Hepatocellular carcinoma is a highly heterogeneous and complex malignant tumor, posing significant challenges for precise diagnosis and treatment. Existing methods face limitations in multimodal data integration and multi-phase feature modeling, making it difficult to fully exploit the complementary information from multi-phase CT images, structured clinical data, and medical texts. On one hand, most methods rely solely on single-modal data (e.g. CT images or clinical reports), failing to effectively utilize the complementary characteristics of multimodal data. On the other hand, traditional approaches typically concatenate multi-phase CT images as input to the model, ignoring the dynamic evolution features between different phases, leading to information loss and limited predictive performance. To address these issues, we propose a novel Tri-modal Phase-aware Contrastive Learning Framework (TPCL), which incorporates a Phase-aware Attention Fusion Network (PAAF-Net) and a Phase-Conditional Prompt Network (PCPN) to achieve deep alignment and integration of multimodal features. Additionally, we design a Multi-modal Contrastive Loss to further optimize the consistency of feature distributions across different modalities. Experimental results on multiple public and private datasets demonstrate that TPCL significantly outperforms existing methods, achieving up to an$18.54\%$improvement in ACC and a$10.09\%$improvement in AUC.

Subtype classification of medical images is a key challenge in medical diagnosis. However, due to scarce labeled data, data quality issues, and poor generalization ability of existing models, certain limitations exist. This study proposes a dynamic prototype multi-model fusion (DPF) framework based on a feature screening mechanism. The framework integrates ResNet and Vision Transformer architectures to extract complementary feature representations and constructs robust prototypes through distance-based threshold setting and voting mechanisms. A feature screening layer distinguishes high-quality from lowquality unlabeled samples using similarity and occurrence maps, while a collaborative training strategy utilizes both sample types to optimize prototype representations. The dynamic prototype expansion mechanism enables real-time adaptation to data distribution changes through self-training. Comprehensive experiments are conducted on two medical imaging datasets, including one public breast MRI dataset and one private CT liver dataset. The proposed framework achieves superior performance with accuracy improvements over supervised baselines and existing semi-supervised methods. The results verify the framework's effectiveness in addressing medical image heterogeneity, sample quality variations, and limited labeled data availability.

Multi-parametric magnetic resonance imaging (mpMRI) is widely used in the diagnosis of orbital lymphoproliferative disorders (OLPDs) due to its non-invasive nature. However, in clinical practice, contrast-enhanced T1-weighted (T1C) images are often unavailable due to contraindications to gadolinium-based contrast agents, meanwhile T2-weighted (T2w) images may also be omitted for time-sensitive diagnoses, making it a challenge to generate these images from T1-weighted (T1 w) image alone for multimodal differential diagnosis. Generative adversarial network (GAN)-based models partially address the issue of missing modalities in medical image analysis; however, they often suffer from unstable generation of missing images and lack integration with subsequent diagnostic tasks. To this end, we propose a One-to-Many Generative Adversarial Network (OMGAN) for diagnosing OLPDs in incomplete mpMRI, consisting of a cross-modal generator and a self-representation module, enabling multimodal diagnosis using pre-contrast images alone within a single model. Specifically, we first design an image-modality fusion module that incorporates trigonometric function coding and mixup augmentation to effectively guide the generation from T1 w to T2w and T1 C within one model. Then, we construct a cross-modal generator with a semantic disambiguation block to synthesize the missing images. Meanwhile, we use a self-representation module with a classification-guided branch to effectively extract task-relevant image features. Finally, multimodal features are fused to accomplish the differential diagnosis of OLPDs in the downstream task. Experiments on internal datasets demonstrated that OMGAN outperforms state-of-the-art GAN-based models, with the area-under-the-curve and accuracy improving by 8.18-14.04% and 12.53-16.39%, respectively. Codes are available at https://github.com/3Iasticheart/OMGAN.

Bioluminescence tomography (BLT) can provide non-invasive quantitative three-dimensional tumor information which has been widely applied in pre-clinical studies. Meanwhile, in recent years, deep learning methods have significantly improved the reconstruction resolution and speed by establishing a non-linear mapping relationship between surface-measured bioluminescence and light source distribution. However, this mapping relationship only works for specific biological tissues and light transmission processes under fixed wavelengths, resulting in poor stability and generalizability. To meet the requirements of diverse practical scenarios and inspired by more effective sparse regularization and graph representation theory, we propose a novel Graph Attention Iterative Contraction Network (GAICN) to conduct a finite element mesh spatial representation study. In the GAICN framework, two learnable spatial topological transforms based on the graph attention mechanism and an iterative contraction activation function were devised to achieve non-local feature aggregation and dynamic adjustment of weights between first-order neighboring nodes in the mesh. As a deep unrolling method, GAICN naturally inherits the coherence of surface bioluminescence with the light source in Forward-Backward Splitting (FBS), thus enhancing the generalizability, stability and interpretability of the network. Both simulation and in-vivo experiments further indicated that GAICN achieved superior reconstruction performance in terms of spatial location, dual light source resolution, stability, generalizability, as well as in-vivo practicability.

Automatic infected segmentation on CT enables rapid quantitative analysis of lung involvement in COVID-19 infections. However, it is a challenge to use limited information for segmentation of small infected regions. To deal with the problem, a Multi-Lung Window Fusion Network (MLWF-Net) is proposed to sufficiently utilize CT information. In the first part of MLWF-Net, the multi-lung window block is introduced to address the long-tailed distribution of CT Hounsfield Unit for more features. Then, the adaptive feature aggregation block which combines the attention mechanism with the message passing mechanism is developed to fuse the infection features adaptively from different lung windows CT. The MLWF-Net model is evaluated on a publicly available dataset and a private dataset, compared with a variety of widely-used segmentation models. The MLWF-Net holds on better performance with some benchmarks, especially for small infected areas. The proposed model is an efficient method to alleviate the long tail problem of CT Hounsfield Unit and improve the performance of small infected areas, which can be used for the grading assessment and predictive treatment of COVID-19 patients, with significant clinical value.