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.

Major adverse cardiac events (MACE) pose a high life-threatening risk to patients with arrhythmogenic right ventricular cardiomyopathy (ARVC). Cardiac magnetic resonance (CMR) has been proven to reflect the risk of MACE, but two challenges remain: limited dataset size due to the rarity of ARVC and overlapping image distributions between non-MACE and MACE patients. To address these challenges by fully leveraging the dynamic and spatial information in the limited CMR dataset, a deep learning-based risk prediction model named Three-Tier Spatiotemporal Transformer (TTST) is proposed in this paper, which utilizes three transformer-based tiers to sequentially extract and fuse features from three domains: the 2D spatial domain of each slice, the temporal dimension of slice sequence and the inter-slice depth dimension. In TTST, a pericardial adipose tissue (PAT) embedding unit is proposed to incorporate the dynamic and positional information of PAT, a key biomarker for distinguishing MACE from non-MACE based on its thickening and reduced motion, as prior knowledge to reduce reliance on large-scale datasets. Additionally, a patch voting unit is introduced to pick out local features that highlight more indicative regions in the heart, guided by the PAT embedding information. Experimental results demonstrate that TTST outperforms existing classification methods in MACE prediction (internal: AUC = 0.89, ACC = 84.02%; external: AUC = 0.87, ACC = 86.21%). Clinically, TTST achieves effective risk prediction performance either independently (C-index = 0.744) or in combination with the existing 5-year risk score model (increasing C-index from 0.686 to 0.777). Code and dataset are accessible at https://github.com/DFLAG-NEU.

Longitudinal deformable registration is crucial for precision radiotherapy but is challenged by large-scale organ deformations and anatomical coupling. While existing deep learning registration methods have shown promise, they often face two key limitations: (1) a trade-off in feature extraction between local details and long-range dependencies, and (2) feature distortion from the naive fusion of spatially misaligned features. To overcome these challenges, we propose HCFMorph, a novel, unified coarse-to-fine registration framework. Our framework introduces three key innovations: (1) a Synergistic Mamba-Conv module to model coupled organ motion by capturing global-local features; (2) a Contribution-Aware Feature Fusion module to adaptively fuse features from the fixed and moving images, thereby mitigating feature misalignment; and (3) a Deformation-Guided Hierarchical Refinement strategy in the decoder, which addresses large-scale deformations by recovering the field in a coarse-to-fine manner and leverages low-resolution fields to guide and pre-align high-resolution features before the CAFF stage, forming a novel feedback loop. Experiments on a clinical dataset of 432 longitudinal CT scans from 97 cervical cancer patients show that HCFMorph significantly outperforms state-of-the-art methods. It achieves dominant anatomical accuracy while maintaining excellent topological integrity with a near-zero Jacobian folding rate, striking a superior balance between accuracy and plausibility.

Deep learning methods have become a staple in the field of medical image segmentation. While traditional segmentation methods have achieved significant success, challenges such as the labelling of medical image datasets and concerns regarding patient privacy have constrained their practical application. In recent years, incremental learning has garnered considerable attention due to its ability to address these limitations through a step-by-step learning approach. Nevertheless, the inextricable nature of medical images between organisations can result in their catastrophic oblivion. In light of the aforementioned reasons, this study proposes a deep learning framework, CCL, with the objective of systematically addressing the challenge of incremental learning of categories in cardiac magnetic resonance imaging (CMR) segmentation tasks. Rather than employing the reutilisation of pre-existing images, the method deploys the utilisation of pseudo-labels generated from legacy models to facilitate the reutilisation of acquired knowledge. This approach does not involve the reutilisation of preexisting images; rather, it employs pseudo-labels generated by preceding models to reclaim knowledge that has been previously acquired. Firstly, a confidence-based pseudolabel strategy is employed to retrieve old category pixels from the background. Secondly, a contrastive learning mechanism is proposed, based on the characteristics of the teacher-student model. During the training of new models, the teacher model corrects errors of forgetting and consistency caused by incorrect pseudolabels. This ensures maximum retention of prior knowledge without compromising the acquisition of new knowledge. Two publicly available cardiac magnetic resonance datasets were utilised for the purpose of comparison between the proposed methodology and classical approaches. The findings demonstrated that the CCL method outperformed several other excellent methods. The findings of the quantitative analysis suggest that the proposed feature contrast mechanism has a significant effect on the forgetting of old class features, without affecting the learning ability of new classes. The code is available at https://anonymous.4open.science/r/ccl-3C56/.

Deep learning-based cross-modality image translation has demonstrated great potential in supporting clinical diagnosis and medical research. Among existing approaches, CycleGAN has been widely adopted due to its effective unsupervised training mechanism. However, it suffers from two major limitations: (1) its cycle-consistency loss provides only an indirect constraint, often leading to anatomical structure distortions during translation; and (2) its limited capacity to capture global contextual information results in synthetic artifacts in the generated images. To address these issues, we propose a novel and efficient unsupervised framework named ConSSMGAN, which is built upon a cycle-consistent architecture and introduce two key modules: (1) a Contrastive Perception Module, which introduces BoNCE loss based on the contrastive learning to enhance training constraints and alleviate anatomical distortions during translation; and (2) a SSM-Based Global Enhancement Module, which enhances the model's ability to capture global contextual features and fuses them with local representations to reduce synthetic artifacts in the generated results. We validate the effectiveness of our method on both a public pelvic dataset and a private clinical pelvic dataset, comparing it with six state-of-the-art approaches across two key evaluation dimensions: image similarity and anatomical accuracy. Experimental results demonstrate that our method achieves superior performance, highlighting its robustness and potential for clinical application.

Abstract Lane detection is a fundamental problem in autonomous driving, which provides vehicles with essential road information. Despite the attention from scholars and engineers, lane detection based on LiDAR meets challenges such as unsatisfactory detection accuracy and significant computation overhead. In this paper, the authors propose LLFormer4D to overcome these technical challenges by leveraging the strengths of both Convolutional Neural Network and Transformer networks. Specifically, the Temporal Feature Fusion module is introduced to enhance accuracy and robustness by integrating features from multi‐frame point clouds. In addition, a sparse Transformer decoder based on Lane Key‐point Query is designed, which introduces key‐point supervision for each lane line to streamline the post‐processing. The authors conduct experiments and evaluate the proposed method on the K‐Lane and nuScenes map datasets respectively. The results demonstrate the effectiveness of the presented method, achieving second place with an F1 score of 82.39 and a processing speed of 16.03 Frames Per Seconds on the K‐Lane dataset. Furthermore, this algorithm attains the best mAP of 70.66 for lane detection on the nuScenes map dataset.

Due to the association with higher incidence of left ventricular dysfunction and complications, segmentation of left ventricle and related pathological tissues: microvascular obstruction and myocardial infarction from late gadolinium enhancement cardiac magnetic resonance images is crucially important. However, lack of datasets, diverse shapes and locations, extreme imbalanced class, severe intensity distribution overlapping are the main challenges. We first release a late gadolinium enhancement cardiac magnetic resonance benchmark dataset LGE-LVP containing 140 patients with left ventricle myocardial infarction and concomitant microvascular obstruction. Then, a progressive deep learning model LVPSegNet is proposed to segment the left ventricle and its pathologies via adaptive region of interest extraction, sample augmentation, curriculum learning, and multiple receptive field fusion in dealing with the challenges. Comprehensive comparisons with state-of-the-art models on the internal and external datasets demonstrate that the proposed model performs the best on both geometric and clinical metrics and it most closely matched the clinician's performance. Overall, the released LGE-LVP dataset alongside the LVPSegNet we proposed offer a practical solution for automated left ventricular and its pathologies segmentation by providing data support and facilitating effective segmentation. The dataset and source codes will be released via https://github.com/DFLAG-NEU/LVPSegNet.

Bias fields adversely affect various automatic analysis technologies. Therefore, bias field correction is essential. However, deep learning based methods encounter challenges in obtaining ground truth. Although existing methods attempt to address this problem by using training-free techniques or constructing datasets with approximations of the ground truth, the lack of task-oriented guidance, training instability, and inappropriate use of approximations still impact performance. Additionally, different approaches for bias field correction, i.e., estimating bias fields versus directly restoring clean images, exhibit different performances. However, there is no consensus on the best way, resulting in limited performance in some cases. To address these problems, we propose a bias field generation method to construct the dataset and provide task-oriented information. We then propose the concept of Equivalent of Residual Mapping (ERM) to analyze the advantages of estimating bias fields. According to ERM, we propose the Bias field Estimation Convolutional Neural Network (BECNN). Finally, we propose the Dual Route Implicit Supervised Learning (DRISL) strategy to balance the guidance derived from approximations of the ground truth with over-dependence on them. The proposed method is compared qualitatively and quantitatively with the correlated methods. Experiment results demonstrate that the proposed method performs effectively both on bias field estimation and correction.

Dimensionality reduction and clustering are crucial tasks in single-cell RNA sequencing (scRNA-seq) data analysis, treated independently in the current process, hindering their mutual benefits. The latest methods jointly optimize these tasks through deep clustering. However, contrastive learning, with powerful representation capability, can bridge the gap that common deep clustering methods face, which requires pre-defined cluster centers. Therefore, a dual-level contrastive clustering method with nonuniform sampling (nsDCC) is proposed for scRNA-seq data analysis. Dual-level contrastive clustering, which combines instance-level contrast and cluster-level contrast, jointly optimizes dimensionality reduction and clustering. Multi-positive contrastive learning and unit matrix constraint are introduced in instance- and cluster-level contrast, respectively. Furthermore, the attention mechanism is introduced to capture inter-cellular information, which is beneficial for clustering. The nsDCC focuses on important samples at category boundaries and in minority categories by the proposed nearest boundary sparsest density weight assignment algorithm, making it capable of capturing comprehensive characteristics against imbalanced datasets. Experimental results show that nsDCC outperforms the six other state-of-the-art methods on both real and simulated scRNA-seq data, validating its performance on dimensionality reduction and clustering of scRNA-seq data, especially for imbalanced data. Simulation experiments demonstrate that nsDCC is insensitive to "dropout events" in scRNA-seq. Finally, cluster differential expressed gene analysis confirms the meaningfulness of results from nsDCC. In summary, nsDCC is a new way of analyzing and understanding scRNA-seq data.

Regularization-based continuous learning (RCL) has been proven to be highly effective on struggling against catastrophic forgetting. However, its application in medical image segmentation (MIS) is relatively scarce. In this paper, we provide a unified description of 6 RCL methods (LwF, LwM, EWC, SI, MAS, and PIGWM) using Taylor expansion and investigate their performances in 2 classic MIS scenarios, namely retinal vessel segmentation (RVS) and cardiac left ventricle segmentation (CLVS) on 8 datasets (CHASE, DRHAGIS, RITE and STARE for the former and M&Ms, LVSC, ACDC and SCD for the latter). We also explore the influence of different task orders (easy to hard or hard to easy), optimizers (Adam or SGD), and parameter capacities (2, 3, 4 or 5 down- and up-sampling pairs) on the performance of these methods. Our experimental results show that these methods are capable of mitigating catastrophic forgetting to a certain extent. Comparing to a hard-to-easy order, most of the methods perform better on all of the already known tasks in an easy-to-hard order. Optimizer Adam performs better on RVS and CLVS. Capacity increases are obviously effective for CLVS, but they have no significant impact on RVS.

Increased pericardial adipose tissue (PEAT) is associated with a series of cardiovascular diseases (CVDs) and metabolic syndromes. Quantitative analysis of PEAT by means of image segmentation is of great significance. Although cardiovascular magnetic resonance (CMR) has been utilized as a routine method for non-invasive and non-radioactive CVD diagnosis, segmentation of PEAT in CMR images is challenging and laborious. In practice, no public CMR datasets are available for validating PEAT automatic segmentation. Therefore, we first release a benchmark CMR dataset, MRPEAT, which consists of cardiac short axis (SA) CMR images from 50 hypertrophic cardiomyopathy (HCM), 50 acute myocardial infarction (AMI), and 50 normal control (NC) subjects. We then propose a deep learning model, named as 3SUnet, to segment PEAT on MRPEAT to tackle the challenges that PEAT is relatively small and diverse and its intensities are hard to distinguish from the background. The 3SUnet is a triple-stage network, of which the backbones are all Unet. One Unet is used to extract a region of interest (ROI) for any given image with ventricles and PEAT being contained completely using a multi-task continual learning strategy. Another Unet is adopted to segment PEAT in ROI-cropped images. The third Unet is utilized to refine PEAT segmentation accuracy guided by an image adaptive probability map. The proposed model is qualitatively and quantitatively compared with the state-of-the-art models on the dataset. We obtain the PEAT segmentation results through 3SUnet, assess the robustness of 3SUnet under different pathological conditions, and identify the imaging indications of PEAT in CVDs. The dataset and all source codes are available at https://dflag-neu.github.io/member/csz/research/.

Automated pancreatic cancer segmentation is highly crucial for computer-assisted diagnosis. The general practice is to label images from selected modalities since it is expensive to label all modalities. This practice brought about a significant interest in learning the knowledge transfer from the labeled modalities to unlabeled ones. However, the imaging parameter inconsistency between modalities leads to a domain shift, limiting the transfer learning performance. Therefore, we propose an unsupervised domain adaptation segmentation framework for pancreatic cancer based on GCN and meta-learning strategy. Our model first transforms the source image into a target-like visual appearance through the synergistic collaboration between image and feature adaptation. Specifically, we employ encoders incorporating adversarial learning to separate domain-invariant features from domain-specific ones to achieve visual appearance translation. Then, the meta-learning strategy with good generalization capabilities is exploited to strike a reasonable balance in the training of the source and transformed images. Thus, the model acquires more correlated features and improve the adaptability to the target images. Moreover, a GCN is introduced to supervise the high-dimensional abstract features directly related to the segmentation outcomes, and hence ensure the integrity of key structural features. Extensive experiments on four multi-parameter pancreatic-cancer magnetic resonance imaging datasets demonstrate improved performance in all adaptation directions, confirming our model's effectiveness for unlabeled pancreatic cancer images. The results are promising for reducing the burden of annotation and improving the performance of computer-aided diagnosis of pancreatic cancer. Our source codes will be released at https://github.com/SJTUBME-QianLab/UDAseg, once this manuscript is accepted for publication.