boundary-sensitive loss

Medical image segmentation models trained with region-based losses, such as Dice loss, commonly suffer from poor calibration and overconfident predictions, hindering their deployment in high-stakes clinical settings. This work is the first to uncover the root cause of this issue from a gradient perspective and introduces a novel “gradient vector field surgery” approach. By incorporating a scaling factor—linearly dependent on prediction error—into the partial derivatives of the loss function, the method enhances model calibration without altering the original loss structure. The proposed technique is highly generalizable and consistently improves calibration across diverse 2D and 3D medical segmentation tasks while preserving state-of-the-art segmentation accuracy.

This study addresses the problem of recovering unknown object boundaries from noisy, unlabeled images in an unsupervised and nonparametric setting. To this end, the authors propose a boundary detection method that integrates a continuous hinge-type surrogate loss with deep neural networks, embedded within a robust Gibbs posterior framework based on a thresholded misclassification loss. Theoretical analysis demonstrates that the resulting estimator achieves minimax optimal convergence rates—up to logarithmic factors—for piecewise smooth boundaries that may include corners and kinks, while also enjoying Fisher consistency and calibration properties. Extensive experiments confirm the method’s stability and superior performance across varying noise levels and complex boundary shapes, significantly outperforming existing unsupervised boundary detection approaches.

Existing weighted binary cross-entropy (WBCE) loss treats all non-edge pixels uniformly, ignoring their structural heterogeneity—leading to blurred edge predictions. To address this, we propose the Edge-Boundary-Texture (EBT) loss, the first to explicitly categorize pixels into three semantically distinct classes—edge, boundary, and texture—and assign class-specific supervision weights. EBT constitutes a structured generalization of WBCE, with theoretical proof confirming WBCE as its special case. Leveraging a differentiable three-class weighting scheme, EBT is framework-agnostic, requires no architecture modification, and supports consistent hyperparameters across models and datasets. Extensive experiments on multiple benchmarks demonstrate that EBT significantly improves edge localization accuracy and structural fidelity, outperforming state-of-the-art methods both quantitatively and visually. It exhibits strong robustness, cross-dataset generalizability, and plug-and-play practicality.

In semantic segmentation, slender structures and densely packed object boundaries often suffer from ambiguous delineation, while small objects are prone to misclassification or omission. Existing distance-transform-based weighted loss functions incur substantial computational overhead and lack prediction adaptability. To address these issues, we propose Multi-scale Adaptive Weighting Loss (MAW-Loss), the first loss formulation incorporating a guided pyramid into loss design. MAW-Loss employs prediction-driven frequency-domain decomposition and cross-scale weight mapping to generate end-to-end differentiable, dynamic weight maps in real time—eliminating explicit distance transforms. Evaluated on SNEMI3D, GlaS, and DRIVE benchmarks, MAW-Loss consistently outperforms 11 state-of-the-art loss functions, achieving significant gains in Dice score and pixel accuracy. Its computational overhead is negligible, and the implementation is publicly available.

Neonatal hypoxic-ischemic encephalopathy (HIE) lesion segmentation in MRI faces challenges including diffuse, multifocal lesions, large inter-lesion volume variability, and severe scarcity of annotated data. To address these, we develop an optimized 3D U-Net framework on the BONBID-HIE dataset and conduct the first systematic evaluation of six loss functions for HIE lesion segmentation. We propose two novel composite losses—Dice-Focal-HausdorffDT and Tversky-HausdorffDT—that jointly optimize regional similarity (via Dice/Tversky/Focal terms) and boundary geometric fidelity (via Hausdorff distance on signed distance transforms). Experimental results demonstrate that Tversky-HausdorffDT achieves the best performance under limited-label settings (Dice = 0.682, Normalized Surface Dice = 0.721), while Dice-Focal-HausdorffDT significantly reduces mean surface distance (1.39 mm). This work establishes a robust, reproducible loss-function design paradigm for HIE lesion segmentation, advancing methodological rigor in neonatal neuroimaging analysis.

This study addresses the extreme class imbalance in whiteboard stroke binary segmentation, where foreground pixels constitute only 1.79% on average, rendering conventional region-based metrics inadequate for evaluating fine-stroke performance. To this end, the authors propose a comprehensive evaluation protocol that integrates region- and boundary-based metrics, fairness analysis across core and fine-stroke subsets, and robustness statistics from multi-run training. Using a DeepLabV3-MobileNetV3 architecture, five loss functions are systematically compared. Results show that overlap-aware losses (e.g., Dice+Focal) improve F1 by over 20 points (0.663 vs. 0.438) compared to cross-entropy, while Tversky loss significantly outperforms the Sauvola method in worst-case F1 (0.565 vs. 0.452). Increasing training resolution further boosts F1 by up to 12.7 points. This work is the first to incorporate non-parametric significance testing and worst-case evaluation, uncovering hidden trade-offs among loss functions in fine-grained segmentation tasks.

This work addresses the challenge of segmenting curvilinear structures in medical images, which is often hindered by imaging artifacts and the difficulty of preserving topological connectivity. The authors propose Topology SegNet, a novel framework that, for the first time, integrates persistence images—a differentiable and learnable topological representation—directly into the U-Net backbone. By fusing topological information during both encoding and decoding stages, the method eliminates the need for handcrafted topological loss functions. Evaluated on three benchmark datasets for curvilinear structure segmentation, Topology SegNet achieves state-of-the-art performance in both pixel-level accuracy and topological fidelity, demonstrating significantly enhanced robustness and generalization under image degradations such as overexposure and blur.

To address segmentation challenges arising from blurred or discontinuous boundaries in noisy images, this paper proposes a hybrid segmentation framework integrating physical priors with deep learning. Methodologically, it introduces a dual-module architecture—F (frequency-domain preprocessing) and T (spatial-domain stabilization)—incorporating an edge detector and an average curvature regularizer, while enforcing a variational PDE constraint derived from a modified Cahn–Hilliard equation. The network adopts a U-Net–like architecture to enable end-to-end, interpretable training. The key contribution lies in the organic unification of frequency-domain analysis, geometric priors, and deep feature representation, achieving enhanced boundary sharpness and segmentation robustness without sacrificing model efficiency. Quantitative evaluation on three benchmark datasets demonstrates superior performance over state-of-the-art CNNs; visual quality matches that of Transformer-based methods, while computational overhead remains significantly lower.

Existing image segmentation evaluation metrics often fail to capture structural and topological consistency, frequently overestimating segmentation quality due to minor boundary errors or spurious holes. To address this limitation, this work introduces the concept of “Jordan-segmentable” masks by integrating digital Jordan theory with homological invariants, grounded in the Jordan curve theorem on digital planes. The proposed framework enables unsupervised assessment of topological plausibility by determining whether a mask partitions the image domain into exactly two connected components. By combining 4/8-connectivity analysis, Betti number computation, and homology theory, the method establishes a mathematically rigorous evaluation paradigm that effectively identifies high-quality segmentations preserving global shape and connectivity—particularly critical in applications such as medical imaging, where topological correctness is paramount.

This work addresses the limitations of traditional minimal path-based image segmentation methods, which suffer from insufficient accuracy in complex backgrounds and high sensitivity to initialization. To overcome these issues, the authors propose a geodesic framework–based mask proposal voting mechanism that generates diverse and reliable mask candidates through adaptive domain partitioning and constructs the final segmentation via a voting strategy incorporating prior knowledge. The approach innovatively integrates region-based min-cut evolution, geodesic distance computation, and mask scoring maps, effectively eliminating dependence on initial seed points. Experimental results demonstrate that the proposed method significantly outperforms existing minimal path segmentation techniques across multiple challenging scenarios, achieving state-of-the-art performance in both robustness and accuracy.