Accurate fetal brain tissue segmentation and biometric measurement are essential for monitoring neurodevelopment and detecting abnormalities in utero. The Fetal Tissue Annotation (FeTA) Challenges have established robust multi-center benchmarks for evaluating state-of-the-art segmentation methods. This paper presents the results of the 2024 challenge edition, which introduced three key innovations. First, we introduced a topology-aware metric based on the Euler characteristic difference (ED) to overcome the performance plateau observed with traditional metrics like Dice or Hausdorff distance (HD), as the performance of the best models in segmentation surpassed the inter-rater variability. While the best teams reached similar scores in Dice (0.81-0.82) and HD95 (2.1-2.3 mm), ED provided greater discriminative power: the winning method achieved an ED of 20.9, representing roughly a 50% improvement over the second- and third-ranked teams despite comparable Dice scores. Second, we introduced a new 0.55T low-field MRI test set, which, when paired with high-quality super-resolution reconstruction, achieved the highest segmentation performance across all test cohorts (Dice=0.86, HD95=1.69, ED=6.26). This provides the first quantitative evidence that low-cost, low-field MRI can match or surpass high-field systems in automated fetal brain segmentation. Third, the new biometry estimation task exposed a clear performance gap: although the best model reached a mean average percentage error (MAPE) of 7.72%, most submissions failed to outperform a simple gestational-age-based linear regression model (MAPE=9.56%), and all remained above inter-rater variability with a MAPE of 5.38%. Finally, by analyzing the top-performing models from FeTA 2024 alongside those from previous challenge editions, we identify ensembles of 3D nnU-Net trained on both real and synthetic data with both image- and anatomy-level augmentations as the most effective approaches for fetal brain segmentation. Our quantitative analysis reveals that acquisition site, super-resolution strategy, and image quality are the primary sources of domain shift, informing recommendations to enhance the robustness and generalizability of automated fetal brain analysis methods.

Accurate segmentation of cervical structures in transvaginal ultrasound (TVS) is critical for assessing the risk of spontaneous preterm birth (PTB), yet the scarcity of labeled data limits the performance of supervised learning approaches. This paper introduces the Fetal Ultrasound Grand Challenge (FUGC), the first benchmark for semi-supervised learning in cervical segmentation, hosted at ISBI 2025. FUGC provides a dataset of 890 TVS images, including 500 training images, 90 validation images, and 300 test images. Methods were evaluated using the Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), and runtime (RT), with a weighted combination of 0.4/0.4/0.2. The challenge attracted 10 teams with 82 participants submitting innovative solutions. The best-performing methods for each individual metric achieved 90.26% mDSC, 38.88 mHD, and 32.85 ms RT, respectively. FUGC establishes a standardized benchmark for cervical segmentation, demonstrates the efficacy of semi-supervised methods with limited labeled data, and provides a foundation for AI-assisted clinical PTB risk assessment.

Image-text multimodal learning in artificial intelli-gence-based fetal cardiac ultrasound imaging analysis is hindered by instance-level alignment. Moreover, instance-level alignment performs poorly in scenarios with scarce annotations and noisy text, which has greatly limited the development of this field. To address this, this paper proposes Prototype-based Category-Level Image-Text Multimodal Learning (PCLIM), which integrates: (1) Prototype-based Labeled Image-Text Contrastive Learning that refines image-text features via Self & Cross-Attention Fusion and optimizes prototype pools through prototype contrastive loss; (2) Dynamic multi-prototype representation using dual-modality prototype pools updated online via EMA and memory buffers to capture evolving intra-class semantics; and (3) Prototype-matched Unlabeled Multimodal Learning that employs an EMA teacher-student framework to match unlabeled samples to category-level pseudo-text anchors with consistency regularization and contrastive learning. Experiments on fetal cardiac ultra-sound datasets EP-CHDT8000 and SYF -CHD30000 demonstrate that PCLIM consistently outperforms state-of-the-art baselines, especially in low-label scenarios. The code is publicly available at https://github.com/SIGMACX/PCLIM.

We propose VRP-ScribFormer, a CNN-Transformer hybrid framework for scribble-supervised medical image segmentation. It introduces a Visual Reference Prototype (VRP) encoder that transfers structural priors from annotated reference images to the Transformer branch, and a Channel-Spatial Attention Module (CSAM) that refines CNN features under Transformer-guided attention. Experiments on the ACDC and MSCMRseg datasets validate the model's effectiveness, demonstrating substantial improvements in segmentation accuracy by leveraging both structural priors and attention-guided feature refinement.

U-Net architectures are widely used in skin lesion segmentation due to their ability to capture fine-grained spatial details and context. However, most models enhance performance by incorporating complex modules, often overlooking the computational resource constraints present in real-world medical environments. Consequently, there is an urgent need to design models for mobile skin lesion segmentation that are efficient in terms of both parameters and computational load. To address this issue, we propose a lightweight skin lesion segmentation network based on a differential evolution algorithm, the Lightweight Differential Evolution-based Boundary-Assisted UNet for Skin Lesion Segmentation (LDEB-UNet). LDEB-UNet introduces three main innovations: (1) We propose the Separable Convolution with GELU and SE module (SCGS), incorporated into Stages 1-3 of the U-shaped architecture. Utilizing depthwise separable convolutions effectively reduces the model's parameter count, while the addition of channel attention and GELU further enhances its performance. (2) We propose the Hybrid Group Attention Shuffle module (HGAS), which enhances feature interaction across different channel segments, improving model performance without increasing the parameter count. (3) We introduce a Differential Evolution-based Boundary-Assisted module(DEB) into segmentation networks, which improves the model's ability to handle blurry boundaries. Comprehensive experiments on the ISIC 2017 and ISIC 2018 datasets demonstrate that LDEB-UNet outperforms existing state-of-the-art methods. Moreover, to our best knowledge, this is the first model with a parameter count limited to just$\mathbf{2 6 K B}$and Giga-Operations Per Second (GFLOPs) limited to 0.081. Our code is available at https://github.com/cjr851/LDEB-UNet.

Recently, U-shaped neural networks (U-Net) and Full resolution convolutional neural networks (F-Net) have been extensively explored for remote sensing image haze removal, achieving excellent performance. However, downsampling in U-Net leads to significant loss of high-frequency information, while F-Net fails to satisfy the large receptive field demand of remote sensing images, resulting in suboptimal dehazing results for both architectures. Moreover, most existing haze removal methods neglect exploring the correlation between spatial and channel information in feature fusion, which is crucial for restoring image texture details and colors. To address these issues, we propose a Dynamic-Routing 3D-Fusion Network (DR3DF-Net), comprising a Dynamic Routing Features Framework (DRFF) and a 3D Perceptual Feature Fusion (3D-PFF) module. Specifically, the DRFF utilizes a Self-generated Constrained Feature Routing (SCFR) mechanism to learn the most representative features extracted from U-Net, F-Net, and their fused features to enhance clear image reconstruction. Furthermore, the 3D-PFF module enhances interaction between spatial and channel information of multiple features, assigning pixel-level weights for feature fusion, improving dehazed image texture details and colors. Experiments on challenging benchmark datasets demonstrate our DR3DF-Net outperforms several state-of-the-art haze removal methods. The source code is available at https://github.com/lslyttx/DR3DF-Net.

Recently, deep neural networks have been extensively explored in remote sensing image haze removal and achieved remarkable performance. However, most existing haze removal methods fail to effectively leverage the fusion of spatial and frequency information, which is crucial for learning more representative features. Moreover, the prevalent perceptual loss used in dehazing model training overlooks the diversity among perceptual channels, leading to performance degradation. To address these issues, we propose a spatial-frequency residual-guided dynamic perceptual network (SFRDP-Net) for remote sensing image haze removal. Specifically, we first propose a residual-guided spatial-frequency interaction (RSFI) module, which incorporates a bidirectional residual complementary mechanism (BRCM) and a frequency residual enhanced attention (FREA). Both BRCM and FREA exploit spatial-frequency complementarity to guide more effective fusion of spatial and frequency information, thus enhancing feature representation capability and improving haze removal performance. Furthermore, a dynamic channel weighting perceptual loss (DCWP-Loss) is developed to impose constraints with varying strengths on different perceptual channels, advancing the reconstruction of high-quality haze-free images. Experiments on challenging benchmark datasets demonstrate our SFRDP-Net outperforms several state-of-the-art haze removal methods. The code is released publicly athttps://github.com/789as-syl/SFRDP-Net.

Deep learning-based scene classification methods have long been a key research area in remote sensing imagery due to their wide-ranging applications. Recently, Transformer models have achieved significant progress in computer vision, making vision transformers (ViTs) a promising direction for scene classification. However, the spatial complexity of remote sensing imagery poses unique challenges for applying Transformers directly. Manually designing Transformers tailored for remote sensing scene classification is time-consuming under model parameter constraints and requires extensive domain expertise. To address this challenge, neural architecture search (NAS) methods provide an effective solution to construct optimal Transformer architectures for remote sensing scene classification automatically. In this work, we propose SceneFormer, an automated Transformer architecture search framework tailored for scene classification tasks. In SceneFormer, we construct a dedicated search space to search for the optimal Transformer. Moreover, we design a supernet training strategy to train numerous candidate architectures within the search space simultaneously. Furthermore, SceneFormer employs the evolutionary search to find the optimal Transformer architecture under specific resource constraints. Experiments on three high-spatial-resolution (HSR) datasets demonstrate the effectiveness of SceneFormer.

Consistent perturbation strategies have emerged as a dominant paradigm in semi-supervised medical image segmentation. Nevertheless, prevailing approaches inadequately address two critical challenges: 1) prediction errors induced by data uncertainty from distribution shifts, and 2) loss instability caused by model uncertainty in parameter generalization. To overcome these limitations, we propose an Uncertainty-Guided Adaptive Correction (UGAC) framework with three key innovations. First, we develop a dual-path uncertainty rectification mechanism that employs normalized entropy measures to detect error-prone regions in unlabeled predictions, followed by bilateral correction through confidence-weighted fusion. Second, we introduce adversarial consistency constraints that leverage labeled data to discriminate authentic segmentation patterns, effectively regularizing uncertainty propagation in unlabeled predictions through spectral normalization. Third, we architect a frequency-aware segmentation backbone through our novel Freqfusion module, which performs adaptive spectral decomposition during feature decoding to explicitly disentangle high-frequency (boundary-aware) and low-frequency (structural) components, thereby enhancing anatomical boundary sensitivity. Comprehensive evaluations on MM-WHS, BUSI, M&Ms and PROMISE12 datasets demonstrate UGAC's superior performance. The proposed framework exhibits robust generalizability across CT, MRI, and ultrasound modalities, while achieving significantly lower computational complexity than baseline UNet implementations. The code will be available at https://github.com/SIGMACX/UGAC.

Hyperspectral images (HSI) contain hundreds of bands, which provide a wealth of spectral information and enable better characterization of features. However, the excessive dimensions and redundant information also cause a dimensional disaster for subsequent processing. Band selection is a widely-used dimension reduction technique for hyperspectral images. Traditional methods mainly consider the hyperspectral band selection problem at the level of data, and maintain the information contained in the data, without considering the spatial structures inside hyperspectral images. To fill the gap, in this work, an unsupervised hyperspectral band selection method through structure-conserved and neighborhood-grouped evolutionary algorithm (SNEA) is proposed. Different from other evolutionary algorithms for hyperspectral band selection, firstly, two spatial-structure related optimization objectives are designed, including the locally spatial structure denoted by the pixel’s spatial consistency with its adjacent neighbors and the globally spatial structure denoted by the affinity graph among pixels. With the designed objectives, the hyperspectral band selection is formulated as the problem of conserving spatial structures. Moreover, a neighborhood-grouped pair-wise learning strategy is proposed to generate high-quality offsprings. In this novel strategy, a neighborhood grouping operation is developed to divide the band space into several groups. The population can be initialized efficiently and the offspring solutions can be generated pairwisely under the guidance of grouping. Compared with 9 state-of-the-art comparison algorithms, experimental results on 3 standard hyperspectral datasets demonstrate that the band subset obtained by our proposed SNEA has a better classification performance than the comparison algorithms.

Hyperspectral endmember extraction (HEE) is an essential part of remote-sensing image processing. There have been recent attempts to model the HEE as a multiobjective optimization problem and apply multiobjective evolutionary algorithms to solve the problem. However, because of the large HEE search space, it is difficult for the current algorithms to achieve exploration–exploitation balance, and they easily stall prematurely. To address these issues, this article proposes a multistrategy evolutionary multiobjective method based on roulette wheel selection and the genetic algorithm (RWS-GA) for endmember extraction. This method designs two parallel algorithms corresponding to global exploration and local exploitation. In the RWS-GA, an improved NSGA-II method, adopting a novel method to sort individuals on the same front instead of the crowding distance, is proposed to divide individuals into superior and inferior subpopulations. Thereafter, different modified population update strategies are utilized for subpopulations based on characteristics. Pixels that appear more frequently in the population are considered to perform better to have a higher probability of forming an endmember set with other pixels. In addition, excellent individuals often exhibit a higher probability of including endmembers compared with inferior individuals. Considering the abovementioned opinions, roulette wheel selection is performed on the inferior subpopulation for global search. Meanwhile, the superior subpopulation is responsible for local search based on the genetic algorithm (GA). Furthermore, an offspring complement mechanism (OCM) is presented to prevent duplicate individuals from appearing in historical archives. Numerous comparative experiments show that the proposed method is superior to other endmember extraction methods in three real-world datasets.

Endmember extraction (EE) plays a crucial part in the hyperspectral unmixing (HU) process. To obtain satisfactory EE results, the EE can be considered as the multiobjective optimization problem to optimize the volume maximization (VM) and root-mean-square error (RMSE) simultaneously. However, it is often quite challenging to balance the conflict of these objectives. In order to tackle the challenges of multiobjective EE, we present a (μ + λ) multiobjective differential evolution algorithm ((μ + λ)-MODE) based on ranking multiple mutations. In the (μ + λ)-MODE algorithm, ranking multiple mutations are adopted to create the mutant vectors via the scaling factor pool to enhance the population diversity. Moreover, mutant vectors employ the binary crossover operator to generate the trial vectors through a crossover control parameter pool in (μ + λ)-MODE to take advantage of the good information of the population. In addition, (μ + λ)-MODE utilizes the fast nondominated sorting approach to sort the parent and trial vectors, and then selects the elitism offspring as the next population via the (μ + λ) selection strategy. Eventually, experimental comparative results in three real HSIs reveal that our proposed (μ + λ)-MODE is superior to other EE methods.

Endmember extraction (EE) is a significant task in hyperspectral unmixing. From a multiobjective optimization perspective, this task is extremely challenging because objectives often conflict with each other. Currently, a multiobjective discrete particle swarm optimization algorithm (MODPSO) is applied to handle the multiobjective optimization EE problem such as the root-mean-square error (RMSE) and the volume maximization (VM). However, in MODPSO, the minimization of RMSE by unconstrained least squares (Ucls) may lack accuracy, the update of velocity by the predefined random selection probability p can also affect the exploration and exploitation, and it may lose good solution in the process of the update of particles when the particles are randomly chosen in the nondominated relationship. To address these issues, we present an improved MODPSO (IMODPSO) for hyperspectral EE. IMODPSO employs nonnegative constrained least squares (Ncls) to enhance the accuracy of RMSE. Moreover, IMODPSO eliminates the effects of probability p and combines the restart mechanism to achieve a balance of the exploration and exploitation. In addition, IMODPSO utilizes the archive strategy to reserve good nondominated particles to strengthen the population diversity. The experiments have been conducted on three real hyperspectral images and the results have demonstrated that IMODPSO obtains best performances for EE.

Particle swarm optimization (PSO) is a population-based stochastic optimization technique that can be applied to solve optimization problems. However, there are some defects for PSO, such as easily trapping into local optimum, slow velocity of convergence. This paper presents the simple butterfly particle swarm optimization algorithm with the fitness-based adaptive inertia weight and the opposition-based learning average elite strategy (SBPSO) to accelerate convergence speed and jump out of local optimum. SBPSO has the advantages of the simple butterfly particle swarm optimizer to increase the probability of finding the global optimum in the course of searching. Moreover, SBPSO benefits from the simple particle swarm (sPSO) to accelerate convergence speed. Furthermore, SBPSO adopts the opposition-based learning average elite to enhance the diversity of the particles in order to jump out of local optimum. Additionally, SBPSO generates the fitness-based adaptive inertia weight ω to adapt to the evolution process. Eventually, SBPSO presents a approach of random mutation location to enhance the diversity of the population in case of the position out of range. Experiments have been conducted with eleven benchmark optimization functions. The results have demonstrated that SBPSO outperforms than that of the other five recent proposed PSO in obtaining the global optimum and accelerating the velocity of convergence.