Semi-supervised learning (SSL) has shown strong potential in reducing the reliance on large-scale voxel-level annotations for 3D medical image segmentation. However, existing SSL methods often suffer from unstable training and limited generalization due to unreliable pseudo-labels and insufficient structural modeling in unlabeled data. These challenges are especially evident in volumetric contexts, where anatomical structures exhibit high inter-class imbalance and complex spatial dependencies. To address these issues, we propose a semi-supervised framework built upon a single-network architecture that integrates feature learning, consistency regularization, and pseudo-label reliability modeling in a unified manner. The framework comprises three key components: 1) a Confidence-aware Multi-level Fusion Network (CMFN) for capturing robust multi-scale semantic representations; 2) a Semantic-Enhanced Center Alignment (SECA) module to align feature distributions of group-level anatomical structures and mitigate semantic drift in pseudo-labels; and 3) a Group-Guided Reliability Assessment (GGRA) module that enhances pseudo-label reliability by modeling confidence errors in a group-aware structural context. Together, these modules enhance both feature discriminability and the reliability of pseudo-labels.We evaluate our framework on three public 3D medical image segmentation benchmarks: LA, BTCV, and BraTS19. Extensive experiments demonstrate that our method consistently outperforms state-of-the-art approaches under limited annotation, achieving superior accuracy and generalization across diverse anatomical structures and segmentation tasks.

The exploitation of label hierarchy is crucial for effective brain tumor segmentation. Nevertheless, existing methods grapple with two key limitations. First, they lack the hierarchical dependency of predictions across different label levels, rendering the network outputs less interpretable. Second, they fail to exploit the hierarchical similarity among labels, thus hindering potential accuracy enhancement. To address these limitations, we present a novel framework termed deep hierarchy-aware segmentation (DHAS), which achieves both hierarchically interpretable and high-accuracy predictions. Specifically, to generate hierarchical predictions, the network is designed to output pixel-wise probability conditional upon the parent label and is hybrid-trained from conditional to unconditional probability. To utilize the label similarity, we propose a tree-triplet loss, which imposes the hierarchy-induced distance within the feature embedding space. Experimental results on three datasets, BraTS2018, BraTS2019 and BraTS2020, show that our proposed framework achieves significantly better performance than other hierarchy-exploiting methods, and it ranks fifth top among 383 participating methods in Brats2020 Challenge. The improved performance and interpretable predictions promise the potential of DHAS for clinical applications in brain tumor segmentation. Furthermore, its generalization is demonstrated for cardiac segmentation on ACDC dataset.

Hierarchical approaches have been tremendously successful at multi-label segmentation. However, it has been shown they may seriously suffer from the problem of only imposing constraints on shallow layers while ignoring deep relationships in the label space. In this paper we overcome this limitation through a hierarchical multi-class group correlation learning (HMGC). Thus, we first transform regional constraints into voxel vector correlations in a high-dimensional space. After performing transformation, we compute a voxel vector correlation matrix to group voxel vectors to reduce disparities between erroneous and valid vectors. We then introduce two loss functions: intra-class group loss, which minimizes differences within the same class, and inter-class group loss, which adjusts distances between class group centers and voxel vectors. This, in turn, can be used to mitigate bias propagation and improve segmentation accuracy. The effectiveness of our method is demonstrated on three Brain Tumor Segmentation Challenge datasets: BraTS2018, BraTS2019, and BraTS2020. Moreover, generalization of our method is evaluated on the ACDC MICCAI'17 Challenge Dataset. Our HMGC model ranks first in overall score on Brats2020 and achieves one of the most competitive results in cardiac segmentation.

Radiomics refers to the high-throughput extraction of quantitative features from medical images, and is widely used to construct machine learning models for the prediction of clinical outcomes, while feature engineering is the most important work in radiomics. However, current feature engineering methods fail to fully and effectively utilize the heterogeneity of features when dealing with different kinds of radiomics features. In this work, latent representation learning is first presented as a novel feature engineering approach to reconstruct a set of latent space features from original shape, intensity and texture features. This proposed method projects features into a subspace called latent space, in which the latent space features are obtained by minimizing a unique hybrid loss function including a clustering-like loss and a reconstruction loss. The former one ensures the separability among each class while the latter one narrows the gap between the original features and latent space features. Experiments were performed on a multi-center non-small cell lung cancer (NSCLC) subtype classification dataset from 8 international open databases. Results showed that compared with four traditional feature engineering methods (baseline, PCA, Lasso and L2,1-norm minimization), latent representation learning could significantly improve the classification performance of various machine learning classifiers on the independent test set (all p<0.001). Further on two additional test sets, latent representation learning also showed a significant improvement in generalization performance. Our research shows that latent representation learning is a more effective feature engineering method, which has the potential to be used as a general technology in a wide range of radiomics researches.

Deep neural network (DNN) approaches have shown remarkable progress in automatic Chest X-rays classification. However, existing methods use a training scheme that simultaneously trains all abnormalities without considering their learning priority. Inspired by the clinical practice of radiologists progressively recognizing more abnormalities and the observation that existing curriculum learning (CL) methods based on image difficulty may not be suitable for disease diagnosis, we propose a novel CL paradigm, named multi-label local to global (ML-LGL). This approach iteratively trains DNN models on gradually increasing abnormalities within the dataset, i,e, from fewer abnormalities (local) to more ones (global). At each iteration, we first build the local category by adding high-priority abnormalities for training, and the abnormality's priority is determined by our three proposed clinical knowledge-leveraged selection functions. Then, images containing abnormalities in the local category are gathered to form a new training set. The model is lastly trained on this set using a dynamic loss. Additionally, we demonstrate the superiority of ML-LGL from the perspective of the model's initial stability during training. Experimental results on three open-source datasets, PLCO, ChestX-ray14 and CheXpert show that our proposed learning paradigm outperforms baselines and achieves comparable results to state-of-the-art methods. The improved performance promises potential applications in multi-label Chest X-ray classification.

Due to bone deformation and joint space narrowing in diseased hips, accurate segmentation for pelvis, and femur from hip computed tomography (CT) images remains a challenging task. Therefore, the paper presents a fully automatic segmentation framework for the pelvis and femur in both of healthy and diseased hips. The framework involves three steps: preprocessing, coarse segmentation, and refinement. It starts with a preprocessing procedure to extract the volume of interest (VOI) from original CT images. Then, a coarse segmentation of bone has been obtained by classifying the VOI as bone and nonbone parts based on conditional random field (CRF) model. Finally, the bone is further divided into the pelvis and femur using a patch-based refinement method. The innovation of this study is the novel patch-based refinement method that is particularly suitable for diseased hips. The refinement method starts from the boundary of coarse segmentation, and propagates to the neighbors only when the label is not consistent with the label of CRF-based classification, it increases the reliability of segmentation for diseased hips with bone deformation. We incorporate neighborhood information to label fusion so that final label estimation is more accurate and robust for diseased hips with joint space narrowing. In total, 60 CT data sets, which included 78 healthy hemi-hips and 42 diseased hemi-hips, were used, and three-fold cross validations were carried out. Compared to two state-of-the-art methods, our method achieved significantly increased segmentation accuracy for the diseased hemi-hips, and is, therefore, more suited for automatic segmentation of diseased hips.

We describe an active contour framework with accurate shape and size constraints on the vessel cross-sectional planes to produce the vessel segmentation. It starts with a multiscale vessel axis tracing in a 3D computed tomography (CT) data, followed by vessel boundary delineation on the cross-sectional planes derived from the extracted axis. The vessel boundary surface is deformed under constrained movements on the cross sections and is voxelized to produce the final vascular segmentation. The novelty of this paper lies in the accurate contour point detection of thin vessels based on the CT scanning model, in the efficient implementation of missing contour points in the problematic regions and in the active contour model with accurate shape and size constraints. The main advantage of our framework is that it avoids disconnected and incomplete segmentation of the vessels in the problematic regions that contain touching vessels (vessels in close proximity to each other), diseased portions (pathologic structure attached to a vessel), and thin vessels. It is particularly suitable for accurate segmentation of thin and low contrast vessels. Our method is evaluated and demonstrated on CT data sets from our partner site, and its results are compared with three related methods. Our method is also tested on two publicly available databases and its results are compared with the recently published method. The applicability of the proposed method to some challenging clinical problems, the segmentation of the vessels in the problematic regions, is demonstrated with good results on both quantitative and qualitative experimentations; our segmentation algorithm can delineate vessel boundaries that have level of variability similar to those obtained manually.