Boundary-Sensitive Deep Supervision Loss

• Boundary-sensitive deep supervision loss is a training objective that emphasizes errors at region interfaces by propagating boundary signals across multiple intermediate scales.
• It employs methods such as signed-distance maps, entropy weighting, and gradient-derived weights to address class imbalance and capture fine structures.
• Empirical studies show that this approach improves contour sharpness, reduces false positives, and enhances overall segmentation performance in diverse applications.

Searching arXiv for the cited and closely related papers on boundary-sensitive supervision. Boundary-sensitive Deep Supervision Loss denotes a family of training objectives in which supervision is concentrated on interfaces between regions, and that supervision is propagated not only to the final prediction but also to intermediate outputs or scales. The common motivation is that widely used regional losses such as cross-entropy, Dice, or IoU optimize integrals over regions, whereas boundary errors often occupy a small fraction of pixels or frames and can therefore contribute weakly to backpropagation, especially under severe class imbalance or when fine structures are thin, ambiguous, or noisy. In the literature, this family includes signed-distance boundary losses, entropy-weighted boundary objectives, differentiable boundary-overlap surrogates, transformation-aware contour losses, and temporally localized boundary regression, with deep supervision used either as an implemented component or as a natural multi-scale extension (Kervadec et al., 2018, Ehab et al., 12 Apr 2026).

1. Motivation and problem setting

A central premise of boundary-sensitive supervision is that region losses and boundary quality are not equivalent. In highly unbalanced segmentation, foreground and background terms can differ by orders of magnitude, which yields noisy gradients for rare labels and unstable training. The boundary-loss formulation of Kervadec et al. therefore measures discrepancy at interfaces rather than over unbalanced volumes, framing the objective as a distance on the space of contours rather than regions (Kervadec et al., 2018).

Related motivations recur across domains. In monocular depth estimation, boundary blur is attributed to two factors: low-level boundary and structure information may be lost in deeper networks, and boundary pixels form only a small fraction of the image, so the model can ignore the errors introduced by the boundary area during backpropagation (Yang et al., 2021). In remote sensing, multiple instances of one class with precisely defined boundaries are often the case, and the accuracy of boundary delineation influences the quality of the segmented areas explicitly (Bokhovkin et al., 2019). In plant seedling segmentation, intricate backgrounds and fine leaf structures make leaf edges ambiguous, and pixels along leaf edges tend to be ambiguous, with probabilities near $0.5$, which yield high entropy (Ehab et al., 12 Apr 2026).

The same logic extends beyond spatial image segmentation. In fine-grained Temporal Action Segmentation, the boundary region is supervised within a tolerance window around action transitions, because sparse transition frames otherwise create an imbalance between boundary and non-boundary supervision (Mitsuoka et al., 2 Apr 2026). In breast ultrasound, lesion images and no-lesion images present two distinct failure modes—boundary leakage and false-positive activations in normal images—and the training objective is modified so that contour penalties are concentrated only where a ground-truth boundary exists and the network remains uncertain (Alsaid et al., 21 Jun 2026).

A common misconception is that boundary sensitivity necessarily requires explicit edge masks or specialized edge decoders. Some methods do rely on explicit boundary maps or signed distance maps, but others use predictive entropy or region-only reformulations as boundary surrogates. This suggests that “boundary-sensitive” refers more to the geometry of the optimization signal than to any single implementation pattern.

2. Mathematical forms of boundary-sensitive supervision

One major formulation is the signed-distance boundary loss. Let $\phi_G$ denote the signed distance map of the ground-truth foreground region $G$, negative inside $G$ and positive outside. The binary loss is

$$\mathcal{L}_B(\theta) = \int_{\Omega} \phi_G(x)\, s_\theta(x)\, dx,$$

and the multi-class extension sums class-specific signed distance maps against softmax probabilities. Its gradient with respect to the foreground probability is simply $\phi_G(x)$, so pixels inside $G$ push predictions upward and pixels outside push them downward, with magnitude determined by distance to the boundary (Kervadec et al., 2018).

A second family uses predictive uncertainty as a boundary proxy. In UGDA-Net, the binary predictive entropy is

$H_i = -[p_i \log(p_i) + (1 - p_i) \log(1 - p_i)],$

with weight

$w_i = 1 + \beta H_i, \qquad \beta = 0.3,$

applied to BCE but not to Dice. The weighted BCE and hybrid loss are

$$L^{WBCE} = \frac{1}{N} \sum_{i=1}^{N} w_i L_i^{BCE}, \qquad L_{hybrid} = \lambda_1 L^{WBCE} + \lambda_2 L^{Dice},$$

with $\phi_G$0 and $\phi_G$1, and the first 3 epochs use standard BCE to stabilize early optimization. The key claim is that no morphological boundary masks or gradient operators are used; uncertainty correlates with boundaries because probabilities are most ambiguous where the classifier transitions between classes (Ehab et al., 12 Apr 2026).

Entropy-guided boundary supervision in breast ultrasound uses a different gating rule. The predictive entropy is again

$\phi_G$2

but the weight is

$\phi_G$3

where $\phi_G$4 is the ground-truth boundary map. The uncertainty-guided boundary loss is

$\phi_G$5

Because $\phi_G$6 on normal images, the boundary term contributes no gradient there by design (Alsaid et al., 21 Jun 2026).

A third family uses gradient-derived weights. The Boundary Aware Depth loss is

$\phi_G$7

with

$\phi_G$8

Here Sobel gradients define a continuous, per-pixel boundary-aware weight without binary masking (Yang et al., 2021).

Other formulations act directly on boundary overlap or structured misalignment. The remote-sensing BF1 surrogate extracts thin boundaries and expanded boundary bands by max-pooling on inverted masks, computes surrogate precision and recall, and defines $\phi_G$9 (Bokhovkin et al., 2019). InverseForm defines a structured contour distance by partitioning boundary maps into tiles, estimating a transformation $G$0 between predicted and ground-truth boundary tiles via a frozen inverse-transformation network, and summing tile distances $G$1 relative to identity (Borse et al., 2021). Boundary DoU instead uses only region calculation,

$G$2

with $G$3, $G$4, and $G$5; the paper presents it as a stable boundary-oriented loss that does not require additional losses (Sun et al., 2023). For crisp boundary detection, DRNet uses a different boundary-sensitive objective, namely Adaptive Weighting Loss,

$G$6

where $G$7 is a soft boundary label map formed by annotator averaging, and $G$8 and $G$9 are trainable scalars initialized to $G$0 (Cao et al., 2021).

3. Deep supervision and multi-scale propagation

When boundary-sensitive losses are combined with deep supervision, the same contour-aware signal is applied to intermediate predictions rather than only to the final output. In UGDA-Net, two auxiliary segmentation heads are attached to encoder stages 4 and 5 in a ResNet34 encoder. Each head is a $G$1 convolution producing a single-channel logit map, upsampled by bilinear interpolation to input resolution, and supervised with the same entropy-weighted hybrid loss. The total objective is

$G$2

with $G$3, $G$4, and $G$5. The stated effect is multi-scale boundary supervision: deeper features receive stronger guidance, encouraging early layers to encode boundary-aware semantics that propagate forward and up the decoder (Ehab et al., 12 Apr 2026).

The general deep-supervision form for signed-distance boundary loss is

$G$6

with

$G$7

The weighting coefficients $G$8 control the contribution of each output, while $G$9 and $$\mathcal{L}_B(\theta) = \int_{\Omega} \phi_G(x)\, s_\theta(x)\, dx,$$0 control regional and boundary components at all scales (Kervadec et al., 2018).

For the BF1 surrogate, a corresponding deep-supervision formulation applies $$\mathcal{L}_B(\theta) = \int_{\Omega} \phi_G(x)\, s_\theta(x)\, dx,$$1 and $$\mathcal{L}_B(\theta) = \int_{\Omega} \phi_G(x)\, s_\theta(x)\, dx,$$2 at each decoder stage, with scale-dependent boundary kernels $$\mathcal{L}_B(\theta) = \int_{\Omega} \phi_G(x)\, s_\theta(x)\, dx,$$3 and $$\mathcal{L}_B(\theta) = \int_{\Omega} \phi_G(x)\, s_\theta(x)\, dx,$$4 chosen according to downsampling. The aggregate loss is

$$\mathcal{L}_B(\theta) = \int_{\Omega} \phi_G(x)\, s_\theta(x)\, dx,$$5

The purpose is to avoid vanishing boundary signals at coarse scales while shifting attention from region coverage to boundary alignment (Bokhovkin et al., 2019).

Several works are boundary-sensitive but not deeply supervised in their reported implementation. The depth-estimation model using $$\mathcal{L}_B(\theta) = \int_{\Omega} \phi_G(x)\, s_\theta(x)\, dx,$$6 applies the loss to the final output depth map, while a principled adaptation to side outputs is described by defining $$\mathcal{L}_B(\theta) = \int_{\Omega} \phi_G(x)\, s_\theta(x)\, dx,$$7 and summing scale-wise terms with coefficients $$\mathcal{L}_B(\theta) = \int_{\Omega} \phi_G(x)\, s_\theta(x)\, dx,$$8 (Yang et al., 2021). The breast-ultrasound entropy-guided boundary term is also applied only at the final decoder output; if side outputs are added, the same boundary loss can be applied per scale with appropriate downsampled boundary maps and scale weights (Alsaid et al., 21 Jun 2026). DRNet supervises the final fused one-channel boundary map rather than side outputs, and its use as deep supervision is explicitly presented as an implementation extension rather than a reported component of DRNet itself (Cao et al., 2021). In Temporal Action Segmentation, the dual auxiliary losses are added to the final output head, although stage-wise boundary and CDF losses with weights $$\mathcal{L}_B(\theta) = \int_{\Omega} \phi_G(x)\, s_\theta(x)\, dx,$$9 are described as an optional adaptation (Mitsuoka et al., 2 Apr 2026).

4. Architectural couplings and boundary-aware feature refinement

Boundary-sensitive losses are often paired with modules that preserve or amplify boundary-rich features. UGDA-Net couples its deep supervision loss with Uncertainty-Guided Dual Attention. For a feature tensor $\phi_G(x)$0, the uncertainty map is derived from channel variance: $\phi_G(x)$1

$\phi_G(x)$2

Channel and spatial attention are then modulated by uncertainty,

$\phi_G(x)$3

and the residual refinement is

$\phi_G(x)$4

with $\phi_G(x)$5 a learnable scalar initialized to $\phi_G(x)$6. UGDA blocks are inserted after encoder stages 2–5, while auxiliary heads remain attached to stages 4 and 5 (Ehab et al., 12 Apr 2026).

The monocular-depth model addresses boundary blur at the feature level through a Scene Understanding module and a Scale Transform module. SU aggregates multi-scale encoder features, channel-compresses them, resamples them to a common resolution, concatenates them, and fuses them into a 128-channel global scene information tensor. ST resamples this global information to each decoder stage and applies channel attention so that relevant global information is injected into multiple decoding phases (Yang et al., 2021). This pairing suggests a division of labor in which architectural modules preserve and route low-level structure while the loss upweights boundary errors.

InverseForm uses a different coupling. A boundary head is attached to intermediate features $\phi_G(x)$7, ground-truth boundaries are obtained with a Sobel operator, and a frozen inverse-transform network $\phi_G(x)$8 receives predicted and target boundary tiles and returns a transformation estimate. The total objective combines segmentation cross-entropy, boundary-weighted cross-entropy, and the transformation-aware contour loss,

$\phi_G(x)$9

The paper emphasizes that the boundary head and $G$0 are removed at test time, so the method adds no inference overhead (Borse et al., 2021).

Other architectures minimize modification. The Temporal Action Segmentation framework adds only one extra class-agnostic boundary channel to the final prediction layer, changing the output dimension to $G$1, with no extra branch or inference-time refinement (Mitsuoka et al., 2 Apr 2026). DRNet instead deepens the refinement path horizontally with stacked refine blocks, skip and adjacent connections, weight convolution, and multi-scale image pyramids, while applying its boundary-sensitive Adaptive Weighting Loss to the final contour map (Cao et al., 2021).

5. Empirical behavior across domains

Reported results show that boundary-sensitive supervision is primarily associated with sharper contour alignment, reduced false positives near boundaries, or improved segmental consistency, although the magnitude and type of improvement vary by task. In plant seedling segmentation, UGDA-Net improves U-Net from $G$2 Dice/IoU to $G$3, and LinkNet from $G$4 to $G$5. The largest single ablation gain comes from the loss-only setting, with $G$6 pp Dice for U-Net and $G$7 pp Dice for LinkNet, while the full model achieves the best overall results. Qualitative overlays show fewer false positives near soil or container edges, and entropy heatmaps align tightly with leaf boundaries and complex morphology (Ehab et al., 12 Apr 2026).

For the signed-distance boundary loss, adding $G$8 to generalized Dice improves ISLES DSC from $G$9 to $H_i = -[p_i \log(p_i) + (1 - p_i) \log(1 - p_i)],$0 with 3D $H_i = -[p_i \log(p_i) + (1 - p_i) \log(1 - p_i)],$1, and reduces HD95 from $H_i = -[p_i \log(p_i) + (1 - p_i) \log(1 - p_i)],$2 mm to $H_i = -[p_i \log(p_i) + (1 - p_i) \log(1 - p_i)],$3 mm; on WMH, DSC improves from $H_i = -[p_i \log(p_i) + (1 - p_i) \log(1 - p_i)],$4 to $H_i = -[p_i \log(p_i) + (1 - p_i) \log(1 - p_i)],$5 and HD95 from $H_i = -[p_i \log(p_i) + (1 - p_i) \log(1 - p_i)],$6 mm to $H_i = -[p_i \log(p_i) + (1 - p_i) \log(1 - p_i)],$7 mm. The paper also reports improved training stability and recovery of small or rare lesions (Kervadec et al., 2018).

In remote sensing, the pooling-based BF1 surrogate improves both IoU and BF1 on synthetic AICD, ISPRS Potsdam, and INRIA Aerial Image Labeling. For ISPRS Potsdam with UNet-ResNet34, IoU rises from $H_i = -[p_i \log(p_i) + (1 - p_i) \log(1 - p_i)],$8 to $H_i = -[p_i \log(p_i) + (1 - p_i) \log(1 - p_i)],$9 and BF1 from $w_i = 1 + \beta H_i, \qquad \beta = 0.3,$0 to $w_i = 1 + \beta H_i, \qquad \beta = 0.3,$1; for INRIA AIL with UNet-Inception-ResNet-v2, IoU rises from $w_i = 1 + \beta H_i, \qquad \beta = 0.3,$2 to $w_i = 1 + \beta H_i, \qquad \beta = 0.3,$3 and BF1 from $w_i = 1 + \beta H_i, \qquad \beta = 0.3,$4 to $w_i = 1 + \beta H_i, \qquad \beta = 0.3,$5. Qualitatively, the paper reports sharper boundaries, better corner alignment, and improved separation of adjacent buildings (Bokhovkin et al., 2019).

In monocular depth estimation, adding BAD to the baseline improves $w_i = 1 + \beta H_i, \qquad \beta = 0.3,$6 from $w_i = 1 + \beta H_i, \qquad \beta = 0.3,$7 to $w_i = 1 + \beta H_i, \qquad \beta = 0.3,$8, AbsRel from $w_i = 1 + \beta H_i, \qquad \beta = 0.3,$9 to $$L^{WBCE} = \frac{1}{N} \sum_{i=1}^{N} w_i L_i^{BCE}, \qquad L_{hybrid} = \lambda_1 L^{WBCE} + \lambda_2 L^{Dice},$$0, RMSE from $$L^{WBCE} = \frac{1}{N} \sum_{i=1}^{N} w_i L_i^{BCE}, \qquad L_{hybrid} = \lambda_1 L^{WBCE} + \lambda_2 L^{Dice},$$1 to $$L^{WBCE} = \frac{1}{N} \sum_{i=1}^{N} w_i L_i^{BCE}, \qquad L_{hybrid} = \lambda_1 L^{WBCE} + \lambda_2 L^{Dice},$$2, and edge F1 at threshold $$L^{WBCE} = \frac{1}{N} \sum_{i=1}^{N} w_i L_i^{BCE}, \qquad L_{hybrid} = \lambda_1 L^{WBCE} + \lambda_2 L^{Dice},$$3 from $$L^{WBCE} = \frac{1}{N} \sum_{i=1}^{N} w_i L_i^{BCE}, \qquad L_{hybrid} = \lambda_1 L^{WBCE} + \lambda_2 L^{Dice},$$4 to $$L^{WBCE} = \frac{1}{N} \sum_{i=1}^{N} w_i L_i^{BCE}, \qquad L_{hybrid} = \lambda_1 L^{WBCE} + \lambda_2 L^{Dice},$$5. With the full SU+ST+BAD model, $$L^{WBCE} = \frac{1}{N} \sum_{i=1}^{N} w_i L_i^{BCE}, \qquad L_{hybrid} = \lambda_1 L^{WBCE} + \lambda_2 L^{Dice},$$6, RMSE $$L^{WBCE} = \frac{1}{N} \sum_{i=1}^{N} w_i L_i^{BCE}, \qquad L_{hybrid} = \lambda_1 L^{WBCE} + \lambda_2 L^{Dice},$$7, and edge F1 at threshold $$L^{WBCE} = \frac{1}{N} \sum_{i=1}^{N} w_i L_i^{BCE}, \qquad L_{hybrid} = \lambda_1 L^{WBCE} + \lambda_2 L^{Dice},$$8, with qualitative reductions in “flying pixels” and clearer object contours (Yang et al., 2021).

In fine-grained Temporal Action Segmentation, the lightweight dual-loss framework improves segmental metrics across MS-TCN, C2F-TCN, and FACT while leaving frame-wise accuracy largely unchanged. For MS-TCN on GTEA, Edit rises from $$L^{WBCE} = \frac{1}{N} \sum_{i=1}^{N} w_i L_i^{BCE}, \qquad L_{hybrid} = \lambda_1 L^{WBCE} + \lambda_2 L^{Dice},$$9 to $\phi_G$00 and F1@10 from $\phi_G$01 to $\phi_G$02; on 50Salads, Edit rises from $\phi_G$03 to $\phi_G$04; on Breakfast, F1@10 rises from $\phi_G$05 to $\phi_G$06 (Mitsuoka et al., 2 Apr 2026).

In breast ultrasound, the proposed entropy-guided boundary supervision preserves lesion Dice on lesion-containing images—$\phi_G$07 versus $\phi_G$08 for the no-boundary baseline, with paired Wilcoxon $\phi_G$09—but changes specificity on normal images substantially: false-positive activations fall from $\phi_G$10 and $\phi_G$11 for the two baselines to $\phi_G$12. Spatial temperature scaling further reduces ECE from $\phi_G$13 to $\phi_G$14 without altering thresholded masks (Alsaid et al., 21 Jun 2026).

Transformation-aware boundary supervision in InverseForm improves both region and boundary metrics across NYU-Depth-v2, PASCAL, and Cityscapes. For example, on NYU-Depth-v2 with HRNet-w48 in single-task segmentation, mIoU rises from $\phi_G$15 to $\phi_G$16 and mean Boundary Accuracy from $\phi_G$17 to $\phi_G$18; on Cityscapes, HRNet-48+OCR improves from $\phi_G$19 mIoU with SegFix to $\phi_G$20 with InverseForm, and from $\phi_G$21 to $\phi_G$22 when combined with HMS attention (Borse et al., 2021). Boundary DoU likewise reports consistent improvements across UNet, TransUNet, and Swin-UNet on Synapse and ACDC, including Dice, HD, and Boundary IoU gains while using a region-only formulation (Sun et al., 2023).

6. Limitations, design choices, and scope

Boundary-sensitive deep supervision is not a single loss but a design space with recurring trade-offs. One major design choice is whether the boundary signal is explicit or implicit. Signed distance maps, dilated contour masks, gradient magnitudes, Sobel edges, and pooled BF1 surrogates are explicit constructions; entropy-weighted losses use uncertainty as an adaptive boundary surrogate, either with no explicit boundary masks at all or with gating by a ground-truth boundary map (Kervadec et al., 2018, Ehab et al., 12 Apr 2026).

Another design choice is whether the boundary term can stand alone. Kervadec et al. state that boundary loss alone can admit trivial low-gradient solutions near empty foregrounds and recommend combining it with a regional term and using a schedule for $\phi_G$23 (Kervadec et al., 2018). By contrast, Boundary DoU is presented as a stable loss that does not need any additional losses (Sun et al., 2023). UGDA-Net stabilizes its uncertainty-weighted objective with a BCE warm-up and a small residual-attention initialization, while noting that extremely thin structures or highly noisy backgrounds may remain challenging when entropy is uniformly high (Ehab et al., 12 Apr 2026).

Deep supervision itself is not universal. Some approaches explicitly supervise multiple scales, but others supervise only the final output and merely admit a principled multi-scale extension. This distinction matters because intermediate boundary supervision changes optimization dynamics even when the per-output loss is unchanged. A plausible implication is that the term “boundary-sensitive deep supervision loss” should be reserved most strictly for objectives in which boundary-aware terms are actually aggregated across intermediate outputs, rather than for any boundary-aware loss applied at a single final head.

Annotation fidelity and discretization are recurrent limitations. Signed distance maps depend on accurate contours and correct spacing; coarse voxel spacing or aggressive downsampling can blur boundaries (Kervadec et al., 2018). Boundary penalties can overfit to noisy hand-annotated edges in remote sensing or ultrasound, and the BUS study notes persistent failure on extreme boundary leakage and on normal images with large posterior acoustic shadows (Bokhovkin et al., 2019, Alsaid et al., 21 Jun 2026). InverseForm’s geodesic variant is theoretically grounded for homographies but prone to instability, whereas its Euclidean affine variant is reported as more robust in practice (Borse et al., 2021).

Across these variants, the shared principle is consistent: boundary-sensitive deep supervision augments standard supervision with signals that are spatially or temporally localized near interfaces, and distributes those signals across the prediction hierarchy. This yields a training objective that is more responsive to contour alignment, thin structures, transition frames, and ambiguity at class boundaries than purely regional supervision.