Background: Liver cancer remains the leading cause of cancer death globally, and the treatment strategies are distinct for each type of malignant hepatic tumors. However, the differential diagnosis before surgery is challenging and subjective. This study aims to build an automatic diagnostic model for differentiating malignant hepatic tumors based on patients’ multimodal medical data including multi-phase contrast-enhanced computed tomography and clinical features. Methods: Our study consisted of 723 patients from two centers, who were pathologically diagnosed with HCC, ICC or metastatic liver cancer. The training set and the test set consisted of 499 and 113 patients from center 1, respectively. The external test set consisted of 111 patients from center 2. We proposed a deep learning model with the modular design of SpatialExtractor-TemporalEncoder-Integration-Classifier (STIC), which take the advantage of deep CNN and gated RNN to effectively extract and integrate the diagnosis-related radiological and clinical features of patients. The code is publicly available at https://github.com/ruitian-olivia/STIC-model. Results: The STIC model achieved an accuracy of 86.2% and AUC of 0.893 for classifying HCC and ICC on the test set. When extended to differential diagnosis of malignant hepatic tumors, the STIC model achieved an accuracy of 72.6% on the test set, comparable with the diagnostic level of doctors’ consensus (70.8%). With the assistance of the STIC model, doctors achieved better performance than doctors’ consensus diagnosis, with an increase of 8.3% in accuracy and 26.9% in sensitivity for ICC diagnosis on average. On the external test set from center 2, the STIC model achieved an accuracy of 82.9%, which verify the model’s generalization ability. Conclusions: We incorporated deep CNN and gated RNN in the STIC model design for differentiating malignant hepatic tumors based on multi-phase CECT and clinical features. Our model can assist doctors to achieve better diagnostic performance, which is expected to serve as an AI assistance system and promote the precise treatment of liver cancer.

Computational pathology, utilizing whole slide images (WSIs) for pathological diagnosis, has advanced the development of intelligent healthcare. However, the scarcity of annotated data and histological differences hinder the general application of existing methods. Extensive histopathological data and the robustness of self-supervised models in small-scale data demonstrate promising prospects for developing foundation pathology models. Here we show BEPH (BEiT-based model Pre-training on Histopathological image), a foundation model that leverages self-supervised learning to learn meaningful representations from 11 million unlabeled histopathological images. These representations are then efficiently adapted to various tasks, including patch-level cancer diagnosis, WSI-level cancer classification, and survival prediction for multiple cancer subtypes. By leveraging the masked image modeling (MIM) pre-training approach, BEPH offers an efficient solution to enhance model performance, reduce the reliance on expert annotations, and facilitate the broader application of artificial intelligence in clinical settings. The pre-trained model is available at https://github.com/Zhcyoung/BEPH.

Background: Identifying individuals with mild cognitive impairment (MCI) at risk of progressing to Alzheimer’s disease (AD) provides a unique opportunity for early interventions. Therefore, accurate and long-term prediction of the conversion from MCI to AD is desired but, to date, remains challenging. Here, we developed an interpretable deep learning model featuring a novel design that incorporates interaction effects and multimodality to improve the prediction accuracy and horizon for MCI-to-AD progression. Methods: This multi-center, multi-cohort retrospective study collected structural magnetic resonance imaging (sMRI), clinical assessments, and genetic polymorphism data of 252 patients with MCI at baseline from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database. Our deep learning model was cross-validated on the ADNI-1 and ADNI-2/GO cohorts and further generalized in the ongoing ADNI-3 cohort. We evaluated the model performance using the area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, and F1 score. Results: On the cross-validation set, our model achieved superior results for predicting MCI conversion within 4 years (AUC, 0.962; accuracy, 92.92%; sensitivity, 88.89%; specificity, 95.33%) compared to all existing studies. In the independent test, our model exhibited consistent performance with an AUC of 0.939 and an accuracy of 92.86%. Integrating interaction effects and multimodal data into the model significantly increased prediction accuracy by 4.76% (P = 0.01) and 4.29% (P = 0.03), respectively. Furthermore, our model demonstrated robustness to inter-center and inter-scanner variability, while generating interpretable predictions by quantifying the contribution of multimodal biomarkers. Conclusions: The proposed deep learning model presents a novel perspective by combining interaction effects and multimodality, leading to more accurate and longer-term predictions of AD progression, which promises to improve pre-dementia patient care.

Mediation analysis with high-dimensional DNA methylation markers is important in identifying epigenetic pathways between environmental exposures and health outcomes. There have been some methodology developments of mediation analysis with high-dimensional mediators. However, high-dimensional mediation analysis methods for time-to-event outcome data are still yet to be developed. To address these challenges, we propose a new high-dimensional mediation analysis procedure for survival models by incorporating sure independent screening and minimax concave penalty techniques for variable selection, with the Sobel and the joint method for significance test of indirect effect. The simulation studies show good performance in identifying correct biomarkers, false discovery rate control, and minimum estimation bias of the proposed procedure. We also apply this approach to study the causal pathway from smoking to overall survival among lung cancer patients potentially mediated by 365,307 DNA methylations in the TCGA lung cancer cohort. Mediation analysis using a Cox proportional hazards model estimates that patients who have serious smoking history increase the risk of lung cancer through methylation markers including cg21926276, cg27042065, and cg26387355 with significant hazard ratios of 1.2497(95%CI: 1.1121, 1.4045), 1.0920(95%CI: 1.0170, 1.1726), and 1.1489(95%CI: 1.0518, 1.2550), respectively. The three methylation sites locate in the three genes which have been showed to be associated with lung cancer event or overall survival. However, the three CpG sites (cg21926276, cg27042065 and cg26387355) have not been reported, which are newly identified as the potential novel epigenetic markers linking smoking and survival of lung cancer patients. Collectively, the proposed high-dimensional mediation analysis procedure has good performance in mediator selection and indirect effect estimation. In this research, we established an efficient procedure for high-dimensional mediation analysis with time-to-event data to select DNA methylation markers and estimate the mediation effects. To evaluate the performance of the proposed procedure, we conducted extensive simulation studies and analyzed a lung cancer data set using our method. We demonstrated the validity and utility of our method under a variety of scenarios.

Identifying spatially variable genes (SVGs) is crucial for understanding the spatiotemporal characteristics of diseases and tissue structures, posing a distinctive challenge in spatial transcriptomics research. We propose HEARTSVG, a distribution-free, test-based method for fast and accurately identifying spatially variable genes in large-scale spatial transcriptomic data. Extensive simulations demonstrate that HEARTSVG outperforms state-of-the-art methods with higher $${F}_{1}$$F 1 scores (average $${F}_{1}$$F 1 Score=0.948), improved computational efficiency, scalability, and reduced false positives (FPs). Through analysis of twelve real datasets from various spatial transcriptomic technologies, HEARTSVG identifies a greater number of biologically significant SVGs (average AUC = 0.792) than other comparative methods without prespecifying spatial patterns. Furthermore, by clustering SVGs, we uncover two distinct tumor spatial domains characterized by unique spatial expression patterns, spatial-temporal locations, and biological functions in human colorectal cancer data, unraveling the complexity of tumors.

Spatial transcriptomics (ST) provides insights into the tumor microenvironment (TME), which is closely associated with cancer prognosis, but ST has limited clinical availability. In this study, we provide a powerful deep learning system to augment TME information based on histological images for patients without ST data, thereby empowering precise cancer prognosis. The system provides two connections to bridge existing gaps. The first is the integrated graph and image deep learning (IGI-DL) model, which predicts ST expression based on histological images with a 0.171 increase in mean correlation across three cancer types compared with five existing methods. The second connection is the cancer prognosis prediction model, based on TME depicted by spatial gene expression. Our survival model, using graphs with predicted ST features, achieves superior accuracy with a concordance index of 0.747 and 0.725 for The Cancer Genome Atlas breast cancer and colorectal cancer cohorts, outperforming other survival models. For the external Molecular and Cellular Oncology colorectal cancer cohort, our survival model maintains a stable advantage.