BackgroundYoung women with localized breast cancer represent a clinically distinct population with heterogeneous outcomes, yet age-specific prognostic models remain limited. Conventional risk stratification tools derived from mixed-age cohorts may fail to capture the complex interactions between tumor biology and treatment response in this group.MethodsWe conducted a single-center retrospective cohort study including 1,060 women aged ≤40 years diagnosed with stage I–III breast cancer between 2000 and 2023. Overall survival (OS) was analyzed using Kaplan–Meier estimates and multivariable Cox regression. To enable data-driven risk prediction beyond linear assumptions, a machine learning–based Random Survival Forest (RSF) model was developed to identify key prognostic features, quantify variable importance, and stratify patients into distinct risk groups.ResultsAmong 1,060 eligible patients, 110 deaths (10.4%) occurred during a median follow-up of 79.8 months. Invasive pathological subtype (hazard ratio [HR] = 5.23, 95% confidence interval [CI] 1.18–23.22; p = 0.030), nipple invasion (HR = 3.95, 95% CI 2.14–7.27; p < 0.001), and advanced T stage (T2-4 vs. Tis/T1; HR = 1.60, 95% CI 1.03–2.48; p = 0.036) were independently associated with worse OS. By contrast, receipt of endocrine therapy (HR = 0.54, 95% CI 0.36–0.80; p = 0.002) and radiotherapy (HR = 0.53, 95% CI 0.32–0.86; p = 0.010) were associated with better OS. Notably, high Ki67 expression (≥35%; HR = 0.39, 95% CI 0.21–0.71; p = 0.002) was associated with improved OS. The RSF model confirmed these predictors, ranked radiotherapy as the most influential variable, and provided effective risk stratification (C-index = 0.723).ConclusionBy integrating clinicopathological variables with machine learning–based survival modeling, this study identified key prognostic factors associated with OS in young women with localized breast cancer. The findings highlight the prognostic importance of treatment-related factors and reveal an unexpected association between high Ki-67 expression and better survival in this population. These data-driven risk stratification approaches may contribute to more personalized prognostic assessment and warrant validation in prospective multicenter studies.

Background:Intrahepatic cholangiocarcinoma (ICC) literature suggests that hypervascularity on arterial contrast agent enhancement (CE) is associated with better prognosis, while rich stroma is associated with poorer prognosis. This exploratory study evaluates the relationships among slow diffusion coefficient (SDC) and diffusion-derived 'vessel density' (DDVD) signal features, Gadoxetate delayed enhancement, and survival in patients with mass-forming ICC. An ICC with higher SDC may be associated with potentially better patient survival due to less stromal fibrosis. Methods:Data were prospectively acquired in two centers [Center 1 (Sun Yat-Sen Memorial Hospital): ICC 21 cases; Center 2 (Fifth Affiliated Hospital of Anhui Medical University): ICC 3 cases]. SDC maps were derived with b=500 and 800 s/mm2 images for Center 1 data, and with b=400 and 600 s/mm2 images for Center 2 data. DDVD maps were calculated from b=0 and 10 s/mm2 images, ADC was calculated from b=0 and 800 s/mm2 images (Center 1) or b=0 and 600 s/mm2 images (Center 2). Relative to the liver signal, tumor SDC and DDVD signals were assigned to six semi-quantitative score (SQS) categories: low signal (scored as '0'), iso-signal (scored as '1'), slightly high signal (scored as '1.5'), high signal (scored as '2'), higher signal (scored as '2.5'), and markedly high signal (scored as '3'). SDC and ADC were also quantitatively measured. Delayed phase magnetic resonance imaging (MRI) scan was obtained at a median of 3 min 51 sec after the contrast agent injection. Percentage area of CE during delayed phase ('% delay CE') was estimated visually and also quantified with the spleen CE signal intensity as the reference. Follow-up data were available in 20 cases from Center 1 (n=9 alive till the last follow-up, n=11 died during the follow-up), and we conducted additional histopathological grading for 14 cases from Center 1. Results:Tumor aggressiveness was negatively correlated with quantitative '% delay CE' (Spearman rs =0.703, P=0.008). Based on visual assessment, both SDC SQS and DDVD SQS were moderately and positively correlated with '% delay CE', with a Spearman rs of 0.504 (P=0.02) and 0.549 (P=0.03), respectively. Correlation of quantitative '% delay CE' with relative threshold of spleen CE signal intensity and ln(SDCICC/SDCliver) derived a Pearson r of 0.481 (P=0.03). The alive group (n=9) had a SDC SQS higher than the dead group (n=11, mean: 2.89 vs. 2.23, P=0.007). The dead group had 4 cases with SDC SQS ≥2.5, three of them had tumor sizes >4,000 mm2, and an additional patient had an advanced age of 76 years. Conclusions:ICCs with higher SDC score (≥2.5) and small/intermediate presenting tumor size (<4,000 cm2 in the largest section) are associated with better survival potential. ICCs with lower SDC score (≤2.0) are likely associated with a poor prognosis.

Accurate delineation of nasopharyngeal carcinoma (NPC) lesions is crucial for radiotherapy planning, but remains challenging in clinical MRI due to irregular tumor morphology, indistinct boundaries, and complex surrounding anatomy. Existing CNN-based and CNN-Transformer segmentation models improve local and global feature representation, but decoderside reconstruction may still suffer from background interference and unstable lesion boundaries. In this paper, we propose EMATransUNet, a 3D TransUNet-based segmentation model enhanced with a decoder-side Enhanced Multi-Scale Attention (EMA) module at the third and fourth decoder stages. EMA projects grouped 3D features into a 2D in-plane representation and broadcasts the learned attention cues back to the 3D feature space, thereby strengthening lesion-related responses while suppressing redundant background interference. Experiments on a private clinical NPC MRI dataset show that EMA-TransUNet achieves a DSC of 74.64%, the highest Recall of 77.24%, and the lowest HD95 of 5.51 mm among the compared methods. Compared with 3D TransUNet, EMA-TransUNet reduces HD95 by 0.92 mm while introducing only 0.013M additional parameters and 0.294G additional FLOPs.

Designing a targeted magnetic resonance imaging (MRI) probe is challenging due to the intrinsically micromolar-millimolar sensitivity of MRI and the high background signal associated with traditional gadolinium complexes. To address the challenges, we introduce a bioorthogonal catalysis-activated FeII-based MRI probe. The FeII probe exhibited near zero relaxivity (0.01 mM-1 s-1) in its low-spin state. Upon bioorthogonal catalysis mediated by a ruthenium complex, the probe undergoes a spin-state conversion from low-spin to high-spin followed by spontaneous oxidation to high-spin FeIII, yielding a mixture of high-spin FeII and FeIII species under photophysical conditions. This transition led to over a 100-fold enhancement in relaxivity with a high turnover number (>200), generating robust MRI contrast. The ruthenium complex was conjugated to bovine serum albumin and predelivered to the tumor site. In vivo bioorthogonal catalysis was subsequently demonstrated in a subcutaneous tumor model. At the tumor site, the catalyst specifically triggers a "turn-on" in relaxivity of the MRI probe, producing markedly enhanced MRI contrast via bioorthogonally catalyzed activation. By combining the specificity of bioorthogonal chemistry with the MRI probe with a tunable spin state, this work establishes the potential of the spin-modulated probe for targeted imaging with low background signals and high imaging contrast.

Manganese(III) porphyrins have been extensively investigated as both superoxide dismutase (SOD) mimics and magnetic resonance imaging contrast agents. However, a fundamental "charge contradiction" exists between their therapeutic and diagnostic applications: high SOD activity requires a cationic character, whereas effective MRI probes are typically anionic. To address this challenge, we designed zwitterionic Mn(III) porphyrins and evaluated them in vitro and in vivo for MRI contrast efficacy and SOD-like activity in an acute kidney injury (AKI) model. These complexes showed a relaxivity of 8-9 mM-1s-1 (1.41 T) and biocompatibility comparable to anionic analogues while retaining SOD-like activity similar to cationic ones. Cytotoxicity and uptake studies revealed balanced charge mitigation and modulated internalization. Zwitterionic MnT-4-PyP-COOH demonstrated pronounced MRI signal enhancement, excellent biocompatibility, and potent antioxidant activity in a lipopolysaccharide-induced murine model of AKI. This work offers a rational design strategy for dual-functional agents that enable combined MR imaging and antioxidant therapy.

Accurate detection of microhepatocellular carcinoma (HCC) remains a major clinical challenge owing to the limited specificity and sensitivity of current imaging modalities. Herein, we present a dual-injection magnetic resonance imaging (MRI) peptidic probe based on in vivo membrane engineering, achieving in situ signal amplification and molecularly precise imaging. The first injection of programmable nanoparticles coassembled from two peptide monomers, incorporating a GPC3-targeting ligand, a β sheet-forming motif, a dibenzocyclooctyne (DBCO) handle, and porphyrin IX (PpIX) for fluorescence tracking. Following systemic administration, the nanoparticles high-specifically bind to GPC3-overexpressing tumor membranes and transform into surface-anchored nanofibrils, exposing confined DBCO groups. A second injection of azide-modified Gd-DOTA enables rapid copper-free click conjugation on the nanofibrillar scaffold, yielding a nearly fourfold increase in longitudinal relaxivity. MRI demonstrated strong T1-weighted signal enhancement and high tumor-to-liver contrast in Hepa1-6 tumor-bearing mice. In vivo membrane engineering strategy establishes a generalizable platform for receptor-guided molecular imaging and early cancer detection.

An integrated diagnostic strategy of preoperative identification of sentinel lymph node (SLN) metastasis, SLN metastatic burden, and non-SLN (NSLN) metastasis in breast cancer remains to be developed to guide axillary surgery de-escalation. Here we develop a magnetic resonance imaging-based hierarchical multitask deep learning model, breast cancer axillary lymph node network (BCALN-Net) to predict SLN metastasis, SLN metastatic burden, and NSLN metastasis in 6,271 breast cancer patients. BCALN-Net achieves high performance in predicting SLN metastasis, SLN metastatic burden, and NSLN metastasis and exhibits robust performance across molecular subtypes, clinical tumor stages, clinical node stages, estrogen receptor statuses, human epidermal growth factor receptor 2 statuses, menopausal statuses, and SLN metastatic burdens. In pooled analysis of 4,081 patients, BCALN-Net also shows superior performance in predicting the omission of axillary invasive procedures and added value over clinical criteria. BCALN-Net holds the potential to provide an integrated diagnostic strategy of ALN status to help axillary surgery de-escalation in breast cancer patients.

Hydrogen-bonded organic frameworks (HOFs) and covalent organic frameworks (COFs) are two prominent classes of crystalline porous materials assembled from organic building blocks through hydrogen bonds and covalent bonds, respectively. Their well-defined structures, tunable porosity, and designable functionalities have attracted extensive attention in various fields. In particular, luminescent HOFs and COFs exhibit remarkable advantages in biomedical applications, including excellent photostability, low toxicity, modular design for functionalization, and strong luminescent signals suitable for cancer therapy, bioimaging, antibacterial, and biosensing. This review summarizes the fundamental characteristics and recent research progress of luminescent HOFs and COFs and their underlying emission mechanisms. Furthermore, it provides an in-depth discussion on the design strategies for constructing photoactive HOFs/COFs and highlights their applications in cancer therapy, bioimaging, antibacterial, and biosensing. Finally, a perspective is offered on the existing challenges and future directions of luminescent HOFs/COFs in biomedical applications. This review aims to provide valuable insights for researchers dedicated to advancing luminescent HOFs/COFs-based materials for biomedical applications.

The application of artificial intelligence in gynecological magnetic resonance imaging (MRI) has been limited by the scarcity of labeled data and the "task silo" dilemma, resulting in poor generalization ability. We introduced GynoMRFound, the first gynecological MRI foundation model, trained on a large-scale multi-center dataset of 28,755 patients and 435,471 sequences. The model employed a dual-path paradigm combining 3D masked image reconstruction with explicit learning from structured report metadata, enabling accurate visual-semantic alignment without dense annotations. When evaluated across 42 clinical tasks, including diagnosis, staging, subtyping, biomarker prediction, prognosis, report generation, and segmentation, GynoMRFound outperformed baseline methods on 40 tasks under a frozen backbone setting, demonstrating strong generalizability. The model and code were open-sourced, providing a reproducible foundation tool to advance precise diagnosis and treatment in gynecologic oncology.

Background:This study investigates the differentiation of liver focal nodular hyperplasia (FNH) from liver malignant tumor (MT) by a combination of T2-weighted imaging (T2WI), diffusion-derived vessel density (DDVD), slow diffusion coefficient (SDC), and apparent diffusion coefficient (ADC). Based on the odds ratio (OR) for a sign to suggest the possibility of a lesion being FNH, we propose a liver mass sum score (LiverMss-FNH) scheme to facilitate the diagnosis. Methods:Liver diffusion-weighted magnetic resonance imaging included 13 cases of FNH and 82 cases of MT. DDVD was calculated from b=0 and b=10 s/mm2 images, SDC was calculated from b=500 and b=800 s/mm2 images, and ADC was calculated from b=0 and b=800 s/mm2 images. For liver semi-quantitative analysis, relative to the adjacent liver signal, a liver lesion's signal was assigned to five categories: low signal, iso-signal, slightly high signal, high signal, and markedly high signal. The lesion on T2WI being not high signal was assigned a sub-score "1" (otherwise scored 0); the lesion being iso-signal on DDVD was assigned a sub-score "1.5" (otherwise scored 0); the lesion on SDC being not high signal was assigned a sub-score "1" (otherwise scored 0); the lesion on ADC being not low signal was assigned a sub-score "0.5" (otherwise scored 0); the existence of stellate scar was assigned a sub-score "0.5" (otherwise scored 0). The sum of these five sub-scores was termed LiverMss-FNH. Results:A total of 26 MT cases had large (median 8.1 cm, standard deviation: 4.2 cm) and very heterogeneous masses which were very unlikely to be FNH. The remaining 13 FNH cases (median 3.8 cm, standard deviation: 1.7 cm) and 56 MT cases (median 4.9 cm, standard deviation: 4.3 cm; hepatocellular carcinoma, n=40; metastasis, n=12; intrahepatic cholangiocarcinoma, n=4) were evaluated with LiverMss. Liver lass lesion being not high signal on T2WI, being iso-signal on DDVD, being not high signal on SDC, being not low signal on ADC, and the existence of stellate scar had ORs of 49.1, 45.8, 30, 8.5, and 13.3, respectively, favoring the diagnosis of FNH. A total of 69.2% (9/13) of the FNH had LiverMss-FNH ≥4.0, while the remaining 4 cases (30.8%) all had a LiverMss-FNH of 3.0. A total of 89.3% (50/56) of the MT had LiverMss-FNH ≤1.5. Conclusions:Liver FNH tend to have lower DDVD signal and lower SDC signal than liver MT. A LiverMss ≥4 can strongly suggest the diagnosis for a liver mass being FNH, and while a LiverMss-FNH ≤1.5 can strongly suggest the diagnosis for a liver mass being MT.

Ovarian cancer (OC) remains one of the most lethal gynecologic malignancies often resistant to immune checkpoint blockade (ICB) due to poor infiltration of cytotoxic CD8⁺ T cells and type 1 conventional dendritic cells (cDC1s) into the tumor microenvironment (TME). To overcome this, we engineered magnetic resonance imaging (MRI)-visible mesenchymal stem cells (MSCs) to co-express interleukin-15 (IL-15) for T-cell activation and XC motif chemokine ligand 1 (XCL1) for cDC1 recruitment, aiming to remodel the TME and enhance therapeutic outcomes. MSCs were engineered via lentiviral transduction to stably express Il15, Xcl1, and a ferritin reporter for MRI tracking. In vitro validation included assays for gene expression, cytokine secretion, T-cell proliferation, and DC migration. Therapeutic efficacy was evaluated in subcutaneous (ID8) and disseminated (intraperitoneal FLUC-eGFP-ID8 and intra-omental FLUC-eGFP-OVHM) murine OC models. Mice received peritumoral (subcutaneous model) or intraperitoneal (disseminated models) injections of engineered MSCs (1×107 cells). Anti-PD-1 antibody (10 mg/kg, twice weekly) was administered intraperitoneally in disseminated models. Tumor progression was monitored by MRI, bioluminescence imaging, and survival analysis. Immune cell infiltration and phenotypes were assessed using flow cytometry, qPCR, immunofluorescence, and immunohistochemistry. Engineered MSCs sustainably secreted IL‑15 and XCL1, enhancing T cell proliferation and cDC1 migration in vitro. In vivo MRI confirmed efficient MSCs homing to subcutaneous tumors, suppressing tumor growth. In disseminated models, multi-armored MSC therapy inhibited tumor progression and prolonged survival, with combination therapy achieving superior outcomes. Mechanistically, this treatment drove a robust infiltration of CD8⁺ T cells and cDC1s into the TME. Flow cytometry revealed a beneficial shift in the CD8⁺ T-cell compartment toward progenitor-like, proliferative, and effector phenotypes. Furthermore, tumor-infiltrating cDC1s displayed elevated expression of co-stimulatory molecules CD80 and CD86, indicating enhanced activation. MRI-visible MSCs co-expressing IL-15 and XCL1 effectively target ovarian tumors and remodel the immune microenvironment to foster potent anti-tumor immunity. By recruiting activated cDC1s and promoting durable, functional CD8⁺ T-cell responses, these multi-armored MSCs synergize with ICB to overcome therapeutic resistance. This cellular immunotherapy represents a promising strategy for ICB-resistant OC and warrants clinical translation.

Automated segmentation of liver lesions on non-contrast computed tomography (NCCT) is clinically important but fundamentally challenging, particularly in low-resource settings across Africa and Asia where contrast agents are frequently unavailable. Progress has been limited by the absence of annotated NCCT benchmarks. Here we describe the TriALS challenge for automated liver lesion segmentation under contrast-limited conditions, supported by a multi-centre dataset of 150 cases with four-phase CT acquisitions (600 volumes) from Egyptian and Chinese institutions. Algorithms were evaluated on 70 cases from three institutions, including an independent external cohort. The top-performing method achieved a mean venous-phase Dice of 0.754, consistent with human-level performance, yet dropped to 0.57 on NCCT. On external validation, the leading method outperformed off-the-shelf models by up to 28

Intestinal drug delivery is a crucial route for rheumatoid arthritis (RA) therapy. However, its effectiveness is often hampered by the viscosity gradient of the mucus layer and the selective degradation and efflux functions of the epithelial barrier. To address these challenges, we developed a nano-microsphere system featuring charge- and stiffness-tunable multilayered vesicles (MLVs) encapsulated within pH-sensitive microspheres. The MLVs are engineered to traverse the negatively charged, viscosity-gradient mucus by sequentially shedding their flexible shells and undergoing a positive-to-negative charge reversal. This exposes a rigid, neutral core that enables multimechanistic endocytosis with potential for transcellular transport. The dynamic tunability of the MLVs is attributed to the incorporation of di-artesunate-phosphatidylcholine (DAPC) in the vesicle shell. LC-MS/MS analysis reveals that DAPC undergoes terminal hydrophobic chain carboxylation via a non-classical hydrolysis pathway, facilitating the observed charge reversal. Encapsulation within pH-sensitive microspheres further protects the MLVs from premature degradation and ensures targeted intestinal delivery. Cryo-electron microscopy and in vitro studies confirmed the multilayered architecture, dynamic adaptability, and effective penetration of intestinal barriers by the MLVs and the nano-microspheres. In vivo, this system achieved a 1.6-fold increase in maximum blood concentration compared to conventional carriers, alongside significantly enhanced therapeutic efficacy for RA. In summary, we present a dynamically adaptive, intestinal barrier-penetrating nano-microsphere platform that offers a promising strategy for RA treatment.

Two‑photon (TP) excited fluorescence has emerged as a key technology for high‑resolution imaging in the life sciences, owing to its advantages of near‑infrared excitation, deep‑tissue penetration, minimized light scattering, and low phototoxicity. Driven by the growing demand for integrated diagnosis and therapy in precision medicine, the combination of TP imaging with controlled drug delivery and multimodal synergistic therapy has become a leading research frontier. Metal‑organic frameworks (MOFs) and metal‑organic cages (MOCs), characterized by their structurally tunable architectures, functional pores/cavities, and excellent photophysical properties, provide ideal platforms for constructing high‑performance two‑photon theranostic agents. This review provides a systematic overview of two-photon luminescent MOF/MOC (TP‑MOF/MOC) systems, covering the design and assembly strategy, photophysical properties, and biomedical applications containing bioimaging, biosensing, drug delivery and therapy, which highlights how rational structural engineering optimizes their TP luminescence to achieve multifunctionality and deeper tissue penetration for advanced biomedical applications including tumor diagnostics, neuroscience, and microenvironment‑responsive therapy, thereby accelerating potential clinical translation of TP‑MOF/MOC system.

Background:Focal nodular hyperplasia (FNH) and liver cancers are commonly differentiated by contrast enhanced scan, particularly with the application of hepatobiliary-specific contrast agents. This study aims to investigate the diffusion-derived vessel density (DDVD) difference between liver FNH and liver malignant lesions [hepatocellular carcinoma (HCC) and metastasis]. Methods:The liver diffusion-weighted magnetic resonance imaging (MRI) dataset-1 had 8 cases of FNH, 56 cases of HCC, and 14 cases of liver metastases. Liver diffusion MRI dataset-2 had 10 cases of FNH, 78 cases of HCC. For dataset-1, DDVDb10 and DDVDb20 were calculated from b=0 and b=10 s/mm2 images, b=0 and b=20 s/mm2 images, respectively. For dataset-2, the measurement was conducted on b=0, b=2, and b=10 s/mm2 diffusion-weighted imaging (DWI) images. The ratios of lesion to adjacent liver tissue were taken as: DDVD ratio (DDVDr) = lesion DDVD/liver DDVD. For semi-quantitative analysis on b=0 s/mm2 DWI image and DDVD map, relative to the adjacent liver signal, a liver lesion signal was assigned to five categories: low signal, iso-signal, slightly high signal, high signal, and markedly high signal. Results:FNH tended to have a lower DDVDr value than malignant lesions, both for dataset-1 (mean DDVDrb10 value, FNH: 1.672, HCC: 5.807, metastases: 7.944) and dataset-2 (mean DDVDrb2 value, FNH: 1.141 HCC: 3.340). For dataset-1, DDVDrb10 had an area under receiver operating characteristic curve (AUROC) of 0.864, and a cutpoint value of >1.923 had a sensitivity of 81.4% and a specificity of 87.5% in suggesting malignancy. For dataset-2, DDVDrb2 had an AUROC of 0.912, and a cutpoint value of >1.845 had a sensitivity of 79.7% and a specificity of 90% in suggesting malignancy. Consistent with quantitative measurement, semi-quantitative scoring showed that a drop from DWI high signal or slightly high signal to DDVD iso-signal suggested the diagnosis of FNH. Dataset-1 showed metastases had a higher DDVD signal than HCC, with markedly high signal on both DWI and DDVD map favoring the diagnosis of metastases. Conclusions:FNH has a lower DDVD measure compared to HCC and Mets. A drop from DWI high signal or slightly high signal to DDVD iso-signal suggests the diagnosis of FNH.

BackgroundThe multiparametric magnetic resonance imaging (mpMRI)-based Prostate Imaging for Recurrence Reporting (PI-RR) system has been proposed to evaluate local recurrence in patients with prostate cancer (PCa) who have been treated with radiation therapy (RT) or radical prostatectomy (RP).PurposeTo evaluate the diagnostic performance and interreader agreement of the PI-RR system in the diagnosis of locally recurrent PCa remains.Material and MethodsA total of 110 patients who have biochemically recurrent PCa after RT (n = 35) or RP (n = 75) were included in this retrospective study. All patients underwent mpMRI, PSMA-PET/CT, and biopsy. Four radiologists with varying levels of expertise independently assessed the local recurrence of PCa using PI-RR. The reference standard was the biopsy pathology. The receiver operating characteristic (ROC) curve was used to evaluate the performance of PI-RR and PSMA-PET/CT, and areas under the ROC curve (AUC) were calculated. Interreader agreement across four readers was evaluated using the intraclass correlation coefficient (ICC).ResultsAmong 110 patients with biochemically recurrent PCa, 28 had local recurrence and 82 had no local recurrence. Using a cutoff of 4, the AUCs of PI-RR in the diagnosis of local recurrence were in the range of 0.61-0.84 in patients treated with RT and 0.71-0.89 in patients treated with RP. The ICC was 0.86 (95% confidence interval = 0.81-0.91).ConclusionPI-RR using a cutoff of 4 has a favorable diagnostic performance and interreader agreement, which might be alternatively used for detecting local recurrence in patients with biochemically recurrent PCa treated with RT or RP.

Several medical Multimodal Large Languange Models (MLLMs) have been developed to address tasks involving visual images with textual instructions across various medical modalities, achieving impressive results. Most current medical generalist models are region-agnostic, treating the entire image as a holistic representation. However, they struggle to identify which specific regions they are focusing on when generating a sentence.To mimic the behavior of doctors, who typically begin by reviewing the entire image before concentrating on specific regions for a thorough evaluation, we aim to enhance the capability of medical MLLMs in understanding anatomical regions within entire medical scans.To achieve it, we first formulate \textbf{Region-Centric tasks} and construct a \textbf{large-scale dataset, MedRegInstruct,} to incorporate regional information into training. Combining our collected dataset with other medical multimodal corpora for training, we propose a \textbf{Region-Aware medical MLLM, MedRegA}, which is the first bilingual generalist medical AI system to simultaneously handle image-level and region-level medical vision-language tasks across a broad range of modalities. Our MedRegA not only enables three region-centric tasks, but also achieves the best performance for visual question answering, report generation and medical image classification over 8 modalities, showcasing significant versatility. Experiments demonstrate that our model can not only accomplish powerful performance across various medical vision-language tasks in bilingual settings, but also recognize and detect structures in multimodal medical scans, boosting the interpretability and user interactivity of medical MLLMs. The codes and model will be made publicly available.

Breast cancer is one of the most common malignancies among women globally. Magnetic resonance imaging (MRI), as the final non-invasive diagnostic tool before biopsy, provides detailed free-text reports that support clinical decision-making. Therefore, the effective utilization of the information in MRI reports to make reliable decisions is crucial for patient care. This study proposes a novel method for BI-RADS classification using breast MRI reports. Large language models are employed to transform free-text reports into structured reports. Specifically, missing category information (MCI) that is absent in the free-text reports is supplemented by assigning default values to the missing categories in the structured reports. To ensure data privacy, a locally deployed Qwen-Chat model is employed. Furthermore, to enhance the domain-specific adaptability, a knowledge-driven prompt is designed. The Qwen-7B-Chat model is fine-tuned specifically for structuring breast MRI reports. To prevent information loss and enable comprehensive learning of all report details, a fusion strategy is introduced, combining free-text and structured reports to train the classification model. Experimental results show that the proposed BI-RADS classification method outperforms existing report classification methods across multiple evaluation metrics. Furthermore, an external test set from a different hospital is used to validate the robustness of the proposed approach. The proposed structured method surpasses GPT-4o in terms of performance. Ablation experiments confirm that the knowledge-driven prompt, MCI, and the fusion strategy are crucial to the model's performance.

Vision foundation models have demonstrated vast potential in achieving generalist medical segmentation capability, providing a versatile, task-agnostic solution through a single model. However, current generalist models involve simple pre-training on various medical data containing irrelevant information, often resulting in the negative transfer phenomenon and degenerated performance. Furthermore, the practical applicability of foundation models across diverse open-world scenarios, especially in out-of-distribution (OOD) settings, has not been extensively evaluated. Here we construct a publicly accessible database, MedSegDB, based on a tree-structured hierarchy and annotated from 129 public medical segmentation repositories and 5 in-house datasets. We further propose a Generalist Medical Segmentation model (MedSegX), a vision foundation model trained with a model-agnostic Contextual Mixture of Adapter Experts (ConMoAE) for open-world segmentation. We conduct a comprehensive evaluation of MedSegX across a range of medical segmentation tasks. Experimental results indicate that MedSegX achieves state-of-the-art performance across various modalities and organ systems in in-distribution (ID) settings. In OOD and real-world clinical settings, MedSegX consistently maintains its performance in both zero-shot and data-efficient generalization, outperforming other foundation models. MedSegX is a vision foundation model for open-world medical image segmentation, and its accompanying dataset covers a large number of segmentation tasks across 39 organs and tissues.

Prediction of axillary lymph node metastasis in breast cancer is a critical factor in determining the prognosis of breast cancer patients and the necessity of axillary intervention. However, current research methods are limited to single-parameter MRI sequences, only use images from the primary tumor area or the axillary lymph node region, and the methods for achieving feature fusion pay relatively little attention to modal interactivity. To address these, we propose a multi-modal fusion model that uses a Cross-Sequence Module to capture relationships between different sequences, while the interaction between the primary tumor and axillary lymph node regions of interest (ROIs) is achieved through a module we refer to as the Dual-ROI Dependency Module, in order to predict axillary lymph node metastasis. Experimental results demonstrate that this method achieves strong predictive performance (ACC=0.887, AUC=0.931), outperforming single-parametric sequence models, single-region models, and models based on simple feature concatenation, validating the superiority of our approach.