Topology-aware loss for segmentation in computed tomography images

CrackDualMamba: A lightweight dual-stream Mamba with novel focal dice balanced loss for vehicle-based road crack segmentation

Thin crack segmentation in building facades is particularly challenging under limited data conditions due to the extremely narrow geometry and weak contrast of crack structures. Conventional overlap-based loss functions, such as Binary Cross-Entropy (BCE) and Dice loss, optimize pixel-wise agreement but do not explicitly account for spatial boundary relevance, often resulting in fragmented predictions and geometric misalignment. This study introduces a lightweight distance-aware weighting extension of Dice loss designed to improve boundary alignment without modifying the network architecture. The proposed approach integrates Euclidean distance information derived from ground-truth masks to assign higher importance to prediction errors near crack boundaries while reducing the influence of distant background regions. The method is evaluated on a real-world dataset of 108 facade images (87 training and 21 validation) using a standard U-Net architecture under identical training conditions. Experimental results demonstrate a consistent reduction in geometric boundary error. The 95th percentile Hausdorff Distance (HD95) decreases from 230.08 px with BCE and 217.55 px with Dice loss to 148.28 px with the proposed distance-aware formulation, corresponding to reductions of approximately 81.8 px and 69.3 px, respectively. In addition, the proposed loss improves overlap-based metrics, achieving IoU@0.1 = 0.2844 and Dice = 0.4087 on the validation set. These results indicate that incorporating spatial distance information into the optimization objective improves geometric alignment and structural continuity of thin crack predictions. The findings suggest that integrating lightweight distance-aware weighting into conventional loss formulations can improve segmentation quality for thin structures in constrained-data scenarios while maintaining computational simplicity.

Fuzzy Attention SegNet Model with Serval Ablation Optimization and Log-Cosh Softmax Loss for Effective Brain Tumor Segmentation Using MRI Imaging

• Proposed a novel Multi-Stage Ensemble Deep Learning Framework for Crack Segmentation and Feature-Based Power Loss Projection (MSC-FPL) using EL images. • Developed an Ensemble-based Binary Classification (EBC) CNN model to classify EL images of PV cells into two classes (Normal and Defective) in the first stage of the proposed MSC-FPL. • Developed an Ensemble-based Multi-Class Semantic Segmentation (EMSS) CNN model in the second stage of the proposed MSC-FPL to perform pixel-level classification of the defective class obtained from the first stage into six classes: Cross, Diagonal cracks relative to the busbar, Parallel cracks relative to the busbar, Perpendicular cracks relative to the busbar, Multiple Direction cracks, and Busbars as a key feature of the PV cell. • Introduced a novel feature-based approach that integrates cell-level crack features, specifically orientation and size, into a structured power loss projection framework. • Quantitative and qualitative results were presented for each stage to validate the proposed framework. Crack detection in Photovoltaic (PV) cells using Electroluminescence (EL) imaging has become a primary research focus due to cracks being one of the most common defects that cause power loss from PV modules. Moreover, the extent of power output loss differs based on the size and orientation of the cracks. Existing deep learning approaches exhibit low predictive performance and offer limited analysis of segmented crack features. To the best of our knowledge, no previous deep learning–based study has considered the correlation between cell-level crack features and feature-based power loss projection using quantitative measurements. Therefore, we proposed a novel Multi-Stage Ensemble Deep Learning Framework for Crack Segmentation and Feature-Based Power Loss Projection in PV Cells (MSC-FPL) to address the aforementioned limitations. In Stage 1, we developed an Ensemble-based Binary Classification (EBC) CNN model that achieved an accuracy of 98.19%. The output of Stage 1 is used as input for Stage 2, where an Ensemble-based Multi-Class Semantic Segmentation (EMSS) CNN model achieved a Mean Intersection over Union (mIoU) of 85.01%. The subsequent stages utilize the output from Stage 2, which provides crack segmentation based on orientation. In Stage 3, crack size features are calculated, and in Stage 4, these features are integrated within a feature-based power loss projection framework using a novel interaction-based formulation. Our results confirm the effectiveness of the proposed framework through comparison with existing approaches, improve PV module quality assessment, and provide an intelligent support system for PV module monitoring, contributing to the sustainability of solar energy systems.

Segmentation-aware Generative Reinforcement Network (GRN) for tissue layer segmentation in 3-D ultrasound images for chronic low-back pain (cLBP) assessment.

A novel method for tackling the problem of imbalanced data in medical image segmentation is proposed in this work. In Balanced Cross Entropy (BCE) loss, which is a type of weighted CE loss, the weight assigned to each class is the inverse of the class frequency. These balancing weights are expected to equalize the effect of each class on the overall loss and prevent the model from being biased towards the majority class. But, as it has been shown in previous studies, this method degrades the performance by a large margin. Therefore, Balanced CE is not a popular loss in medical segmentation tasks, and usually a region-based loss, like the Dice loss, is used to address the class imbalance problem. Existing weighting CE loss approaches, like Balanced CE, are usually adapted from classification tasks. In this study we propose a weighting method which is designed for segmentation and takes advantage of spatial and geometrical properties of the labels. In the proposed method, the weighting of cross entropy loss for each class is based on a dilated area of each class mask, and balancing weights are assigned to each class together with its surrounding pixels. The goal of this study is to show that the performance of Balanced CE loss can be greatly improved my modifying its weighting strategy. Experiments on different datasets show that the proposed Dilated Balanced CE (DBCE) loss outperforms the Balanced CE loss by a large margin and produces superior results compared to CE loss, and its performance is similar or better compared to the performance of the combination of Dice and CE loss. Also, the performance of different combination losses can be improved by replacing the CE loss with the proposed DBCE loss. This study shows that a weighted cross entropy loss with the right weighing strategy can be as effective as a region-based loss in handling the class imbalance problem.

ABSTRACT Accurate segmentation of skin lesions is crucial for dependable computer‐aided diagnosis of melanoma. However, many existing deep learning models still have difficulty dealing with vague lesion borders, uneven appearance, and unstable performance when applied to new datasets. This paper proposes a dual‐scale boundary‐aware network (U 2 ‐DBA) for dermoscopic image segmentation. The model includes a nested U‐in‐U encoder that captures both local and global features, a dual‐branch gating module that balances semantic and structural information, and a decoder that focuses on preserving boundary details. We further propose a novel Feature‐Boundary‐Skeleton (FBS) loss function, which integrates region overlap, edge gradient, and skeleton‐level shape constraints to enhance segmentation accuracy and structural consistency. To evaluate model efficiency, we introduce the Smooth Accuracy‐Compactness Score (SACS), combining Dice and IoU metrics with a logarithmic penalty on model size. Experiments conducted on the ISIC 2018 dataset demonstrate that U 2 ‐DBA achieves high performance (Dice = 0.884, IoU = 0.799) and outperforms six state‐of‐the‐art models in SACS. When directly evaluated on PH2 and HAM10000 without fine‐tuning, the model retains strong performance. These findings indicate that U 2 ‐DBA is not only accurate and compact but also generalizes effectively across diverse datasets, offering a practical and deployable solution for clinical dermoscopic lesion segmentation. The code is available at https://github.com/kid‐od/U2‐DBA .

Fuzzy rough set loss for deep learning-based precise medical image segmentation.

The availability of large, high-quality annotated datasets in the medical domain poses a substantial challenge in segmentation tasks. To mitigate the reliance on annotated training data, self-supervised pre-training strategies have emerged, particularly employing contrastive learning methods on dense pixel-level representations. In this work, we proposed to capitalize on intrinsic anatomical similarities within medical image data and develop a semantic segmentation framework through a self-supervised fusion network, where the availability of annotated volumes is limited. In a unified training phase, we combine segmentation loss with contrastive loss, enhancing the distinction between significant anatomical regions that adhere to the available annotations. To further improve the segmentation performance, we introduce an efficient parallel transformer module that leverages Multiview multiscale feature fusion and depth-wise features. The proposed transformer architecture, based on multiple encoders, is trained in a self-supervised manner using contrastive loss. Initially, the transformer is trained using an unlabeled dataset. We then fine-tune one encoder using data from the first stage and another encoder using a small set of annotated segmentation masks. These encoder features are subsequently concatenated for the purpose of brain tumor segmentation. The multiencoder-based transformer model yields significantly better outcomes across three medical image segmentation tasks. We validated our proposed solution by fusing images across diverse medical image segmentation challenge datasets, demonstrating its efficacy by outperforming state-of-the-art methodologies.

Objective.Ultrasound shear wave imaging enables noninvasive, quantitative assessment of tissue pathology with mechanical elasticity measurements. However, shear wave elastography (SWE) reconstructions are challenged by noise sensitivity, inefficient multi-push strategies for scalable region of interest coverage, and limited annotated data, leading to suboptimal reconstruction and unreliable inclusion segmentation.Approach.In this work, we present a novel two-stage deep learning framework that addresses these limitations through a convolutional neural network (CNN)-based multi-nested-LSTM reconstruction network followed by a compound-loss-driven CNN-denoiser. The reconstruction stage begins with a ResNet3D-encoder that extracts spatiotemporal features from sequential multi-push acoustic radiation force data. These features are temporally windowed with Nested CNN-LSTM, converted from 3D to 2D with temporal attention module, and enhanced by fast Fourier transform-based frequency attention. The resulting 2D maps are subsequently decoded into primary 2D elasticity reconstructions. To mitigate data-scarcity and improve generalization, a patch-based training regime is also proposed. The second stage introduces a dual-decoder denoising network that separately processes inclusion and background stiffness features, followed by a fusion module that produces a denoised modulus map and a segmentation mask. A multi-objective compound loss is designed to accommodate the denoising, fusing, and mask generation. The method is validated on sequential multi-push (simulation and experimental) SWE motion data with multiple overlapping regions.Results.The method was tested on simulated and CIRS phantom datasets with four overlapping push regions, yielding 26.33 dB peak-signal-to-noise-ratio (PSNR), 30.73 dB contrast-to-noise-ratio (CNR), and 0.813 intersection over union (IoU) in simulation, and 22.44 dB PSNR, 36.88 dB CNR, and 0.781 IoU experimentally. Evaluation on anex vivoswine liver confirmed elasticity estimates within reported biological stiffness ranges. Compared to DSWE-Net and spatio-temporal CNNs, our approach shows superior reconstruction, segmentation, and noise insensitivity.Significance.This framework provides a robust approach to SWE reconstruction and inclusion segmentation, demonstrating strong potential for clinical translation.

Pancreatic ductal adenocarcinoma (PDAC) presents significant segmentation challenges due to its heterogeneous structure and low contrast with surrounding tissues. Manual segmentation is time-consuming and inconsistent, while conventional deep learning models often lack clinical context and struggle with small tumor regions. The proposed CAAG-UNet3D is a residual 3D U-Net architecture that incorporates synthetic multi-modal imaging (T1, T1ce, T2, FLAIR) and condition-specific attention mechanisms. Residual convolution blocks and attention gates are integrated into the encoder-decoder structure to enhance feature learning and focus on tumor-specific regions. A hybrid loss function combining Focal Tversky Loss and Dice Loss is employed to handle class imbalance and improve segmentation of small tumor areas. The model is trained using the AdamW optimizer with OneCycleLR scheduling and evaluated on standard segmentation metrics. On a synthetic multi-modal CT dataset, CAAG-UNet3D achieved a mean Dice score of 0.832, IoU of 0.766, and VOE of 12%, outperforming benchmark models such as TransBTS, nnUNet, and 3D-ResUNet. Ablation studies confirmed the contribution of multi-modality, attention mechanisms, and residual learning to overall performance. CAAG-UNet3D demonstrates that the integration of synthetic multi-modal radiomic inputs and attention-guided residual learning yields superior automated PDAC segmentation, paving the way for more reliable quantitative imaging biomarkers in clinical radiomics.

Synthetic aperture radar (SAR) provides the high- resolution imaging capability and can obtain the ocean and land target information under complex observation conditions. Dual-frequency SAR images exhibit the complementary properties in the resolution and scattering characteristics, thus fusing dual- frequency SAR images can combine these advantages to provide the richer and more comprehensive information for downstream semantic segmentation tasks. However, existing fusion algorithms typically assume preregistered images and focus on the good visual effects while neglecting the semantic requirements of high-level vision tasks. To address this problem, a joint registration and fusion network for the dual-frequency SAR images based on the semantic information is proposed, which can achieve the high-precision cross-frequency registration and high-quality image fusion. First, in the image registration stage, the microwave vision data (MVD) semantic pseudo-color maps are introduced as the auxiliary information, and the semantic-consistency loss is used to guide the registration process, finally the fusion loss is fed back to the image registration as an optimization signal. Second, in the image fusion stage, the global spatial attention mechanism is used for the feature extraction and fusion, and then the semantic segmentation loss is incorporated to guide the image fusion so that the results can better conform to the high-level semantic structures. Experimental results demonstrate the effectiveness and correctness of the proposed method, which can outperform state-of-the-art approaches in the image registration accuracy, image fusion quality and downstream semantic segmentation performance.

Accurate terrain perception is essential for safe rover operations and reliable geotechnical interpretation of Martian surfaces. The heterogeneous scales, colors, and textures of Martian terrain present significant challenges for semantic segmentation. We present MarsTerrNet, a dual-backbone segmentation framework that combines Progressive Residual Blocks (PRB) with a Swin Transformer to jointly capture fine-grained local details and global contextual dependencies. To further enhance discrimination among geologically correlated classes, we design a feature-guided loss that aligns representative features across terrain categories and reduces confusion between visually similar but physically distinct types. For comprehensive evaluation, we establish MarsTerr2024, an extended dataset derived from the Curiosity rover, providing diverse geological scenes for terrain understanding. Experimental results show that MarsTerrNet achieves state-of-the-art performance and produces geologically consistent segmentation results, supporting automated mapping and geotechnical assessment for future Mars exploration missions.

ABSTRACT In medical image diagnosis, in order to address the problem of limited diagnostic performance caused by insufficient cross‐domain feature adaptation and poor interpretability of hidden features, this paper constructs a Mask Region‐based Convolutional Neural Network (Mask R‐CNN) framework based on transfer learning. The gradient reversal layer (GRL) is used to achieve region‐adaptive feature alignment, and the attention mechanism is combined with feature deconvolution to achieve precise lesion segmentation and clinically understandable pathological feature expression. The gradient reversal layer is embedded in the backbone network of Mask R‐CNN, and the feature distribution difference between the source domain and the target domain is minimized through adversarial training. The channel‐spatial attention module is applied into the feature pyramid network (FPN), and features of different scales are dynamically weighted to enhance the significant expression of the lesion area. A deconvolution network is cascaded after the segmentation head to map the high‐dimensional hidden features into a visual heat map, and the key pathological areas are located through gradient‐weighted class activation mapping (Grad‐CAM) to assist clinical decision‐making. A weighted combination of domain classification loss, segmentation loss, and feature reconstruction loss is designed to balance the accuracy requirements of cross‐domain adaptation and lesion segmentation. Experimental results show that the Dice coefficient of the Mask R‐CNN in this paper is between 0.88 and 0.95 for five tumor types, and the Intersection over Union (IoU) is between 0.78 and 0.9, which has a high lesion segmentation accuracy and improves the diagnostic performance. The GRL‐embedded Mask R‐CNN exhibits a unique two‐stage convergence characteristic. The Maximum Mean Discrepancy (MMD) distance in the first 20 iterations decreases sharply (0.423 → 0.345) and enters the fine‐tuning stage (0.345 → 0.337) in the later stage, with significant cross‐domain adaptation effect. The Spearman rank correlation coefficient of the model in this paper is above 0.8 in the five pathological features, which verifies the breakthrough ability of this paper in hidden feature extraction.

In recent years, the semantic segmentation of multimodal remote-sensing images using convolutional methods has received significant attention. Owing to the localized nature of convolutional operations, existing methods use the attention method to obtain global relations. However, effectively complementing and eliminating the global and local relations between two modalities has become a major challenge. In addition, category imbalance often affects the model performance as a result of the high resolution of remote-sensing images. In this study, we propose an asymmetric dual encoder network with clustering mutual contrast loss. Specifically, we use a convolutional neural network and Transformer as the backbone to extract two modal features in parallel to learn the local and global information, respectively. Next, our proposed multimodal hierarchical interaction module and dynamic weight inspired block efficiently fuse the multimodal features to complement the local and global information. The fused features are fed into our proposed local context extraction module and global context extraction module. Furthermore, to address the challenge presented by feature class imbalances, we apply a clustering algorithm to classify each pixel, which is subsequently inter-supervised using the inter-contrast loss. Extensive experiments on benchmark datasets show that the proposed model is extremely effective in the semantic segmentation of remotely sensed images and that it outperforms current state-of-the-art networks both quantitatively and qualitatively. Our code and results can be found at https://github.com/LYZ00918/AMCNet.

CRISP: A modular platform for cryo-EM image segmentation and processing with Conditional Random Field.

Background/Objectives: Spinal MRI segmentation has become increasingly important with the prevalence of disc herniation and vertebral injuries. Artificial intelligence can help orthopedic surgeons and radiologists automate the process of segmentation. Currently, there are few tools for T1-weighted spinal MRI segmentation, with most focusing on T2-weighted imaging. This paper focuses on creating an automatic lumbar spinal MRI segmentation tool for T1-weighted images using deep learning. Methods: An Attention U-Net was employed as the main algorithm because the architecture has shown success in other segmentation applications. Segmentation loss functions were compared, focusing on the difference between BCE and MSE loss. Two board-certified radiologists scored the output of the Attention U-Net versus four other algorithms to assess clinical relevance and segmentation accuracy. Results: The Attention U-Net achieved superior results, with SSIM and DICE coefficients of 0.998 and 0.93, outperforming other architectures. Both radiologists agreed that the Attention U-Net segmented lumbar spinal images with the highest accuracy on the Likert Scale (3.7 ± 0.82). Cohen's Kappa coefficient was measured at 0.31, indicating a fair level of agreement. MSE loss outperformed BCE with respect to both SSIM and DICE, serving as the loss function of choice. Conclusions: Qualitative observations showed that the Attention U-Net and U-Net++ were the top performing networks. However, the Attention U-Net minimized external noise and focused on internal spinal preservation, demonstrating strong segmentation performance for T1-weighted lumbar spinal MRI.

Accurate segmentation in medical imaging requires loss functions that capture both regional overlap and boundary alignment. This study evaluates composite losses combining binary cross-entropy (BCE) and a boundary-based term under fixed and adaptive weighting schemes, using U-Net and SwinUNETR on the MM-WHS dataset. For U-Net, a small boundary contribution with adaptive weighting yielded the best results: Standard SoftAdapt (90/10 BCE + BoundaryDoU) achieved the highest Dice score (0.932 pm 0.01), surpassing both the baseline (0.923 pm 0.01) and fixed ratios. In contrast, SwinUNETR achieved its strongest performance with a fixed 70% BCE + 10% boundary ratio (0.919 ± 0.02). The result showed that combining a boundary-based loss term helps improve the segmentation accuracy. However, the performance gain is dependent on the architecture of the segmentation model; convolution-based U-Net benefited from the adaptive loss weighting scheme, whereas Transformer-based SwinUNETR without strong inductive bias did not benefit from increased influence of the boundary loss term.

Optical coherence tomography (OCT) is an essential tool for diagnosing retinal diseases because of its high-resolution, three-dimensional structural and functional imaging of the retina. Automatic segmentation and quantification of the retinal biomarkers provide clinicians with reliable diagnostic references and improve the accuracy and efficiency of diagnosis. However, the diverse lesions, artifacts, and missing normal retinal structures in the OCT images of patients with macular edema severely affect the accuracy of the segmentation model. Moreover, most deep learning segmentation models require a considerable amount of annotated data, which increases the development cost of medical image segmentation models. To address these issues, we propose a structural prior-guided and feature-enhanced transformer with masked imaging modeling pretraining (SPFET-MIMP) to segment the retinal layers and fluid in macular edema OCT B-scans. The segmentation network employs a transformer architecture combining shifted-windowing multi-head self-attention and axial attention to enhance the extraction of contextual information and multiscale features. To focus on the physiological order of the retinal layers and their positional relationships with fluid, a customized multi-class synergistic segmentation (MCSS) loss is incorporated into the loss function. The loss value reflects the prior knowledge of relative positions and topological structures in the retina, which helps maintain the correct order and completeness of the retinal layers. We also utilize a self-supervised pretraining framework, SimMIM, to pretrain a segmentation model on a large-scale unlabeled OCT dataset to enhance the robustness of the model for images with low contrast or shadow artifacts. Our method achieved average Dice coefficients of 94.35% and 90.19% on the AROI dataset and a private diabetic macular edema dataset, respectively, both outperforming other state-of-the-art technologies.