Find the papers that actually matter

To predict cartilage tumor malignancy from radiographic images combined with readily available non-imaging information based on a vision-language foundation model. This single-institution study assembled a dataset of 3336 radiographs from 455 patients with enchondroma or chondrosarcoma that was assembled from two sources: (1) patients with histopathology-confirmed diagnoses of enchondroma or chondrosarcoma, and (2) patients with imaging-stable enchondroma without biopsy, confirmed through long-term imaging follow-up. An adapted vision-language foundation model based on the pre-trained CLIP (Contrastive Language-Image Pretraining) architecture was fine-tuned with our proposed Medical Knowledge Adapters and evaluated using 10-fold patient-level cross-validation to predict cartilage tumor malignancy from plain radiographs and demographic information. Using radiographs alone, the model achieved an Areas Under the receiver operating characteristic Curve (AUC) of 0.91 ± 0.04. Incorporating demographics improved the AUC to 0.94 ± 0.02. Subgroup analysis demonstrated robust generalizability across tumor grades with an AUC of 0.91 ± 0.07 in distinguishing atypical cartilaginous tumors (ACT) previously known as low grade chondrosarcomas, and 0.95 ± 0.02 in differentiating high-grade chondrosarcomas from enchondromas. Within the clinically challenging extremity subgroup (enchondroma vs ACT/LGCS), the model achieved an AUC of 0.79 ± 0.14, reflecting diagnostic difficulty observed in clinical practice. This foundation model-based approach demonstrates strong performance using accessible data sources, offering a non-invasive, cost-efficient, and scalable solution for cartilage tumor assessment in musculoskeletal oncology.

Brain MR imaging at 10.5 T ultra-high field offers significant improvements in signal-to-noise ratio (SNR), but faces challenges with B1 + inhomogeneity. Parallel-transmission (pTx) can be used to achieve a more uniform RF field distribution, but necessitates the use of B1 + calibration in the region of interest. This study explores a universal B1 + shim solution on 10.5 T that could eliminate the need for time-consuming subject-specific B1 + calibration. B1 + data from 7 participants (19 sessions) were used to develop the universal B1 + shim, which was then validated against traditional subject-specific approaches using T1-weighted MP2RAGE structural images in 5 participants (6 sessions). Statistical comparisons of tissue and subcortical segmentations were conducted using popular neuroimaging tools SPM and FreeSurfer, respectively. The universal shim rapidly converged with a small training dataset, likely due to consistent positioning and the simplicity of B1 + shimming used for head imaging. Whole-brain tissue segmentation showed no statistically significant differences between universal and subject-specific solutions, with only minor variations near the ventricles and inferior brain regions in the detailed subcortical segmentation. The proposed universal B1 + shim reduces examination time by removing the need for separate data acquisition and optimization. These findings suggest that the universal B1 + shim is a viable substitute for subject-specific approaches, offering a more efficient solution for neuroimaging applications. Additionally, it confirms that 10.5 T MRI can produce reliable structural brain imaging data, paving the way for broader adoption of ultra-high field MRI in neuroimaging research.

Clathrin is a key structural protein in intracellular vesicle transport, mainly mediating clathrin-mediated endocytosis (CME) through trimeric assembly. Its functional abnormalities are closely associated with various diseases, including neurodegenerative disorders, tumor metastasis, and immune system dysregulation. Traditional experimental methods for identifying the presence of Clathrin have limitations such as high cost and time consumption. Therefore, it is particularly urgent to develop efficient and reliable computational methods to assist in Clathrin recognition. In this study, we propose a model named ClathPLM, which integrates sequence embeddings from three pre-trained protein language models (PPLMs), i.e., ProtT5, ProtBert, and ESM-3, and performs deep representation learning on each feature through an independent branch composed of a convolutional neural network (CNN) and a multi-head Attention (MHA) mechanism, finally fusing the representations of the three views to accomplish the classification task. To validate the effectiveness of this design, we further examined variants of the fusion strategy and attention mechanism. Evaluation results show that ClathPLM demonstrates excellent overall classification performance and robustness, surpassing current state-of-the-art methods. Moreover, the model performs strongly on an additional case-study dataset and shows good scalability on an extra vesicular transport proteins (VTPs) dataset. We anticipate that ClathPLM may contribute to a deeper understanding of the role of Clathrin in cellular regulation and disease mechanisms, and facilitate future biological studies as well as potential clinical applications.

Nanozymes are engineered nanomaterials designed at the atomic scale to fine-tune their structure, composition, and electronic properties, thereby creating active sites that mimic those of natural enzymes. Among these materials, Metal-organic frameworks (MOFs) are notable for their well-defined, porous frameworks, which are created by connecting metal ions or clusters with organic linkers. Their large surface area, adjustable porosity, and superior biocompatibility enable excellent catalytic activity. By containing specific catalytic functionalities, MOF-based nanozymes can mimic peroxidase, oxidase, catalase, and superoxide dismutase activities. These properties make them genuinely promising for biomedical applications, particularly in cancer diagnosis and therapy. Recent progress in synthetic design, post-synthetic modification, and machine learning-assisted optimization has enhanced their structural precision and catalytic efficiency. Furthermore, MOFs serve as multifunctional therapeutic platforms capable of supporting combined treatment strategies and producing synergistic therapeutic effects, thereby establishing their potential as next-generation systems for targeted cancer treatment and diagnostic integration.

Evaluate how small lesion filtering and different intersection-over-union (IoU) thresholds influence deep learning segmentation of Ga-68-PSMA-11 PET images in prostate cancer. A 3D U-Net was trained on 115 patient scans with manual contours as ground truth. Performance was assessed at voxel, lesion and patient levels. Lesions less than 8 voxels (195 mm 3 ) or less than 27 voxels (658 mm 3 ) were optionally discarded, and lesion-level metrics were computed across 10-50% IoU thresholds. Excluding lesions less than 27 voxels increased voxel-level Dice from 0.7975 to 0.8173 and improved precision. Lesion-level metrics were stable across IoU thresholds of 20-40% after 27-voxel filtering, while sensitivity declined at higher overlap thresholds. Patient-level sensitivity and positive predictive value reached 96.6 and 94.5%, respectively. Implementing small lesion filtering criteria improves the reliability of automated segmentation outputs, particularly for quantitative evaluation. For lesion-level metrics, defining true positives within the range of 20-40% IoU thresholds is optimal. Further validation through multicenter studies and larger datasets is essential to ensure the generalizability of these findings.

Matrix metalloproteinase-9 (MMP-9) is crucial for extracellular matrix remodeling, and its dysregulation is associated with inflammatory diseases and different forms of cancer. Conventional MMP-9 detection methods such as enzyme-linked immunosorbent assay (ELISA) are limited by complexity, expensive equipment, and lengthy antibody incubation times. These limitations prevent their use in point-of-care testing. An MMP-9 responsive photonic crystal (PC/PEG-M9SP) hydrogel has been developed to address these challenges. The hydrogel is synthesized from 4-arm polyethylene glycol-acrylate and an MMP-9 sensitive peptide via Michael-type addition reaction. Upon MMP-9-specific enzymatic cleavage, the hydrogel undergoes a structural reconfiguration, resulting in a distinct color shift. Integrated with a deep learning-based smartphone app, this platform enables both visual and quantitative detection within 10 min, achieving high sensitivity (10.60 nm mL/ng) and a detection limit of 0.62 ng/mL. Validation in complex biological fluids demonstrated strong concordance with ELISA, confirming the analytical reliability of the hydrogel. This system provides a rapid, portable, and cost-effective solution for accurate MMP-9 detection, with strong potential for clinical and point-of-care applications.

Contrastive language-image pretraining has greatly enhanced visual representation learning and enabled zero-shot classification. Vision-language language models (VLM) have succeeded in few-shot learning by leveraging adaptation modules fine-tuned for specific downstream tasks. In computational pathology (CPath), accurate whole-slide image (WSI) prediction is crucial for aiding in cancer diagnosis, and multiple instance learning (MIL) remains essential for managing the gigapixel scale of WSIs. In the intersection between CPath and VLMs, the literature still lacks specific adapters that handle the particular complexity of the slides. To solve this gap, we introduce MIL-Adapter, a novel approach designed to obtain consistent slide-level classification under few-shot learning scenarios. In particular, our framework is the first to combine trainable MIL aggregation functions and lightweight visual-language adapters to improve the performance of histopathological VLMs. MIL-Adapter relies on textual ensemble learning to construct discriminative zero-shot prototypes. It is serves as a solid starting point, surpassing MIL models with randomly initialized classifiers in data-constrained settings. With our experimentation, we demonstrate the value of textual ensemble learning and the robust predictive performance of MIL-Adapter through diverse datasets and configurations of few-shot scenarios, while providing crucial insights on model interpretability. The code is publicly accessible in https://github.com/cvblab/MIL-Adapter.

This study aimed to build a multimodal ultrasound (color Doppler flow imaging/shear wave elastography/contrast-enhanced ultrasound) combined with machine learning (ML) model, evaluating random forest (RF) for early ovarian malignancy diagnosis. A retrospective analysis included 130 patients (72 benign, 58 malignant) pathologically confirmed ovarian lesions. Patients were split 7:3 into training/test sets. Data included demographics, lab tests and ultrasound features (32 variables). Key predictors were selected via univariate analysis, RF-recursive feature elimination and multivariate logistic regression. Six ML models were built and evaluated with 10-fold cross-validation. Significant differences were observed between the benign and malignant groups in the training set for 24 indicators (all p < 0.005), including age, menopausal, CA125 and so on. After feature selection, five core predictors were identified: Peak intensity (PIy), maximum elasticity (Emax), CA125, human epididymis protein 4 (HE4) and internal composition. The RF model achieved area under the curves of 0.986 (training set) and 0.886 (test set), significantly outperforming other algorithms. Decision curve analysis demonstrated its highest net benefit within the 0-0.74 threshold probability range and the lowest Brier score (0.014 for training, 0.128 for test). SHapley Additive exPlanations (SHAP) analysis revealed that Emax, PIy and internal composition were the key features influencing model decisions, with the solid component having the largest impact on the malignant probability (ΔSHAP = -0.125). The multimodal ultrasound-RF model constructed in this study exhibits excellent diagnostic performance and quantifies the contribution of key features, providing a reliable imaging tool for the early and precise diagnosis of ovarian malignancies.

Reliable detection of robust biomarkers from high-dimensional transcriptomic data remains a major challenge in computational oncology. Traditional approaches often suffer from overfitting and poor generalization due to the high dimensionality of genomic data and limited sample sizes. This study aims to identify an optimal, biologically meaningful subset of mRNA biomarkers capable of distinguishing ovarian cancer samples from healthy controls using an integrated machine learning-based feature selection framework. We analyzed mRNA expression data encompassing approximately 63,000 transcripts from ovarian cancer and control samples derived from cell lines. A hybrid feature selection pipeline combining statistical filtering, recursive elimination, and regularization was implemented under stratified cross-validation to derive stable biomarkers. Model validation was performed using Logistic Regression, Random Forest, XGBoost, and Support Vector Machine classifiers, while experimental validation was conducted through droplet digital PCR (ddPCR). Statistical analyses included ANOVA, t-tests, and pathway enrichment. The pipeline identified 80 discriminative mRNA biomarkers with exceptionally high classification performance (accuracy = 1.00, sensitivity = 1.00, specificity = 1.00 for top models). ddPCR confirmed consistent expression patterns, with significant downregulation of ADAMTS12, FN1, and ABI3BP and overexpression of EPCAM, COX6C, and TMT1B in ovarian cancer. Pathway enrichment revealed involvement in DNA repair, RNA processing, protein translation, immune regulation, and metabolic reprogramming. This hybrid feature selection framework applied to patient derived cell lines, effectively reduces dimensionality, enhances biomarker reliability, and uncovers biologically interpretable mRNA signatures associated with ovarian cancer, demonstrating potential for diagnostic and therapeutic applications.

Fatty acid synthase (FASN), which is highly expressed in multiple cancers, contributes critically to cancer cell survival, proliferation, and metastasis, rendering it a promising target for therapeutic intervention. To develop novel and efficient FASN inhibitors, a series of lipophilic amide fragments were introduced into the natural inhibitor mangiferin (MGF) to synthesize new MGF derivatives. Among these derivatives, compound 4 demonstrated notable antiproliferative activity against human hepatocellular carcinoma cell lines with high FASN expression. In particular, 4 exhibited the most potent activity against Hep-G2 cells (IC50 = 0.47 ± 0.06 μM), demonstrating 185-fold greater potency than MGF (IC50 = 87.24 ± 2.06 μM). The capability to bind to FASN and inhibit its activity was significantly stronger than that of MGF. Further investigations revealed that 4 was involved in blocking the activation of PI3K/AKT pathway, thereby inducing reactive oxygen species production and promoting cancer cells apoptosis. Moreover, 4 exhibited a high selectivity index toward Hep-G2 cells (SI = 260.00) and inhibited the migration and invasion of Hep-G2 cells. These findings may serve as a valuable reference for the development of novel FASN inhibitors exhibiting potent anti-hepatocellular carcinoma activity.