Multi-Scale Boundary Loss Methods

• Multi-scale boundary loss is a training objective for dense prediction tasks that enhances boundary fidelity by integrating supervision at multiple spatial scales.
• It employs methods such as CWMI, SPW, and cbDice to preserve thin structures, improve contour continuity, and address class imbalance in segmentation.
• Empirical studies demonstrate significant improvements in metrics like Hausdorff Distance and segmentation accuracy on benchmarks such as SNEMI3D, DRIVE, and Cityscapes.

Searching arXiv for papers on multi-scale boundary-aware losses and closely related methods. arxiv_search.query({"^^^^^^^^5search_query5 wavelet mutual information loss\" OR ^^^^^^^^5all:\5 pyramid weighted loss\" OR ^^^^^^^^5all:\5 corrected multi-scale fusion network\" OR ^^^^^^^^5all:\5 boundary Dice loss\"","max_results":^^^^5all:\5search_query5^^^^,"sort_by":"submittedDate","sort_order":"descending"}) Reviewing search results for directly relevant arXiv papers. Multi-Scale Boundary Loss denotes a family of training objectives for dense prediction in which boundary supervision is distributed across multiple spatial scales, and often across orientations, centerlines, or geometric neighborhoods, rather than being imposed only at the native pixel grid. In semantic segmentation and related structured prediction tasks, such losses are motivated by the persistent failure modes of standard pixel-wise or region-overlap objectives on thin structures, small instances, and contour localization. Recent formulations instantiate this idea through complex steerable-pyramid mutual information, adaptive steerable-pyramid weighting, online boundary-hard-sample mining in multi-scale fusion networks, centerline- and distance-aware Dice variants, and power-regularized multi-magnitude losses in physics-informed settings (&&&^^^^5search_query5^^^^&&&, &&&^^^^5all:\5^^^^&&&, &&&^^^^5 OR all:\5^^^^&&&, &&&^^^^5 OR all:\5^^^^&&&, &&&^^^^5 OR all:\5^^^^&&&).

^^^^5all:\5^^^^. Conceptual scope and problem setting

A multi-scale boundary loss is designed for regimes in which segmentation quality depends not only on region overlap but also on contour fidelity, thin-structure continuity, and topological correctness. The motivating observation shared across the literature is that conventional losses such as Cross-Entropy, BCE, Focal, or Dice either optimize each pixel independently or emphasize global set overlap, and therefore tend to underweight boundary pixels, especially under class imbalance and instance imbalance. Smaller instances and thin boundaries are often overshadowed by larger structures, producing jagged, eroded, or missing contours (&&&^^^^5search_query5^^^^&&&).

Within semantic segmentation, two recent ^^^^5 OR all:\5search_query5 OR all:\55^^^^ formulations make this definition explicit. The complex wavelet mutual information (CWMI) loss treats a multi-scale boundary loss as a frequency-domain structural objective: predictions and labels are decomposed into oriented complex bandpass subbands, and mutual information is maximized between corresponding subband responses across scales (&&&^^^^5search_query5^^^^&&&). The steerable pyramid weighted (SPW) loss instead realizes the same concept as a weighting scheme: it up-weights pixels whose ground-truth and current prediction exhibit strong multi-scale, multi-orientation bandpass energy, thereby turning a standard base loss such as cross-entropy into a boundary-focused objective without distance transforms (&&&^^^^5all:\5^^^^&&&).

The same general principle appears in other domains and architectures. In real-time scene parsing, Boundary Corrected Loss (BCL) compensates for the contour degradation induced by low-resolution multi-scale fusion through online selection of boundary-hard samples (&&&^^^^5 OR all:\5^^^^&&&). In vascular segmentation, cbDice introduces scale-adaptive boundary supervision by combining centerline topology with mask-distance and inverse-radius weighting, thereby making the loss responsive to both vessel boundaries and vessel caliber (&&&^^^^5 OR all:\5^^^^&&&). In PINNs for multi-scale PDEs, boundary terms are balanced against interior residuals through power-regularized reconstruction of the loss, so that small-magnitude boundary or initial terms are not ignored during optimization (&&&^^^^5 OR all:\5^^^^&&&).

This suggests that “multi-scale boundary loss” is best understood as a design principle rather than a single formula: the key requirement is that boundary information be represented at more than one spatial or structural scale, and that optimization be biased toward preserving geometrically fragile structures.

^^^^5 OR all:\5^^^^. Frequency-domain formulations: steerable pyramids and complex wavelets

The most explicit frequency-domain formulation is CWMI, introduced in "Complex Wavelet Mutual Information Loss: A Multi-Scale Loss Function for Semantic Segmentation" (&&&^^^^5search_query5^^^^&&&). Its core mechanism is a complex steerable pyramid, described as a redundant, polar-separable, multi-scale, multi-orientation decomposition that produces oriented bandpass subbands across scales together with low- and high-pass components. The stated motivation is that phase and magnitude in these complex subbands preserve local structural information critical for edges, corners, and thin boundaries.

For orientation PRESERVED_PLACEHOLDER_^^^^5search_query5^^^^ in polar frequency coordinates PRESERVED_PLACEHOLDER_^^^^5all:\5^^^^, the bandpass filter is defined as

PRESERVED_PLACEHOLDER_^^^^5 OR all:\5^^^^

with angular component

PRESERVED_PLACEHOLDER_^^^^5 OR all:\5^^^^

To encode local phase, the analytic version discards negative frequencies through

PRESERVED_PLACEHOLDER_^^^^5 OR all:\5^^^^

Subbands are computed by FFT-domain masking followed by inverse FFT, or equivalently by complex spatial convolution,

$B_{s,o}(x)\;=\;(X * g_{s,o})(x).$

For an image $\mathbf{I}$ with $N$ pyramid levels and $K$ orientations, the decomposition is written as

$$\mathbf{I} \;\longrightarrow\; \Big\{ \mathbf{I}_{H_0}\in\mathbb{R}^{H_0\times W_0},\; \mathbf{I}_{B_1}\in\mathbb{C}^{K\times H_0\times W_0},\; \mathbf{I}_{B_2}\in\mathbb{C}^{K\times H_1\times W_1},\;\dots,\; \mathbf{I}_{B_N}\in\mathbb{C}^{K\times H_{N-1}\times W_{N-1}},\; \mathbf{I}_{L_0}\in\mathbb{R}^{H_{N-1}\times W_{N-1}} \Big\}.$$

CWMI uses only the oriented complex bandpass subbands for mutual-information computation, not the low- or high-pass residues. The paper reports that all experiments use PRESERVED_PLACEHOLDER_^^^^5all:\5search_query5^^^^ scales and PRESERVED_PLACEHOLDER_^^^^5all:\5all:\5^^^^ orientations, with uniform aggregation across subbands and no bespoke weighting (&&&^^^^5search_query5^^^^&&&).

SPW, presented in "Steerable Pyramid Weighted Loss: Multi-Scale Adaptive Weighting for Semantic Segmentation" (&&&^^^^5all:\5^^^^&&&), uses a standard steerable pyramid rather than a complex mutual-information objective. Its frequency-domain decomposition has the same multi-scale, multi-orientation role: boundaries are treated as energy-rich structures in oriented subbands, especially at finer scales. For any subband PRESERVED_PLACEHOLDER_^^^^5all:\5 OR all:\5^^^^, the local boundary energy is measured through the analytic signal

PRESERVED_PLACEHOLDER_^^^^5all:\5 OR all:\5^^^^

where PRESERVED_PLACEHOLDER_^^^^5all:\5 OR all:\5^^^^ is the Hilbert transform. The magnitude PRESERVED_PLACEHOLDER_^^^^5all:\55^^^^ is then used as a smooth local amplitude envelope.

A plausible implication is that the frequency-domain branch of multi-scale boundary losses differs internally on what is optimized: CWMI aligns prediction and label through statistical dependence in a complex representation, whereas SPW uses the same spectral evidence to modulate the importance of pixels in a conventional loss. Both, however, rely on the same boundary premise: edges and thin structures are more explicitly represented in oriented, multi-scale bandpass responses than in raw pixel space.

^^^^5 OR all:\5^^^^. Loss constructions and mathematical forms

The principal formulations differ in whether multi-scale boundary information enters the objective directly, through per-pixel weighting, through hard-sample selection, or through topology-aware overlap terms.

CWMI defines mutual information on each bandpass level PRESERVED_PLACEHOLDER_^^^^5all:\5^^^^6 by

PRESERVED_PLACEHOLDER_^^^^5all:\5^^^^7

but implements a differentiable Gaussian approximation, following Zhao et al. (RMI), rather than histogram-based MI (&&&^^^^5search_query5^^^^&&&). The local/regional estimator is

PRESERVED_PLACEHOLDER_^^^^5all:\5^^^^8

with

PRESERVED_PLACEHOLDER_^^^^5all:\5^^^^9

Each pixel contributes a PRESERVED_PLACEHOLDER_^^^^5 OR all:\5search_query5^^^^-dimensional vector of directional responses, and sample covariances are computed across spatial samples in the subband. The loss itself is

PRESERVED_PLACEHOLDER_^^^^5 OR all:\5all:\5^^^^

Training uses a hybrid objective

PRESERVED_PLACEHOLDER_^^^^5 OR all:\5 OR all:\5^^^^

SPW uses a weighting map built from both the label and the current prediction. For an image PRESERVED_PLACEHOLDER_^^^^5 OR all:\5 OR all:\5^^^^, the multi-scale map is

PRESERVED_PLACEHOLDER_^^^^5 OR all:\5 OR all:\5^^^^

where PRESERVED_PLACEHOLDER_^^^^5 OR all:\55^^^^ upsamples each subband back to input resolution. The adaptive SPW map is

PRESERVED_PLACEHOLDER_^^^^5 OR all:\5^^^^6

and the final pixel weight is

PRESERVED_PLACEHOLDER_^^^^5 OR all:\5^^^^7

The paper uses weighted cross-entropy

PRESERVED_PLACEHOLDER_^^^^5 OR all:\5^^^^8

BCL in BCMFNet uses neither spectral decomposition nor distance transforms. Its overall objective is

PRESERVED_PLACEHOLDER_^^^^5 OR all:\5^^^^9

where PRESERVED_PLACEHOLDER_^^^^5 OR all:\5search_query5^^^^ is constructed from row- and column-shifted label maps. For step size PRESERVED_PLACEHOLDER_^^^^5 OR all:\5all:\5^^^^, transformed labels are defined by

PRESERVED_PLACEHOLDER_^^^^5 OR all:\5 OR all:\5^^^^

PRESERVED_PLACEHOLDER_^^^^5 OR all:\5 OR all:\5^^^^

Per-pixel cross-entropies are computed against these transformed labels, the maximum of row and column losses is used as a score map, and non-maximum suppression plus thresholding or top-PRESERVED_PLACEHOLDER_^^^^5 OR all:\5 OR all:\5^^^^ selection yields a boundary-hard mask. The per-step boundary-corrected loss is then

PRESERVED_PLACEHOLDER_^^^^5 OR all:\55^^^^

with multi-step aggregation

PRESERVED_PLACEHOLDER_^^^^5 OR all:\5^^^^6

cbDice, introduced in "Centerline Boundary Dice Loss for Vascular Segmentation" (&&&^^^^5 OR all:\5^^^^&&&), takes a different route. It generalizes clDice through the cl-X-Dice family and instantiates cbDice as the cl-MIN-D case, combining normalized distance-to-boundary maps and inverse normalized radii on skeletons. With PRESERVED_PLACEHOLDER_^^^^5 OR all:\5^^^^7, PRESERVED_PLACEHOLDER_^^^^5 OR all:\5^^^^8, and PRESERVED_PLACEHOLDER_^^^^5 OR all:\5^^^^9, the precision and sensitivity terms are

PRESERVED_PLACEHOLDER_^^^^5 OR all:\5search_query5^^^^

PRESERVED_PLACEHOLDER_^^^^5 OR all:\5all:\5^^^^

and

PRESERVED_PLACEHOLDER_^^^^5 OR all:\5 OR all:\5^^^^

By construction, this loss is boundary-aware through distance maps and scale-adaptive through inverse-radius weighting.

In PINNs, the multi-scale boundary idea refers not to image scale but to loss-term magnitude across scales and frequencies. The modified loss is

PRESERVED_PLACEHOLDER_^^^^5 OR all:\5 OR all:\5^^^^

or, with grouped residuals,

PRESERVED_PLACEHOLDER_^^^^5 OR all:\5 OR all:\5^^^^

Here PRESERVED_PLACEHOLDER_^^^^5 OR all:\55^^^^ includes initial and boundary terms, such as Dirichlet, Neumann, and Robin components. The paper states that the power factors equalize magnitudes so that all terms are optimized synchronously (&&&^^^^5 OR all:\5^^^^&&&). This suggests a broader interpretation of multi-scale boundary loss: scale can refer not only to image resolution but also to magnitude scale in coupled objectives.

^^^^5 OR all:\5^^^^. Mechanisms of boundary sensitivity

The boundary sensitivity of these losses arises from distinct but related mechanisms.

In CWMI, boundary sensitivity is attributed to orientation-selective bandpass filters and to the emphasis of mid/high-frequency layers. The paper states that orientation-selective filters respond strongly to edges at specific angles, and that empirically layers ^^^^5 OR all:\5^^^^ and ^^^^5 OR all:\5^^^^ contributed most to region overlap and boundary HD, with the ^^^^5 OR all:\5^^^^th layer best for HD. No explicit subband weights are applied, but the ablations show that higher-frequency subbands aid boundary metrics (&&&^^^^5search_query5^^^^&&&). Mutual information is described as robust to nuisance variation such as small phase shifts and mild noise, so the loss can enforce structural consistency even when exact pixel alignment is imperfect.

SPW’s mechanism is more directly local. High-frequency subbands capture abrupt intensity transitions, edges, and thin structures across orientations, and the analytic magnitude produces a positive, phase-insensitive estimate of edge strength. Because the weight map is computed from both ground-truth and current prediction and updated on-the-fly, the loss can focus learning on regions where the model presently exhibits strong high-frequency content, including suspicious edges and fine structures (&&&^^^^5all:\5^^^^&&&). The paper explicitly contrasts this with distance-transform weighting, which is described as static unless recomputed.

BCL identifies likely boundary pixels through label shifts rather than explicit edge extraction. The transformed labels induce mismatches mainly near contours; per-pixel losses under these transforms therefore act as proxies for boundary uncertainty. Online hard-example mining via non-maximum suppression and thresholding then concentrates supervision on the selected boundary-hard pixels (&&&^^^^5 OR all:\5^^^^&&&). Unlike spectral or distance-based losses, BCL derives its boundary focus from supervision perturbation.

cbDice derives boundary sensitivity from Euclidean distance transforms on masks. The distance maps PRESERVED_PLACEHOLDER_^^^^5 OR all:\5^^^^6 and PRESERVED_PLACEHOLDER_^^^^5 OR all:\5^^^^7 weight centerline matches by proximity to boundaries, while inverse-radius weights amplify the contribution of thinner vessels so that diameter imbalance does not cause large vessels to dominate the optimization (&&&^^^^5 OR all:\5^^^^&&&). The paper states that cl-M-D is sensitive to translations of the mask within the radius PRESERVED_PLACEHOLDER_^^^^5 OR all:\5^^^^8, unlike clDice, and that cbDice matches the behavior of B-DoU under pure translations.

In PINNs, boundary sensitivity takes the form of gradient rebalancing. For a general loss PRESERVED_PLACEHOLDER_^^^^5 OR all:\5^^^^9, the gradient is

$B_{s,o}(x)\;=\;(X * g_{s,o})(x).$^^^^5search_query5^^^^

With $B_{s,o}(x)\;=\;(X * g_{s,o})(x).$^^^^5all:\5^^^^, large loss terms are attenuated and small ones amplified. The paper argues that this prevents PDE residuals from dominating boundary and initial terms, so boundary satisfaction improves across multi-scale or multi-frequency regimes (&&&^^^^5 OR all:\5^^^^&&&).

A common misconception is that multi-scale boundary loss must be equivalent to distance-transform supervision. The recent literature contradicts this. CWMI and SPW are spectral; BCL is shift-based and OHEM-driven; cbDice is skeleton- and distance-based; the PINN formulation is magnitude-balanced. The unifying property is multi-scale boundary emphasis, not any single geometric representation.

5. Empirical evidence and comparative performance

The clearest direct benchmark for a multi-scale boundary loss appears in CWMI. On SNEMI^^^^5 OR all:\5^^^^D, GlaS, DRIVE, and Massachusetts Roads, the paper reports gains in regional and topological metrics together with major Hausdorff Distance improvements (&&&^^^^5search_query5^^^^&&&). For U-Net, SNEMI^^^^5 OR all:\5^^^^D HD improves from $B_{s,o}(x)\;=\;(X * g_{s,o})(x).$^^^^5 OR all:\5^^^^ with CE to $B_{s,o}(x)\;=\;(X * g_{s,o})(x).$^^^^5 OR all:\5^^^^ with CWMI, a $B_{s,o}(x)\;=\;(X * g_{s,o})(x).$^^^^5 OR all:\5^^^^ absolute change or $B_{s,o}(x)\;=\;(X * g_{s,o})(x).$5. On DRIVE, U-Net HD improves from $B_{s,o}(x)\;=\;(X * g_{s,o})(x).$6 to $B_{s,o}(x)\;=\;(X * g_{s,o})(x).$7, a $B_{s,o}(x)\;=\;(X * g_{s,o})(x).$8 change. On GlaS, U-Net HD improves from $B_{s,o}(x)\;=\;(X * g_{s,o})(x).$9 to $\mathbf{I}$^^^^5search_query5^^^^, or $\mathbf{I}$^^^^5all:\5^^^^. The same paper reports improvements in mIoU, mDice, VI, and ARI across all four datasets.

SPW reports results on SNEMI^^^^5 OR all:\5^^^^D, GlaS, and DRIVE against ^^^^5all:\5all:\5^^^^ baselines: CE, BCE, Dice, Focal, Jaccard, Tversky, WCE, ABW, Skea-topo, RMI, and clDice (&&&^^^^5all:\5^^^^&&&). The reported mean $\mathbf{I}$^^^^5 OR all:\5^^^^ std scores are $\mathbf{I}$^^^^5 OR all:\5^^^^ mIoU and $\mathbf{I}$^^^^5 OR all:\5^^^^ mDice on SNEMI^^^^5 OR all:\5^^^^D, $\mathbf{I}$5 mIoU and $\mathbf{I}$6 mDice on GlaS, and $\mathbf{I}$7 mIoU and $\mathbf{I}$8 mDice on DRIVE. The paper does not report edge F^^^^5all:\5^^^^, Hausdorff, or clDice, but interprets lower VI and higher ARI on thin-structure datasets as evidence of better instance separation and boundary quality.

BCMFNet provides a distinct kind of evidence because its boundary loss is part of a real-time architecture. On Cityscapes, BCMFNet reports $\mathbf{I}$9 mIoU at $N$^^^^5search_query5^^^^ FPS, compared with the DDRNet-s baseline at $N$^^^^5all:\5^^^^ mIoU and $N$^^^^5 OR all:\5^^^^ FPS (&&&^^^^5 OR all:\5^^^^&&&). The ablation isolates the effect of the loss: BASE yields $N$^^^^5 OR all:\5^^^^ mIoU, adding LMFM gives $N$^^^^5 OR all:\5^^^^, and adding both LMFM and BCL gives $N$5. The paper interprets this as contextual improvement from low-resolution fusion plus recovery of boundary quality through BCL.

cbDice reports performance across ^^^^5 OR all:\5^^^^D and ^^^^5 OR all:\5^^^^D vascular segmentation tasks (&&&^^^^5 OR all:\5^^^^&&&). On DRIVE, peak Dice is $N$6, clDice $N$7, and NSD $N$8 at $N$9. On TopCoW ^^^^5 OR all:\5search_query5 OR all:\5 OR all:\5^^^^, nnU-Net with cbDice and $K$^^^^5search_query5^^^^ reports Dice(S) $K$^^^^5all:\5^^^^ versus clDice $K$^^^^5 OR all:\5^^^^, with NSD(S) $K$^^^^5 OR all:\5^^^^. NexToU with cbDice reaches Dice(S) $K$^^^^5 OR all:\5^^^^ and NSD(S) $K$5 for $K$6, or NSD(S) $K$7 with reduced Betti errors for $K$8. The paper emphasizes gains on small communicating arteries, which are precisely the type of fine structure that multi-scale boundary losses aim to protect.

For PINNs, the evidence is not segmentation accuracy but boundary-condition enforcement under multi-magnitude terms. On a heat equation with $K$9, MMPINN-DNN reports relative $$\mathbf{I} \;\longrightarrow\; \Big\{ \mathbf{I}_{H_0}\in\mathbb{R}^{H_0\times W_0},\; \mathbf{I}_{B_1}\in\mathbb{C}^{K\times H_0\times W_0},\; \mathbf{I}_{B_2}\in\mathbb{C}^{K\times H_1\times W_1},\;\dots,\; \mathbf{I}_{B_N}\in\mathbb{C}^{K\times H_{N-1}\times W_{N-1}},\; \mathbf{I}_{L_0}\in\mathbb{R}^{H_{N-1}\times W_{N-1}} \Big\}.$$^^^^5search_query5^^^^ error $$\mathbf{I} \;\longrightarrow\; \Big\{ \mathbf{I}_{H_0}\in\mathbb{R}^{H_0\times W_0},\; \mathbf{I}_{B_1}\in\mathbb{C}^{K\times H_0\times W_0},\; \mathbf{I}_{B_2}\in\mathbb{C}^{K\times H_1\times W_1},\;\dots,\; \mathbf{I}_{B_N}\in\mathbb{C}^{K\times H_{N-1}\times W_{N-1}},\; \mathbf{I}_{L_0}\in\mathbb{R}^{H_{N-1}\times W_{N-1}} \Big\}.$$^^^^5all:\5^^^^, compared with SA-PINN at $$\mathbf{I} \;\longrightarrow\; \Big\{ \mathbf{I}_{H_0}\in\mathbb{R}^{H_0\times W_0},\; \mathbf{I}_{B_1}\in\mathbb{C}^{K\times H_0\times W_0},\; \mathbf{I}_{B_2}\in\mathbb{C}^{K\times H_1\times W_1},\;\dots,\; \mathbf{I}_{B_N}\in\mathbb{C}^{K\times H_{N-1}\times W_{N-1}},\; \mathbf{I}_{L_0}\in\mathbb{R}^{H_{N-1}\times W_{N-1}} \Big\}.$$^^^^5 OR all:\5^^^^ and conventional PINN at approximately $$\mathbf{I} \;\longrightarrow\; \Big\{ \mathbf{I}_{H_0}\in\mathbb{R}^{H_0\times W_0},\; \mathbf{I}_{B_1}\in\mathbb{C}^{K\times H_0\times W_0},\; \mathbf{I}_{B_2}\in\mathbb{C}^{K\times H_1\times W_1},\;\dots,\; \mathbf{I}_{B_N}\in\mathbb{C}^{K\times H_{N-1}\times W_{N-1}},\; \mathbf{I}_{L_0}\in\mathbb{R}^{H_{N-1}\times W_{N-1}} \Big\}.$$^^^^5 OR all:\5^^^^ (&&&^^^^5 OR all:\5^^^^&&&). On ^^^^5all:\5^^^^D multi-frequency Poisson, MMPINN-MFF reaches $$\mathbf{I} \;\longrightarrow\; \Big\{ \mathbf{I}_{H_0}\in\mathbb{R}^{H_0\times W_0},\; \mathbf{I}_{B_1}\in\mathbb{C}^{K\times H_0\times W_0},\; \mathbf{I}_{B_2}\in\mathbb{C}^{K\times H_1\times W_1},\;\dots,\; \mathbf{I}_{B_N}\in\mathbb{C}^{K\times H_{N-1}\times W_{N-1}},\; \mathbf{I}_{L_0}\in\mathbb{R}^{H_{N-1}\times W_{N-1}} \Big\}.$$^^^^5 OR all:\5^^^^ versus MFF at approximately $$\mathbf{I} \;\longrightarrow\; \Big\{ \mathbf{I}_{H_0}\in\mathbb{R}^{H_0\times W_0},\; \mathbf{I}_{B_1}\in\mathbb{C}^{K\times H_0\times W_0},\; \mathbf{I}_{B_2}\in\mathbb{C}^{K\times H_1\times W_1},\;\dots,\; \mathbf{I}_{B_N}\in\mathbb{C}^{K\times H_{N-1}\times W_{N-1}},\; \mathbf{I}_{L_0}\in\mathbb{R}^{H_{N-1}\times W_{N-1}} \Big\}.$$5. The paper states that without power regularization, boundary losses can increase during training, whereas with regularization boundary and residual losses decrease together.

Representative reported results

Method	Setting	Reported result
CWMI	SNEMI^^^^5 OR all:\5^^^^D, U-Net	HD $$\mathbf{I} \;\longrightarrow\; \Big\{ \mathbf{I}_{H_0}\in\mathbb{R}^{H_0\times W_0},\; \mathbf{I}_{B_1}\in\mathbb{C}^{K\times H_0\times W_0},\; \mathbf{I}_{B_2}\in\mathbb{C}^{K\times H_1\times W_1},\;\dots,\; \mathbf{I}_{B_N}\in\mathbb{C}^{K\times H_{N-1}\times W_{N-1}},\; \mathbf{I}_{L_0}\in\mathbb{R}^{H_{N-1}\times W_{N-1}} \Big\}.$$6; mIoU $$\mathbf{I} \;\longrightarrow\; \Big\{ \mathbf{I}_{H_0}\in\mathbb{R}^{H_0\times W_0},\; \mathbf{I}_{B_1}\in\mathbb{C}^{K\times H_0\times W_0},\; \mathbf{I}_{B_2}\in\mathbb{C}^{K\times H_1\times W_1},\;\dots,\; \mathbf{I}_{B_N}\in\mathbb{C}^{K\times H_{N-1}\times W_{N-1}},\; \mathbf{I}_{L_0}\in\mathbb{R}^{H_{N-1}\times W_{N-1}} \Big\}.$$7 (&&&^^^^5search_query5^^^^&&&)
CWMI	DRIVE, U-Net	HD $$\mathbf{I} \;\longrightarrow\; \Big\{ \mathbf{I}_{H_0}\in\mathbb{R}^{H_0\times W_0},\; \mathbf{I}_{B_1}\in\mathbb{C}^{K\times H_0\times W_0},\; \mathbf{I}_{B_2}\in\mathbb{C}^{K\times H_1\times W_1},\;\dots,\; \mathbf{I}_{B_N}\in\mathbb{C}^{K\times H_{N-1}\times W_{N-1}},\; \mathbf{I}_{L_0}\in\mathbb{R}^{H_{N-1}\times W_{N-1}} \Big\}.$$8; mDice $$\mathbf{I} \;\longrightarrow\; \Big\{ \mathbf{I}_{H_0}\in\mathbb{R}^{H_0\times W_0},\; \mathbf{I}_{B_1}\in\mathbb{C}^{K\times H_0\times W_0},\; \mathbf{I}_{B_2}\in\mathbb{C}^{K\times H_1\times W_1},\;\dots,\; \mathbf{I}_{B_N}\in\mathbb{C}^{K\times H_{N-1}\times W_{N-1}},\; \mathbf{I}_{L_0}\in\mathbb{R}^{H_{N-1}\times W_{N-1}} \Big\}.$$9 (&&&^^^^5search_query5^^^^&&&)
SPW	GlaS	mIoU PRESERVED_PLACEHOLDER_^^^^5all:\5search_query5search_query5^^^^; mDice PRESERVED_PLACEHOLDER_^^^^5all:\5search_query5all:\5^^^^ (&&&^^^^5all:\5^^^^&&&)
BCMFNet + BCL	Cityscapes	PRESERVED_PLACEHOLDER_^^^^5all:\5search_query5 OR all:\5^^^^ mIoU at PRESERVED_PLACEHOLDER_^^^^5all:\5search_query5 OR all:\5^^^^ FPS (&&&^^^^5 OR all:\5^^^^&&&)
cbDice	TopCoW, nnU-Net	Dice(S) PRESERVED_PLACEHOLDER_^^^^5all:\5search_query5 OR all:\5^^^^; NSD(S) PRESERVED_PLACEHOLDER_^^^^5all:\5search_query55^^^^ (&&&^^^^5 OR all:\5^^^^&&&)

These results support a consistent interpretation: when the target structures are thin, elongated, densely packed, or topologically fragile, multi-scale boundary losses tend to improve not only contour quality but also global topology-oriented metrics such as VI, ARI, NSD, or Betti-derived scores.

6. Computational properties, implementation, and reproducibility

A recurrent design goal is to increase boundary sensitivity without introducing prohibitive overhead.

CWMI reports explicit complexity estimates. For image size PRESERVED_PLACEHOLDER_^^^^5all:\5search_query5^^^^6 and PRESERVED_PLACEHOLDER_^^^^5all:\5search_query5^^^^7 orientations, the first-scale cost combines FFT, filtering, inverse FFT, and MI terms; summed over pyramid levels, the overall complexity is

PRESERVED_PLACEHOLDER_^^^^5all:\5search_query5^^^^8

With PRESERVED_PLACEHOLDER_^^^^5all:\5search_query5^^^^9, the paper states that complexity scales effectively as PRESERVED_PLACEHOLDER_^^^^5all:\5all:\5search_query5^^^^. On SNEMI^^^^5 OR all:\5^^^^D/U-Net, the measured per-epoch training time is PRESERVED_PLACEHOLDER_^^^^5all:\5all:\5all:\5^^^^ s for CE and PRESERVED_PLACEHOLDER_^^^^5all:\5all:\5 OR all:\5^^^^ s for CWMI, a PRESERVED_PLACEHOLDER_^^^^5all:\5all:\5 OR all:\5^^^^ s overhead (&&&^^^^5search_query5^^^^&&&). Implementation uses PyTorch, FFT-based filtering, and real-valued representations for complex tensors.

SPW reports similar practical efficiency. For an PRESERVED_PLACEHOLDER_^^^^5all:\5all:\5 OR all:\5^^^^ input and PRESERVED_PLACEHOLDER_^^^^5all:\5all:\55^^^^ orientations, decomposition plus envelope computation and upsampling has complexity PRESERVED_PLACEHOLDER_^^^^5all:\5all:\5^^^^6, reducing to PRESERVED_PLACEHOLDER_^^^^5all:\5all:\5^^^^7 in practice for PRESERVED_PLACEHOLDER_^^^^5all:\5all:\5^^^^8 (&&&^^^^5all:\5^^^^&&&). On SNEMI^^^^5 OR all:\5^^^^D with U-Net, CE takes PRESERVED_PLACEHOLDER_^^^^5all:\5all:\5^^^^9 s per epoch and SPW takes PRESERVED_PLACEHOLDER_^^^^5all:\5 OR all:\5search_query5^^^^ s, a PRESERVED_PLACEHOLDER_^^^^5all:\5 OR all:\5all:\5^^^^ s overhead. The paper notes that ABW, clDice, and Skea-topo are slower in the reported setup. Weight computation is performed under torch.no_grad(), and gradients flow only through the base loss, not through the weight map.

The recommended hyperparameters in SPW are PRESERVED_PLACEHOLDER_^^^^5all:\5 OR all:\5 OR all:\5^^^^ scales, PRESERVED_PLACEHOLDER_^^^^5all:\5 OR all:\5 OR all:\5^^^^ orientations, PRESERVED_PLACEHOLDER_^^^^5all:\5 OR all:\5 OR all:\5^^^^, and PRESERVED_PLACEHOLDER_^^^^5all:\5 OR all:\55^^^^, with standard class weights PRESERVED_PLACEHOLDER_^^^^5all:\5 OR all:\5^^^^6 if needed (&&&^^^^5all:\5^^^^&&&). CWMI likewise reports that PRESERVED_PLACEHOLDER_^^^^5all:\5 OR all:\5^^^^7 and PRESERVED_PLACEHOLDER_^^^^5all:\5 OR all:\5^^^^8 gave the best trade-offs in ablation, and uses PRESERVED_PLACEHOLDER_^^^^5all:\5 OR all:\5^^^^9 when combining CWMI with CE (&&&^^^^5search_query5^^^^&&&).

BCMFNet’s computational story is different: BCL is designed to add negligible cost at training time and zero cost at inference. The heavy lifting remains in low-resolution multi-scale fusion, while the loss adds only label shifts, per-pixel cross-entropy maps, non-maximum suppression, and masking (&&&^^^^5 OR all:\5^^^^&&&). This explains why the model can preserve real-time speed.

cbDice requires differentiable skeletonization and Euclidean distance transforms per forward pass. The paper uses PyTorch ^^^^5 OR all:\5^^^^.^^^^5all:\5^^^^ on nnU-Net V^^^^5 OR all:\5^^^^, differentiable skeletonization from Menten et al. (ICCV ^^^^5 OR all:\5search_query5 OR all:\5 OR all:\5^^^^), and cuCIM GPU EDT (&&&^^^^5 OR all:\5^^^^&&&). It characterizes the extra cost as modest and comparable to other auxiliary-map losses. Stability is supported by clipping distances to PRESERVED_PLACEHOLDER_^^^^5all:\5 OR all:\5search_query5^^^^, normalization by PRESERVED_PLACEHOLDER_^^^^5all:\5 OR all:\5all:\5^^^^ and PRESERVED_PLACEHOLDER_^^^^5all:\5 OR all:\5 OR all:\5^^^^, and PRESERVED_PLACEHOLDER_^^^^5all:\5 OR all:\5 OR all:\5^^^^-stabilization for inverse-radius terms.

PINN power-regularization has negligible loss-side overhead because it adds only fractional powers and summations, though grouping residuals across subdomains introduces some bookkeeping (&&&^^^^5 OR all:\5^^^^&&&). The dominant additional cost in that setting comes from architecture choices such as MFF or INN rather than from the loss reconstruction itself.

The reproducibility patterns across these works are notable. CWMI and SPW both use Adam, StepLR with initial learning rate PRESERVED_PLACEHOLDER_^^^^5all:\5 OR all:\5 OR all:\5^^^^, decay PRESERVED_PLACEHOLDER_^^^^5all:\5 OR all:\55^^^^ every ^^^^5all:\5search_query5^^^^ epochs, 5^^^^5search_query5^^^^ epochs, batch size ^^^^5all:\5search_query5^^^^, and early stopping on mIoU (&&&^^^^5search_query5^^^^&&&, &&&^^^^5all:\5^^^^&&&). This suggests that the comparison between direct spectral alignment and spectral weighting was conducted under closely matched training settings.

7. Limitations, trade-offs, and relations to adjacent loss families

Despite their shared motivation, multi-scale boundary losses do not solve the same problem in the same way, and each carries specific limitations.

CWMI depends on covariance-based mutual-information estimation. The paper states that poorly conditioned covariances or extreme class imbalance can affect stability, that many subbands increase memory footprint, and that the method depends on filter design (&&&^^^^5search_query5^^^^&&&). Multi-class and ^^^^5 OR all:\5^^^^D extensions are described as feasible but not validated in that study. The recommendation to add small diagonal regularization to covariance inverses is presented as practical guidance rather than as part of the reported experiments.

SPW is validated on ^^^^5 OR all:\5^^^^D biomedical datasets only and is untested for ^^^^5 OR all:\5^^^^D volumetric segmentation in the paper (&&&^^^^5all:\5^^^^&&&). Because it up-weights high-frequency content, textures unrelated to true boundaries may also be emphasized if they are prominent. The paper notes that PRESERVED_PLACEHOLDER_^^^^5all:\5 OR all:\5^^^^6 and PRESERVED_PLACEHOLDER_^^^^5all:\5 OR all:\5^^^^7 need tuning to avoid over-emphasizing noise, and that SPW augments rather than replaces class weighting.

BCL depends on the choice of step sizes and thresholds. If the shift is too small, true boundary errors may be missed; if it is too large, the selected mask can drift into interiors (&&&^^^^5 OR all:\5^^^^&&&). The paper also notes that extremely thin structures may still be smoothed by low-resolution fusion and that BCL may not fully recover them without some high-resolution processing.

cbDice depends on skeleton quality and uses EDT as a fixed transform per forward pass, so gradients do not propagate through distances themselves (&&&^^^^5 OR all:\5^^^^&&&). Extremely noisy ultra-thin structures may be over-emphasized by inverse-radius weighting, and increasing the boundary/scale weight PRESERVED_PLACEHOLDER_^^^^5all:\5 OR all:\5^^^^8 too much may slightly reduce raw Dice on large vessels. Its formulation is especially natural for tubular or curvilinear structures; this suggests that its transfer to non-tubular objects may require careful reinterpretation of skeletons and local thickness.

In PINNs, the principal limitation is sensitivity to exponent choice. Excessively small powers can flatten gradients and slow or stall learning, and the paper explicitly advises against very large exponents such as PRESERVED_PLACEHOLDER_^^^^5all:\5 OR all:\5^^^^9 (&&&^^^^5 OR all:\5^^^^&&&). Moreover, loss balancing alone cannot compensate for inadequate representational capacity in high-frequency boundary regimes; MFF or INN may still be required.

Relative to adjacent loss families, the papers draw several distinctions. CWMI contrasts itself with boundary-aware geometric losses such as Boundary Loss with SDF, Hausdorff distance loss, and clDice, characterizing them as computationally heavy or specialized, whereas CWMI is frequency-domain, orientation-selective, multi-scale, and robust to small misalignments and noise (&&&^^^^5search_query5^^^^&&&). SPW distinguishes itself from distance-transform weighting by emphasizing prediction adaptivity and FFT efficiency (&&&^^^^5all:\5^^^^&&&). cbDice positions itself as matching the translation sensitivity of B-DoU while preserving the topological advantages of clDice (&&&^^^^5 OR all:\5^^^^&&&). BCMFNet places BCL against contour branches, post-processing methods, and explicit edge optimization, emphasizing its simplicity and inference-time neutrality (&&&^^^^5 OR all:\5^^^^&&&).

Taken together, these comparisons indicate that the field has moved away from a narrow opposition between region losses and boundary losses. Current multi-scale boundary losses increasingly combine topology, frequency structure, distance geometry, and adaptive weighting, and they do so in ways tailored to specific object classes and deployment constraints. A plausible implication is that future formulations will continue to hybridize these ingredients rather than converge on a single canonical objective.