Find the papers that actually matter

Breast cancer's (BC) diverse nature and global impact demand tailored clinical strategies. Conventional screening methods often fall short in early detection and individualized risk assessment. By merging multi-omics technologies such as genomics, transcriptomics, proteomics, and metabolomics with artificial intelligence (AI), clinicians gain powerful tools to navigate this complexity. AI's ability to analyze vast, intricate multi-omics datasets enables precise risk stratification, early diagnosis, and the development of customized treatment plans. Applications range from refining mammographic analysis and forecasting therapy outcomes to uncovering novel biomarkers. However, barriers such as data standardization, model applicability across diverse patient groups, and AI interpretability limit clinical integration. This review provides clinicians with a comprehensive guide to current advances in multi-omics profiling, including genomics, transcriptomics, proteomics, and metabolomics, as well as their integration through AI-driven models to decode tumor heterogeneity and predict treatment response. We discuss cutting-edge computational frameworks, challenges in data integration, and clinical applications that enhance prognostic accuracy and facilitate precision oncology approaches. By embracing the convergence of multidimensional molecular data and AI, clinicians can deliver individualized BC care that optimizes therapeutic outcomes and advances the post-genomic era of oncology.

Cell mechanics, elasticity and viscoelasticity, are key markers of biological states like cancer. Atomic force microscopy (AFM) is ideal for such studies, but its low throughput limits large-scale use. Two solutions exist: automation for higher throughput, or high-density measurements for richer data. The latter enables machine learning (ML)-based classification, with viscoelastic parameters offering unique insights beyond static measures like Young's modulus. This study used dynamic mechanical analysis (DMA) to classify cells, focusing on viscoelastic descriptors (storage/loss moduli) across frequencies. Normal (RWPE-1) and grade IV cancerous (PC3-GFP) prostate cells were probed at 1-200Hz, generating 304 features per cell. The fuzzy logic-based LAMDA algorithm, trained on 19 selected features, classified cells using 40 samples per line. PC3-GFP cells showed higher deformability and heterogeneity, behaving more like viscous fluids at low frequencies. The model achieved 79% classification accuracy. Adding features improved performance, suggesting fewer training samples may suffice with rich datasets. A sensitivity-optimized threshold reduced false negatives in cancer detection. Combining viscoelastic analysis with ML effectively discriminates normal and malignant cells. Future work could refine training and integrate new features, though acquisition time remains a challenge. This approach offers a promising framework for mechanome-based diagnostics, with applications in cancer and stem cell research.

The Ovarian-Adnexal Reporting and Data System (O-RADS) is essential for standardizing the risk stratification of ovarian lesions detected on ultrasound. However, manual assignment of O-RADS scores is time-consuming and can vary between observers. This study investigates an automated method for O-RADS scoring using a large language model (LLM) to analyze narrative ultrasound reports. A two-stage pipeline was developed for automated O-RADS classification. Initially, the Lingshu LLM, specialized in medical language, extracted and embedded features from free-text descriptions of ovarian lesions. It identified key diagnostic features mentioned by sonologists. Subsequently, these features were used to train and evaluate several machine learning algorithms, including logistic regression (LR), support vector machines and random forests, to predict O-RADS scores (1-5). The proposed method was evaluated on a dataset of 513 cases using fivefold cross-validation. The pipeline using Lingshu model embeddings with LR achieved the highest accuracy of 0.803 [95% CI: 0.753, 0.853], a weighted-average F1-score of 0.819 [95% CI: 0.777, 0.861] and a macro-averaged AUROC of 0.948 [95% CI: 0.937, 0.959]. This outperformed the MedGemma model's pipeline, which had an accuracy of 0.760 [95% CI: 0.700, 0.820], F1-score of 0.787 [95% CI: 0.739, 0.835] and AUROC of 0.941 [95% CI: 0.911, 0.971]. This study introduces a novel approach to automate O-RADS scoring using LLMs for feature extraction and traditional machine learning for classification. The results indicate that this method can accurately stratify ovarian cancer risk, potentially improving clinical workflow efficiency and reducing diagnostic variability. This approach may support radiologists in making more consistent and timely assessments.

The precise prediction by MSI plays a key role in the perioperative treatment and prognosis of colorectal cancer (CRC) patients. This study seeks to establish an interpretable deep learning radiomics model using enhanced CT images to improve the preoperative prediction of microsatellite instability (MSI) in CRC. The retrospective study analyzed 873 CRC patients who received curative surgery at three medical centers. This group was separated into a training cohort (Center 1), an internal validation cohort (Center 1), external validation cohort 1 (Centers 2) and external validation cohort 2 (Centers 3). By processing the pre-operative portal venous phase CT enhanced images, deep learning as well as radiomics features was derived and combined with body composition based-clinical risk factors to develop three models to predict the MSI status, namely the deep learning radiomics model (DLR), the clinical model, and the clinical model combining deep learning and radiomics (CDLR), with the use of the random forest algorithm. Model performance was quantified by the areas under the receiver operating characteristic curves through the DeLong test, while calibration and decision curve analyses (DCA) were applied to estimate the potential clinical benefit of the models. Finally, SHAP (Shapley Additive exPlanations) analysis and Gradient-weighted Class Activation Mapping (Grad-CAM) was utilized to illustrate the interpretability and visualizability of the model. Compared with the other models, the CDLR model exhibited superior performance achieving area under the curves (AUCs) of 0.882 in training cohorts, 0.768 in internal validation cohorts, and 0.803 and 0.751 in external validation cohorts 1 and 2, respectively. The clinical model yielded AUCs of 0.730, 0.683, 0.670 as well as 0.607 across the corresponding cohorts and the DLR achieved AUCs of 0.841, 0.740, 0.779, and 0.712. DCA indicated that the CDLR model provided the greatest clinical benefit in MSI prediction in the external validation cohorts. In summary, the interpretable CDLR fusion model based on enhanced CT demonstrates promising potential as a noninvasive pre-screening tool for predicting MSI in CRC, supporting individualized treatment strategies, while further prospective validation is needed before routine clinical adoption.

Differentiating T1-stage nasopharyngeal carcinoma (NPC) from benign hyperplasia (BH) is challenging. This study aims to construct and validate a multimodal model combining magnetic resonance imaging (MRI) and endoscopy to distinguish T1-NPC from BH. Additionally, SHapley Additive exPlanations (SHAP) are used for model interpretability analysis. Data from 161 patients with histologically confirmed diagnoses between 2015 and 2022 were retrospectively collected, including 95 cases of T1-NPC and 66 cases of BH. Regions of interest (ROI) were drawn based on MRI and endoscopy to extract features. Feature selection techniques, such as elastic net, recursive feature elimination, and deep learning, were used to identify the optimal feature subset. Naive Bayes, Adaptive Boosting (AdaBoost), Light Gradient Boosting Machine (LightGBM), k-Nearest Neighbors (kNN), and Multilayer Perceptron (MLP) were applied to establish the MRI radiomics model and the MRI-endoscopy combined model. SHAP was used to perform interpretability analysis of the models. The MRI-endoscopy combined model outperformed the radiomics model, with the MLP-based model showing the best performance. The mean AUC of the test set reached 0.98, with an accuracy of 0.90, precision of 0.90, sensitivity of 0.93, and specificity of 0.86. SHAP analysis revealed that texture features (including GLSZM, GLCM, and GLRLM) and first-order features were critical for distinguishing T1-NPC from BH. Compared to traditional radiomics methods, the multimodal model combining MRI and endoscopy more accurately distinguishes between benign and malignant tissues. SHAP enables visualization of feature contributions and model predictions, highlighting the model's clinical potential.

Diabetic retinopathy (DR) is a leading cause of vision impairment worldwide. Optical coherence tomography (OCT) and OCT angiography (OCTA) provide detailed retinal imaging, enabling early detection of microvascular changes. This study aims to systematically review artificial intelligence (AI), particularly deep learning (DL), applications for DR detection and analysis using OCT and OCTA images. A comprehensive literature search was conducted across PubMed, Web of Science, Scopus, IEEE Xplore, and Embase for studies published up to March 2026. A total of 1007 articles were identified, of which 175 studies met the inclusion criteria following the PRISMA study selection process. DL-based approaches consistently demonstrated superior performance compared to traditional machine learning (ML) methods, with reported AUC values typically ranging from 0.90 to 0.99 across classification and segmentation tasks. Convolutional neural networks (CNNs), Vision Transformers (ViTs), and encoder-decoder architectures such as U-Net showed strong performance in detecting key DR biomarkers, including microaneurysms, macular edema, and neovascularization. However, performance variability was observed depending on dataset size, imaging modality, and annotation quality. AI-driven analysis of OCT and OCTA images offers significant potential for automated DR detection. Despite promising results, challenges such as limited public datasets, lack of cross-institutional validation, and model interpretability remain. Future research should focus on multimodal integration, explainable AI, and large-scale validation to enhance clinical applicability.

Accelerated golden-angle radial acquisitions are widely used for dynamic MRI, but compressed sensing (CS)-based reconstruction presents residual artifacts at high acceleration and long computation times. Supervised deep learning (DL) enables fast reconstruction with improved image quality but usually relies on CS because fully-sampled references are unavailable. The proposed SELFIE (SElf-supervised Learning for Fast dynamIc golden-anglE radial MRI) technique enables self-supervised reconstruction that neither requires fully-sampled nor CS training references. SELFIE operates slice-by-slice in the image domain and thus avoids computationally expensive k-space data-consistency operations, allowing for a single forward-pass inference. Self-supervision is achieved by leveraging two properties of dynamic imaging with golden-angle radial sampling: (i) variable temporal resolution to form data-derived self-references at multiple temporal resolutions and (ii) temporal sparsity. SELFIE was evaluated on dynamic contrast-enhanced (DCE)-MRI in patients with gynecologic cancer and on motion-resolved free-breathing abdominal MRI in patients with liver cancer. Reconstructions were compared against CS and supervised DL using quantitative image-quality metrics and qualitative assessment from radiologists. Across both applications, SELFIE achieved image quality and dynamic fidelity (contrast enhancement and motion depiction) comparable to CS and supervised DL, while reconstructing 3D time series in seconds per case, substantially faster than CS. Ranking analysis and reader study favored SELFIE over CS and were comparable to supervised DL, with no statistically significant differences. Overall, SELFIE provides a fast, reference-free reconstruction framework for dynamic golden-angle radial MRI with competitive performance, representing a viable alternative to existing methods.

Radiomic studies in lung cancer have primarily analyzed single-domain features extracted separately from intranodular (Zone-1) or perinodular (Zone-2) regions, potentially overlooking their biological interdependence. We developed MUSIOMICS (Multiregional Unified and Spatially Integrated Oncologic Model for Imaging-based Connected Structures), a multi-region radiomic framework, and constructed two region-dependent delta-radiomic models (Delta-1 and Delta-2). Their performance was evaluated and validated in classifying primary versus metastatic pulmonary nodules. A total of 443 malignant pulmonary nodules (training set, n = 360; test set, n = 83) were retrospectively analyzed. Zone-1 and Zone-2 vol were delineated using LIFEx software. The MUSIOMICS framework was applied to construct two spatial delta-radiomic models that extracted features from both zones of different biological roles (e.g. Zone-1 and Zone-2). These features of different strengths were then fused into a single effective delta-feature. Predictive models were developed in the training dataset using a two-stage feature selection strategy and three classifiers (Random Forest, AdaBoost, and Support Vector Machine [SVM]). A two-sample t-test was applied to both the training and independent test datasets to identify reproducible statistically significant (RSS) delta-features. SHapley Additive exPlanations (SHAP) analysis was performed to rank feature importance and identify informative delta-features. RSS and informative delta-features together were combined to characterize the spatial delta-radiomic models. In the independent test dataset, spatial delta-radiomic models (Delta-1: 82%, Delta-2: 81%) outperformed single-region models (Zone-1: 75%, Zone-2: 67%), producing a 6-15% improvement in predictive accuracy. Among all classifiers, Delta-1 combined with SVM achieved the highest performance (accuracy, 86%; area under receiver operating characteristic curve, 0.90). The t-test identified 58 and 48 RSS delta-features for Delta-1 and Delta-2, respectively, with 40 overlapping across both models. Among classifiers, GLCM_DV and Intensity_QCD were consistently top-ranked contributors in Delta-1, with Intensity_QCD identified as the most informative feature in both Delta-1 and Delta-2. Spatial delta-radiomics based on MUSIOMICS integrates complementary information from biologically connected intranodular and perinodular compartments, achieving higher and more reproducible predictive performance than conventional single-zone models for characterizing malignant pulmonary nodules.

To evaluate the feasibility and limitations of real-world, text-only inference of PI-RADS v2.1 categories from prostate MRI reports using large language models, with lesion-level and zone-aware analysis. This single-center retrospective study included 1,205 lesion-level entries from 1,118 patients derived from semi-structured prostate MRI reports after removal of all explicit PI-RADS elements. ChatGPT-4o was prompted to assign numeric PI-RADS categories based solely on report text. Agreement with radiologist-assigned reference categories was assessed using exact agreement, Cohen's κ, and class-wise metrics. Analyses were performed overall, by zone (peripheral vs transition), and using collapsed risk strata (1-2/3/4-5). Discordant cases were reviewed to identify error mechanisms and severity. Human interobserver agreement, intra-model reproducibility, temporal stability, and a paired model-version sensitivity analysis comparing ChatGPT-4o with GPT-5.2 were also evaluated. Overall exact agreement was 72.9% (κ = 0.538; macro-F1 = 61.2%), with a systematic tendency toward overcalling. Agreement was higher in the peripheral zone than in the transition zone (κ = 0.476 vs 0.077, reference PI-RADS 3-5). PI-RADS 3 showed the lowest precision and recall, with frequent bidirectional misclassification. Collapsing categories improved agreement (κ = 0.610). Incorrect diffusion-weighted imaging subscores were the most common error mechanism, with zone-specific differences. Clinically high-impact downgrades of PI-RADS 4-5 to 1-2 were rare (1.6%). Human interobserver agreement was excellent (κ = 0.916-0.967). GPT-5.2 outperformed ChatGPT-4o in paired analyses but produced invalid outputs in a minority of cases. Text-only large language models can infer radiologist-assigned PI-RADS v2.1 categories from real-world prostate MRI reports with moderate agreement, but performance is zone dependent and limited around PI-RADS 3, particularly in the transition zone. These models are best suited as supervised tools for quality control rather than autonomous decision-making.

Wee1-like protein kinase (Wee1) is a validated target for cancer therapy, though no inhibitors have yet received regulatory approval. To identify new Wee1 inhibitors, we conducted a virtual screening campaign using the Eurofins-Villapharma library. Based on the initial hit Compound 1 (IC50 = 240 nM in the biochemical assay), we synthesized and tested additional compounds. Analog 20 showed a significant improvement in potency (IC50 = 33 nM). It was also the most potent in cellular assays, inhibiting the proliferation of lung carcinoma cell lines A427 (IC50 = 0.761 μM) and NCI-H23 (IC50 = 0.449 μM), OVCAR-3 (IC50 = 0.627 μM), and MDA-MB-231 (IC50 = 0.842 μM). Compound 18 (cell-free IC50 = 55 nM) showed weaker cellular activity, with IC50 values of 1.67 μM (A427), 1.38 μM (NCI-H23), 1.58 μM (OVCAR-3), and 2.93 μM (MDA-MB-231). In a kinome screen, Compound 18 was the most selective of the tested compounds, with an S(1) score of 0.025. Structure-activity relationship analysis highlighted key activity determinants, including the effects of methyl substitution on the phenylpiperazine and of double‑chlorine halogenation on the terminal phenyl group. Molecular dynamics simulations were used to elucidate the binding interactions and conformational dynamics of the high-affinity compounds. These simulations revealed stabilizing interactions that aligned with the observed SAR trends. In vitro ADME compound profiling demonstrated favorable physicochemical properties and drug-likeness, including high solubility and metabolic stability, supporting their potential as lead candidates.