emd-based consistency 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.

Automatic detection and segmentation of prostate cancer metastases in PSMA PET/CT images remains challenging due to small lesion under-segmentation, blurred boundaries in large lesions, and performance degradation caused by lesion diffusion. Method: We propose the L1-weighted Dice Focal Loss (L1DFL), the first voxel-wise adaptive weighting scheme based on the L1 norm, dynamically mitigating these issues. Our framework employs dual 3D architectures—Attention U-Net and SegResNet—fusing multimodal PET/CT input channels, augmented by an SUV-normalization-guided true-positive assignment strategy. Contribution/Results: Evaluated on 380 clinical cases, our method achieves a test-set F1 score improvement of ≥6% over standard Dice Loss and 34% over Dice Focal Loss. It significantly reduces false positives and enhances robustness across multiscale lesion segmentation, demonstrating superior generalizability and clinical applicability.

Traditional pixel-wise loss functions for cerebral artery segmentation in digital subtraction angiography (DSA) sequences neglect geometric and physical consistency, leading to fragmented boundaries and unstable predictions. To address this, we propose a physics-informed loss (PIL) that, for the first time, incorporates dislocation theory from materials physics into medical image segmentation—modeling elastic interactions along vascular boundaries to physically regularize contour evolution. PIL is architecture-agnostic and seamlessly integrates with mainstream segmentation frameworks—including U-Net, U-Net++, SegFormer, and MedFormer—jointly optimizing pixel-wise overlap and boundary dynamics. Evaluated on the DIAS and DSCA datasets, PIL significantly improves sensitivity, F1 score, and boundary coherence, consistently outperforming cross-entropy, Dice, and active contour losses. The implementation is publicly available.

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.

Longitudinal segmentation of multiple sclerosis (MS) lesions faces severe data and output imbalance challenges, particularly degrading performance on small lesions, boundary regions, and distance-sensitive metrics. To address this, we propose HyTver, a novel hybrid loss function that integrates an enhanced Dice loss with weighted cross-entropy and incorporates distance-aware regularization to jointly optimize segmentation accuracy and geometric consistency. HyTver requires no complex hyperparameter tuning and demonstrates superior training stability when applied to pre-trained models. Evaluated on the public MSLesion dataset, our method achieves a Dice score of 0.659 and significantly outperforms mainstream loss functions in distance-based metrics—including Hausdorff Distance (HD) and Average Symmetric Surface Distance (ASSD)—demonstrating its effectiveness in multi-objective optimization and robustness to lesion-scale variability and boundary ambiguity.

This work addresses the challenge of out-of-distribution generalization in geospatial data across regions, where distributional shifts hinder model transfer and existing approaches lack effective means to quantify domain discrepancies. To this end, we propose GeoSpOT, which for the first time integrates optimal transport theory with geographic coordinate encoding to construct a geospatial inter-domain distance metric using only latitude and longitude. Notably, GeoSpOT requires no downstream task data to quantify domain shift or predict model transfer performance. Experimental results demonstrate that the GeoSpOT distance accurately forecasts cross-regional generalization outcomes and effectively guides data selection and identification of high-risk regions.

Existing image registration methods rely on global smooth regularization, limiting their ability to model regionally heterogeneous deformations inherent in anatomical motion. To address this, we propose SegReg—a segmentation-driven medical image registration framework that leverages anatomical segmentation maps to guide region-adaptive deformation field estimation. SegReg decomposes global regularization into localized subregion optimizations and establishes, for the first time, an approximately linear relationship between registration accuracy and segmentation quality. Specifically, it employs a deep learning-based segmentation network to delineate anatomical subregions and computes local displacement fields in parallel within a unified registration backbone, subsequently fusing them into a globally consistent deformation field. Evaluated on multi-center cardiac, abdominal, and pulmonary datasets, SegReg achieves average improvements of 2–12% over state-of-the-art methods. With ground-truth segmentations, it attains a Dice score of 98.23% for critical structures, significantly enhancing anatomical consistency and registration robustness.

This study addresses the performance degradation in cardiac structure segmentation from multi-source partially labeled echocardiograms, caused by domain shift and diverse missing label patterns. It presents the first systematic comparison of three loss functions—adaptive categorical cross-entropy (aCCE), Boundary Loss, and adaptive binary cross-entropy (aBCE)—across complex partial labeling scenarios, including single-domain versus cross-domain settings and single-label versus multi-label missingness. Experimental results demonstrate that Boundary Loss achieves the best performance in cross-domain tasks with multi-label missingness, while both aBCE and Boundary Loss significantly outperform other methods under cross-domain conditions with single-label missingness. All approaches exhibit robust performance in single-domain tasks. This work provides empirical evidence and practical guidance for selecting loss functions in partially labeled medical image segmentation.