Find the papers that actually matter

Endoscopic endonasal and transcranial approaches are used for the resection of various pathological lesions in neurosurgery, especially pituitary adenomas, craniopharyngiomas, chordomas, or meningiomas. The video feed provided by endoscopes is generally two-dimensional, which can hinder depth perception. Thus, generating three-dimensional imaging without the need for special endoscopes using deep learning might be beneficial for enhanced intraoperative orientation. DINOv2 is a pre-trained deep-learning model published by Meta in 2023. One of its capabilities is to estimate the depth in two-dimensional images. In this study, we explore the application of DINOv2 to the video feed of eight transsphenoidal endonasal surgeries. The results were evaluated for quality by both a senior neurosurgeon and a resident neurosurgeon. Furthermore, depth estimations from a randomly selected subset of 488 images taken from the videos were semi-quantitatively compared against manual segmentations for the estimation of deep, intermediate, and superficial areas. Using DINOv2, numeric depth maps were generated, and colormaps were created for depth visualization. Although these colormaps were not perfect, they aligned well with the subjective assessment of depth in the video feed by a senior neurosurgeon as well as a resident neurosurgeon. Semi-quantitative validation of the model's estimations yielded a mean overall DICE Similarity Index of 0.48. These semi-quantitative results should be interpreted with caution, as the cutoffs used for model depth predictions and manual segmentation are not standardized. Through the application of DINOv2, we were able to estimate depth in endoscopic imaging from transsphenoidal endonasal surgeries by generating numeric maps and depth colormaps. This illustrates the potential of deep learning-based depth estimations, which in the future could contribute to improving intraoperative orientation. It also highlights the opportunities in using artificial intelligence to augment endoscopic video feeds.

Frailty is associated with biological aging, causing poor prognosis for cancer survivors, but has not been considered in life expectancy (LE) estimates. This study aims to estimate frailty-adjusted LE for patients aged ≥50 with stomach cancer, using routine blood tests. We included 8,281 patients aged ≥50 diagnosed with stomach cancer in 2007-2018 at National Cancer Center (NCC), extracted from electronic health records. Frailty was assessed by 27 laboratory tests (FI-Lab) and classified as non-frail and frail. The impact of frailty on mortality was estimated by cause-specific Cox regression models with age as the timescale. The area under the survival curve was used to estimate the remaining LE adjusted for frailty. Frail patients (higher FI-Lab) had 1.4-fold higher risk of cancer and non-cancer mortality and shorter LE than non-frail patients (lower FI-Lab). Frailty affected 31.2% of middle-aged populations, leading to larger LE losses than in older groups. LE losses due to frailty for males at age 50 and 85 were 7.6 and 2.0 years, slightly higher than the losses in females (6.9 and 1.8 years, respectively). The mortality risk and LE loss due to frailty were largest in localized stage, at age 50: 3.9 years for males and 3.0 years for females, compared to <2 years in regional stage and 0.5 years in distant stage for both sexes. FI-Lab-defined frailty was associated with significant LE loss, particularly in male, middle-aged survivors with early-stage stomach cancer. By capturing early signs of biological aging, FI-Lab-based LE estimates may support clinical assessment and individualized survivorship care.

A new generation of pH-responsive magnetic theranostic nanovectors (NV) has been developed to deliver doxorubicin (DOX) to triple negative breast cancer (TNBC) cells which overexpress epidermal growth factor receptor (EGFR). DOX was loaded onto functionalized NV using a pH-sensitive DOX-Fe2+ complex (hereafter called NVscFv-DOX). NVscFv-DOX consist of superparamagnetic iron oxide nanoparticles (SPIONs), labelled with DylightTM 680 fluorophore, and coated with a layer of covalently bound polyethylene-glycol (PEG) which is partially functionalized with anti-EGFR scFvs (average ratio is ≈12 scFvs per nanovector). The physico-chemical characteristics of the new nanovectors were suitable for IV injection: hydrodynamic diameter DH below 150 nm, polydispersity index below 0.3 and slightly negative surface charge (≈ -10 mV). Thanks to the functional grafted scFvs, the NVscFv-DOX were able to recognize the EGFR antigen efficiently. Using preformed DOX-Fe2+ complex which binds to the SPION surface in a pH-dependent manner, about 6.5% w/w (DOX/iron oxide) of the drug was loaded onto the NVscFv-DOX. The drug loading and release in its native form at acidic pH were characterized by surface-enhanced Raman scattering (SERS) spectroscopy. The dual fluorescent response of both DylightTM680 and that of DOX was confirmed, which is promising for theranostic use of the nanovectors. Finally, the in vitro toxicity of NVscFv-DOX on the EGFR-overexpressing TNBC cell line MDA-MB-468 was confirmed and compared to that of free DOX and NVscFv without DOX. Together, all these properties of the NVscFv-DOX are promising for their potential use as theranostic platform for TNBC treatment.

Prognostic variables play a critical role in guiding clinical treatment decisions for cancer patients. However, extracting prognostic information from gigapixel histopathology slides remains a significant challenge. While attention-based deep learning models trained on histologic images have been extensively investigated, existing approaches often fail to effectively model slide-level contextual information or demonstrate generalizability across diverse cancer types and multi-center datasets. We propose a Hypergraph-based Multi-instance Contrastive Reinforcement learning model (HeMiCoRe), which integrates cluster-restricted local features and cross-cluster global representations from 5196 H&E-stained slides across 10 cancer types, leveraging both morphological and spatial relationships. HeMiCoRe employs hypergraph neural networks to predict patient survival outcomes and achieves state-of-the-art (SOTA) performance on 8 cancer types, demonstrating superior generalization compared to existing weakly supervised methods. This framework holds promise for clinical adoption, offering a robust tool for cancer prognosis and supporting treatment decision-making.

Deep learning has advanced pathological image analysis but remains constrained by limited annotated data, especially for fine-grained diagnostic tasks such as tumor subtyping, grading, and cellularity assessment. While data augmentation alleviates this issue, existing methods are restricted to spatial manipulations that lack morphological plausibility and overlook the temporal attributes of pathological state transition. To address this gap, we propose Cellflow, the first temporal-aware generative framework for pathological image augmentation. Cellflow models pathological transition as smooth trajectories on a biological image manifold, generating intermediate states via a stair-based diffusion bridge with classifier-guided probability-flow ordinary differential equations. This design produces morphologically plausible sequences that capture both cellular details and tissue-level architecture. Evaluated on 7 diverse datasets across organs, staining modalities, and diagnostic tasks, Cellflow consistently outperforms 6 spatial augmentation methods and 4 state-of-the-art generative models, yielding improved classification performance, higher image fidelity, and preservation of temporal coherence. Quantitative cellularity analysis provides additional validation of the biological authenticity of transition sequences. By introducing temporal modeling into pathological data augmentation, Cellflow establishes a paradigm shift from spatial manipulations to biologically grounded temporal trajectories that advances robust model training, rare disease exploration, and educational simulation in computational pathology. The Code is available at https://github.com/Rowerliu/Cellflow.

To evaluate the performance of radiologists with artificial intelligence (AI)-based computer-aided detection (CAD) systems on automated breast ultrasound (ABUS) for breast non-mass lesions (NMLs). From July 2020 to April 2022, patients who underwent ABUS examinations and described NMLs were included in this retrospective study. First, we compared the performance of two AI-CAD systems for diagnosing NMLs. Then, the superior-performing was selected with 4 radiologists with different levels of experience to assess the ability to diagnose NMLs with and without AI-CAD systems. A total of 251 patients with 251 NMLs were enrolled, of which 54.2% (136/251) were benign and 45.8% (115/251) were malignant. Comparing BU-CAD and QV-CAD, the sensitivity, specificity, and area under the receiver operating characteristic curve (AUC) were 45.2% versus 64.3% (p < 0.001), 77.2% versus 64.7% (p = 0.005), and 0.63 versus 0.68 (p = 0.17), respectively. Given the clinical characteristic of NMLs having a high risk of missed diagnosis, we prioritized sensitivity over specificity in our considerations. Consequently, we selected QV-CAD for its high sensitivity and numerically superior AUC. With CAD support, the mean AUC was improved from 0.78 to 0.83 (p = 0.04) for all readers; for novice readers, the mean AUC was improved from 0.73 to 0.80 (p < 0.001); for experienced readers, there were no differences in AUC among the two reading modes 0.83 versus 0.85 (p = 0.14). Our study shows that readers can improve their diagnosis of NMLs after using AI-CAD systems, especially for novice readers.

We aimed to develop a novel cardiac magnetic resonance (CMR)-based method for quantifying myocardial synchrony and evaluate its diagnostic value in detecting myocardial dysfunction of coronary artery disease (CAD). Consecutive participants with anatomically/angiographically obstructive CAD (n = 112) and healthy participants (n = 87) undergoing CMR imaging were prospectively enrolled. Myocardial strain was analyzed using feature-tracking, and myocardial synchrony was quantified via Pearson correlation coefficients of segmental strain time series across the cardiac cycle. Machine learning models (strain-only, synchrony-only, combined) were developed and validated in an independent external cohort. Healthy participants exhibited high left ventricular myocardial synchrony (radial: 0.91 [IQR: 0.88, 0.93]; circumferential: 0.90 ± 0.04; longitudinal: 0.97 ± 0.02), significantly reduced in participants with CAD (radial: 0.84 [IQR: 0.75, 0.89]; circumferential: 0.81 ± 0.12; longitudinal: 0.90 ± 0.08), including those with preserved left ventricular ejection fraction (LVEF ≥50%) (radial: 0.86 [IQR: 0.82, 0.90]; circumferential: 0.86 ± 0.07; longitudinal: 0.91 ± 0.07), all p < 0.001. In model analysis, the combined model significantly outperformed individual models (AUC: 0.94 [95% CI: 0.89-1.00] vs. 0.84 [0.75-0.94] for strain model, p = 0.037; vs. 0.79 [0.68-0.90] for synchrony model, p = 0.001). Superiority persisted in CAD with preserved LVEF (AUC: 0.91 [95% CI: 0.83-1.00]) and external validation (AUC: 0.93 [95% CI: 0.84-1.00]). This CMR-derived approach demonstrated the high degree of left ventricular synchrony in healthy populations and significant dyssynchrony in CAD, even in those with preserved LVEF. Integrating myocardial synchrony with strain significantly enhanced CAD myocardial dysfunction detection relative to strain alone, with robust diagnostic performance maintained in CAD with preserved LVEF.

Semi-supervised learning (SSL) has achieved notable progress in medical image segmentation. To achieve effective SSL, a model needs to be able to efficiently learn from limited labeled data and effectively exploit knowledge from abundant unlabeled data. Recent developments in visual foundation models, such as the Segment Anything Model (SAM), have demonstrated remarkable adaptability with improved sample efficiency. To seamlessly harness foundation models in SSL, we propose a SAM-driven cross prompting framework with adaptive sampling and prompt consistency for semi-supervised medical image segmentation, named CPAC-SAM. Our method employs SAM's unique prompt design and innovates a cross prompting strategy within a dual-branch framework to automatically generate prompts and supervision across two decoder branches, enabling effective learning from both scarce labeled and valuable unlabeled data. To ensure the quality of prompts for unlabeled data and provide meaningful supervision in the cross prompting scheme, we propose an innovative prototype-guided grid sampling strategy with adaptive intervals to simultaneously improve the reliability of the prompt selection area and ensure both adequate prompt density and complete target coverage. We further design a novel prompt consistency regularization to reduce SAM's prompt sensitivity and to enhance the output invariance under different prompts. We validate our method on five medical image segmentation tasks, encompassing both 2D and 3D scenarios. The extensive experiments with different labeled-data ratios and modalities demonstrate the superiority of our proposed method over the state-of-the-art SSL methods, with more than 4.1% and 3.8% Dice improvement on the breast cancer segmentation task and left atrium segmentation task, respectively. Our code is available at: https://github.com/JuzhengMiao/CPAC-SAM.

Apigeninidin (APN) is a flavonoid belonging to the 3-deoxyanthocyanidin family, exhibiting diverse biological activities and representing a potential natural antitumor compound. However, the poor lipophilicity and cell membrane permeability of APN limit its bioavailability and antitumor activity. To overcome these limitations, we designed a site-specific propynylation strategy and synthesized two derivatives, APN-A and APN-B, investigating how targeted modification alters APN's antitumor activity. Comparative analysis of the physicochemical properties and bioactivities of these compounds revealed that APN-A exhibited significantly enhanced cell membrane permeability and increased anticancer activity against cervical cancer cells compared to the parent compound APN. In vitro experiments further demonstrated that APN-A can dramatically reduce the viability of cervical cancer cells, inhibited cell proliferation and migration, and synergistically potentiate the antitumor efficacy of 5-fluorouracil (5-FU). In addition, chemical proteomics enrichment analyses indicated that APN-A shows its antitumor effects primarily by targeting and inhibiting processes such as DNA replication and protein transcription-translation in cancer cells via targeting proteins such as PARP-1, EIF3J, and TCEA1. These findings provide a methodological reference and mechanistic insight for the propynyl modification of APN, and highlight its potential applications in the food industry and drug development.

Patient recruitment for clinical trials remains a major challenge, with 86% of trials failing to meet enrollment targets on time. In over 77% of cases, recruitment difficulties stem from matching problems between trials and patients. Case-Based Reasoning (CBR) offers a distinct patient-to-patient approach by determining eligibility through comparison with previously enrolled patients, yet this methodology remains underexplored in contemporary oncology trial matching despite its potential advantages. To compare the performance of two CBR approaches-random forest (RF) and target patient similarity (TPS)-in predicting patient eligibility for recent oncology clinical trials using real-world electronic health record data. We selected three breast cancer clinical trials (2019-2022) from our institutional registry. Patient data were extracted from our clinical data warehouse, including structured data (laboratory results, diagnosis codes, procedures, treatments) and unstructured clinical narratives processed using natural language processing. For each trial, we trained RF classifiers and TPS models using repeated hold-out validation (25 splits, 70/30 train-test). Performance was evaluated using discriminative metrics (AUC, positive precision, recall, F1-score) and ranking metrics (P@5, P@10, MAP, MRR, NDCG@5, NDCG@10). We analyzed model performance across varying numbers of eligible patients in training datasets (2 to 70% of the total number of eligible patients). Both approaches demonstrated strong discriminative performance across three trials, with average AUCs of 84.1 % for RF and 76.4 % for TPS, driven primarily by high recall (82.3 % and 77.7 %, respectively). However, positive precision remained low (13.3 % and 9.9 %), reflecting high false-positive rates due to class imbalance. RF showed superior ranking performance, particularly for the trial with the largest eligible cohort (n = 542; P@5 = 78.6 %, MRR = 88.0 %), compared to TPS (P@5 = 47.9 %, MRR = 69.2 %). Both approaches reached performance plateaus with only around 10 eligible patients in training datasets. Variable importance analysis revealed that treatment-related features, diagnostic codes, and procedures were consistently the most important predictors, with relevant patterns identified even with minimal training data. CBR approaches can effectively support patient pre-screening for oncology clinical trials, with RF demonstrating moderately superior performance over TPS. Both methods show robust discriminative performance with small training datasets, though ranking performance varies substantially across trials. Our findings suggest that CBR approaches may benefit from integration with query-based or prompt-based methods during early recruitment phases when training data is scarce.