domain realism	Adversarial training
Airway structure awareness loss	Enforces lumen-centered depth ordering	Anatomical prior

This architecture bridges the domain gap by combining synthetic geometric supervision with unpaired real-image adaptation. The paper positions this as a response to earlier style-transfer methods that preserve appearance poorly or fail to retain bronchoscopic structure.

4. Airway structure awareness and structural evaluation

The principal anatomical regularizer is the Airway Structure Awareness Loss. The airway region is extracted by thresholding the grayscale-translated image: $h_a, h_b \in (0,1)$4 Given the predicted disparity or depth representation $h_a, h_b \in (0,1)$5, the mean values inside and outside the airway mask are

$h_a, h_b \in (0,1)$6

The loss is then

$h_a, h_b \in (0,1)$7

In the paper’s interpretation, this penalizes anatomically implausible predictions in which the lumen is not deeper than surrounding structures (Zhang et al., 15 Sep 2025).

The framework also introduces Airway Depth Structure Evaluation, motivated by the claim that conventional metrics such as Abs Rel do not measure preservation of global airway geometry. The first component, Lowest Depth Localization Accuracy, defines

$h_a, h_b \in (0,1)$8

and measures whether minimum-depth pixels lie inside the lumen mask: $h_a, h_b \in (0,1)$9 The second component, Depth Contrast Consistency, compares mean depth inside and outside the lumen: $$R_a^* = \frac{D_a}{2 \sin \Phi_a}, \qquad R_b^* = \frac{D_b}{2 \sin \Phi_b},$$0 with z-score

$$R_a^* = \frac{D_a}{2 \sin \Phi_a}, \qquad R_b^* = \frac{D_b}{2 \sin \Phi_b},$$1

The paper states that $$R_a^* = \frac{D_a}{2 \sin \Phi_a}, \qquad R_b^* = \frac{D_b}{2 \sin \Phi_b},$$2 indicates correct localization of the lowest depth in the lumen, while $$R_a^* = \frac{D_a}{2 \sin \Phi_a}, \qquad R_b^* = \frac{D_b}{2 \sin \Phi_b},$$3 indicates that the lumen is significantly deeper (Zhang et al., 15 Sep 2025).

These definitions make explicit that the framework evaluates anatomical realism as an ordering problem, not solely as pixelwise regression. The paper also states that the method preserves smooth transitions and structural integrity, although the only explicitly formalized anatomical loss is the lumen-versus-non-lumen disparity ordering term above.

5. Datasets, baselines, and empirical profile

The paper uses three distinct data resources. First, it introduces an ex vivo human lung dataset collected with an Ambu aScope 4 bronchoscope operated by an expert bronchoscopist. The structural-evaluation subset consists of five navigation sequences, over 20 minutes of data, recorded at 2.5 fps, totaling 3,437 frames. Each frame was manually prompted and then automatically segmented with SegmentAnythingV2 to identify the airway lumen. Second, the training regime uses 55,000 real bronchoscopic images supervised by pseudo-depth from Depth Anything. Third, evaluation on conventional depth metrics uses the Visentini-Scarzanella et al. bronchoscopic dataset, which contains 16 videos, 39,599 frames, a bronchial phantom, ground-truth depth, and CT renderings (Zhang et al., 15 Sep 2025).

The reported baselines are 3cGAN, Depth Anything, Depth Anything V2, EndoDAC, EndoOmni$$R_a^* = \frac{D_a}{2 \sin \Phi_a}, \qquad R_b^* = \frac{D_b}{2 \sin \Phi_b},$$4, and EndoOmni$$R_a^* = \frac{D_a}{2 \sin \Phi_a}, \qquad R_b^* = \frac{D_b}{2 \sin \Phi_b},$$5. For foundation models, the largest versions are used except for EndoOmni, which is evaluated in base and large variants. On the ex vivo dataset, BREA-Depth reports 97.27 DepthCon and 62.36 LocalAccu. On the phantom dataset, after median alignment post-processing, it reports Abs Rel $$R_a^* = \frac{D_a}{2 \sin \Phi_a}, \qquad R_b^* = \frac{D_b}{2 \sin \Phi_b},$$6, Sq Rel $$R_a^* = \frac{D_a}{2 \sin \Phi_a}, \qquad R_b^* = \frac{D_b}{2 \sin \Phi_b},$$7, RMSE $$R_a^* = \frac{D_a}{2 \sin \Phi_a}, \qquad R_b^* = \frac{D_b}{2 \sin \Phi_b},$$8, RMSE log $$R_a^* = \frac{D_a}{2 \sin \Phi_a}, \qquad R_b^* = \frac{D_b}{2 \sin \Phi_b},$$9, and $$\Theta \sim \mathcal{U}(0^\circ, 360^\circ), \qquad L \sim \mathcal{N}(L_{\text{mean}}, 0.3L_{\text{mean}}).$$0 $$\Theta \sim \mathcal{U}(0^\circ, 360^\circ), \qquad L \sim \mathcal{N}(L_{\text{mean}}, 0.3L_{\text{mean}}).$$1. The same paper reports that EndoOmni$$\Theta \sim \mathcal{U}(0^\circ, 360^\circ), \qquad L \sim \mathcal{N}(L_{\text{mean}}, 0.3L_{\text{mean}}).$$2 attains stronger conventional phantom metrics, with Abs Rel $$\Theta \sim \mathcal{U}(0^\circ, 360^\circ), \qquad L \sim \mathcal{N}(L_{\text{mean}}, 0.3L_{\text{mean}}).$$3, Sq Rel $$\Theta \sim \mathcal{U}(0^\circ, 360^\circ), \qquad L \sim \mathcal{N}(L_{\text{mean}}, 0.3L_{\text{mean}}).$$4, RMSE $$\Theta \sim \mathcal{U}(0^\circ, 360^\circ), \qquad L \sim \mathcal{N}(L_{\text{mean}}, 0.3L_{\text{mean}}).$$5, RMSE log $$\Theta \sim \mathcal{U}(0^\circ, 360^\circ), \qquad L \sim \mathcal{N}(L_{\text{mean}}, 0.3L_{\text{mean}}).$$6, and $$\Theta \sim \mathcal{U}(0^\circ, 360^\circ), \qquad L \sim \mathcal{N}(L_{\text{mean}}, 0.3L_{\text{mean}}).$$7 $$\Theta \sim \mathcal{U}(0^\circ, 360^\circ), \qquad L \sim \mathcal{N}(L_{\text{mean}}, 0.3L_{\text{mean}}).$$8, while BREA-Depth is strongest in structural preservation (Zhang et al., 15 Sep 2025).

The ablation study attributes the main gains to both domain adaptation and anatomical regularization. Relative to the full model, removing the CycleGAN yields 68.36 DepthCon, 25.36 LocalAccu, Abs Rel $$\Theta \sim \mathcal{U}(0^\circ, 360^\circ), \qquad L \sim \mathcal{N}(L_{\text{mean}}, 0.3L_{\text{mean}}).$$9, Sq Rel $$X_s = \{I_s, D_s\}, \qquad Z_s = E_s(D_s), \qquad \hat{Y}_s = D_r(Z_s), \qquad \hat{Y}_s = \{\hat{I}_r, \hat{D}_r\},$$0, RMSE $$X_s = \{I_s, D_s\}, \qquad Z_s = E_s(D_s), \qquad \hat{Y}_s = D_r(Z_s), \qquad \hat{Y}_s = \{\hat{I}_r, \hat{D}_r\},$$1, RMSE log $$X_s = \{I_s, D_s\}, \qquad Z_s = E_s(D_s), \qquad \hat{Y}_s = D_r(Z_s), \qquad \hat{Y}_s = \{\hat{I}_r, \hat{D}_r\},$$2, and $$X_s = \{I_s, D_s\}, \qquad Z_s = E_s(D_s), \qquad \hat{Y}_s = D_r(Z_s), \qquad \hat{Y}_s = \{\hat{I}_r, \hat{D}_r\},$$3 $$X_s = \{I_s, D_s\}, \qquad Z_s = E_s(D_s), \qquad \hat{Y}_s = D_r(Z_s), \qquad \hat{Y}_s = \{\hat{I}_r, \hat{D}_r\},$$4. Removing $$X_s = \{I_s, D_s\}, \qquad Z_s = E_s(D_s), \qquad \hat{Y}_s = D_r(Z_s), \qquad \hat{Y}_s = \{\hat{I}_r, \hat{D}_r\},$$5 yields 96.67 DepthCon, 52.06 LocalAccu, Abs Rel $$X_s = \{I_s, D_s\}, \qquad Z_s = E_s(D_s), \qquad \hat{Y}_s = D_r(Z_s), \qquad \hat{Y}_s = \{\hat{I}_r, \hat{D}_r\},$$6, Sq Rel $$X_s = \{I_s, D_s\}, \qquad Z_s = E_s(D_s), \qquad \hat{Y}_s = D_r(Z_s), \qquad \hat{Y}_s = \{\hat{I}_r, \hat{D}_r\},$$7, RMSE $$X_s = \{I_s, D_s\}, \qquad Z_s = E_s(D_s), \qquad \hat{Y}_s = D_r(Z_s), \qquad \hat{Y}_s = \{\hat{I}_r, \hat{D}_r\},$$8, RMSE log $$X_s = \{I_s, D_s\}, \qquad Z_s = E_s(D_s), \qquad \hat{Y}_s = D_r(Z_s), \qquad \hat{Y}_s = \{\hat{I}_r, \hat{D}_r\},$$9, and $$X_r = \{I_r\}, \qquad Z_r = E_r(X_r), \qquad \hat{Y}_r = D_s(Z_r), \qquad \hat{Y}_r = \{\hat{I}_s, \hat{D}_s\},$$0 $$X_r = \{I_r\}, \qquad Z_r = E_r(X_r), \qquad \hat{Y}_r = D_s(Z_r), \qquad \hat{Y}_r = \{\hat{I}_s, \hat{D}_s\},$$1. The reported implementation details are PyTorch, NVIDIA RTX 3080, 30 epochs, batch size 2, learning rate $$X_r = \{I_r\}, \qquad Z_r = E_r(X_r), \qquad \hat{Y}_r = D_s(Z_r), \qquad \hat{Y}_r = \{\hat{I}_s, \hat{D}_s\},$$2, and inference at 60 FPS. The paper also notes that its strongest classical depth metrics are not obtained on the phantom benchmark, which it attributes to low-quality phantom ground truth that may misassign depth in lumen regions and bifurcations.

One numerical inconsistency is notable. The text states that BREA-Depth achieves the highest performance in airway depth structure evaluation, yet the reported structure table assigns a slightly higher raw DepthCon value to 3cGAN, while BREA-Depth remains best in LocalAccu and is described as best overall. The paper does not explain this discrepancy.

6. Significance, interpretation, and limitations

BREA-Depth’s main contribution is to redefine bronchoscopic monocular depth estimation as an anatomically constrained domain-adaptation problem rather than a pure image-regression task. In this framing, synthetic airway geometry supplies structural supervision, foundation-model pseudo-depth supplies real-domain supervision, and the airway structure loss imposes a clinically motivated ordering prior centered on the lumen. This suggests that the method is optimized more for anatomical plausibility and reconstruction utility than for agreement with phantom-derived depth labels alone (Zhang et al., 15 Sep 2025).

The paper links these design choices to downstream use in real-time navigation and 3D airway reconstruction. It explicitly claims that incorporating anatomical priors yields more robust and accurate 3D airway reconstructions and reports online inference at 60 FPS. Its qualitative discussion emphasizes improved lumen highlighting, finer branch detail, and better preservation of bifurcation structure relative to texture-driven baselines. At the same time, the paper does not provide a direct intervention-level safety study; the claimed safety benefits remain downstream implications rather than experimentally validated clinical outcomes.

Several limitations are explicit. The authors identify the lack of high-quality bronchoscopic depth datasets as a central constraint. They also argue that standard phantom benchmarks are structurally imperfect and can underestimate anatomically faithful methods. Some implementation details are omitted, including optimizer choice, scheduler, input resolution, augmentation policy, and an explicit pseudo-depth regression loss. The paper does not present a dedicated failure-case analysis for secretions, severe occlusion, or rare anatomy. Future work is directed toward higher-quality bronchoscopic depth datasets and tighter integration with bronchoscopy camera pose localization and landmark recognition (Zhang et al., 15 Sep 2025).