We propose a novel framework for integrating fragmented multi-modal data in Alzheimer's disease (AD) research using large language models (LLMs) and knowledge graphs. While traditional multimodal analysis requires matched patient IDs across datasets, our approach demonstrates population-level integration of MRI, gene expression, biomarkers, EEG, and clinical indicators from independent cohorts. Statistical analysis identified significant features in each modality, which were connected as nodes in a knowledge graph. LLMs then analyzed the graph to extract potential correlations and generate hypotheses in natural language. This approach revealed several novel relationships, including a potential pathway linking metabolic risk factors to tau protein abnormalities via neuroinflammation (r>0.6, p<0.001), and unexpected correlations between frontal EEG channels and specific gene expression profiles (r=0.42-0.58, p<0.01). Cross-validation with independent datasets confirmed the robustness of major findings, with consistent effect sizes across cohorts (variance <15%). The reproducibility of these findings was further supported by expert review (Cohen's k=0.82) and computational validation. Our framework enables cross modal integration at a conceptual level without requiring patient ID matching, offering new possibilities for understanding AD pathology through fragmented data reuse and generating testable hypotheses for future research.

[¹⁸F]FDG PET/CT scan combined with [¹⁸F]PSMA-1007 PET/CT scan is commonly conducted for detecting bone metastases in prostate cancer (PCa). However, it is expensive and may expose patients to more radiation hazards. This study explores deep learning (DL) techniques to synthesize [¹⁸F]PSMA-1007 PET bone images from CT bone images for the early detection of bone metastases in PCa, which may reduce additional PET/CT scans and relieve the burden on patients. We retrospectively collected paired whole-body (WB) [¹⁸F]PSMA-1007 PET/CT images from 152 patients with clinical and pathological diagnosis results, including 123 PCa and 29 cases of benign lesions. The average age of the patients was 67.48 ± 10.87 years, and the average lesion size was 8.76 ± 15.5 mm. The paired low-dose CT and PET images were preprocessed and segmented to construct the WB bone structure images. 152 subjects were randomly stratified into training, validation, and test groups in the number of 92:41:19. Two generative adversarial network (GAN) models-Pix2pix and Cycle GAN-were trained to synthesize [¹⁸F]PSMA-1007 PET bone images from paired CT bone images. The performance of two synthesis models was evaluated using quantitative metrics of mean absolute error (MAE), mean squared error (MSE), peak signal-to-noise ratio (PSNR), and structural similarity index metrics (SSIM), as well as the target-to-background ratio (TBR). The results of DL-based image synthesis indicated that the synthesis of [¹⁸F]PSMA-1007 PET bone images from low-dose CT bone images was highly feasible. The Pix2pix model performed better with an SSIM of 0.97, PSNR of 44.96, MSE of 0.80, and MAE of 0.10, respectively. The TBRs of bone metastasis lesions calculated on DL-synthesized PET bone images were highly correlated with those of real PET bone images (Pearson's r > 0.90) and had no significant differences (p < 0.05). It is feasible to generate synthetic [¹⁸F]PSMA-1007 PET bone images from CT bone images by using DL techniques with reasonable accuracy, which can provide information for early detection of PCa bone metastases.

With the growing capabilities of language models like ChatGPT to process text and images, this study evaluated their accuracy in detecting supernumerary teeth on periapical radiographs. A customized GPT-4V model (CGPT-4V) was also developed to assess whether domain-specific training could improve diagnostic performance compared to standard GPT-4V and GPT-4o models. One hundred eighty periapical radiographs (90 with and 90 without supernumerary teeth) were evaluated using GPT-4 V, GPT-4o, and a fine-tuned CGPT-4V model. Each image was assessed separately with the standardized prompt "Are there any supernumerary teeth in the radiograph above?" to avoid contextual bias. Three dental experts scored the responses using a three-point Likert scale for positive cases and a binary scale for negatives. Chi-square tests and ROC analysis were used to compare model performances (p < 0.05). Among the three models, CGPT-4 V exhibited the highest accuracy, detecting supernumerary teeth correctly in 91% of cases, compared to 77% for GPT-4o and 63% for GPT-4V. The CGPT-4V model also demonstrated a significantly lower false positive rate (16%) than GPT-4V (42%). A statistically significant difference was found between CGPT-4V and GPT-4o (p < 0.001), while no significant difference was observed between GPT-4V and CGPT-4V or between GPT-4V and GPT-4o. Additionally, CGPT-4V successfully identified multiple supernumerary teeth in radiographs where present. These findings highlight the diagnostic potential of customized GPT models in dental radiology. Future research should focus on multicenter validation, seamless clinical integration, and cost-effectiveness to support real-world implementation.

Medical image captioning is a challenging task that requires generating clinically accurate and semantically meaningful descriptions of radiology images. While recent vision-language models (VLMs) such as BLIP, BLIP2, Gemini and ViT-GPT2 show strong performance on natural image datasets, they often produce generic or imprecise captions when applied to specialized medical domains. In this project, we explore the effectiveness of fine-tuning the BLIP model on the ROCO dataset for improved radiology captioning. We compare the fine-tuned BLIP against its zero-shot version, BLIP-2 base, BLIP-2 Instruct and a ViT-GPT2 transformer baseline. Our results demonstrate that domain-specific fine-tuning on BLIP significantly improves performance across both quantitative and qualitative evaluation metrics. We also visualize decoder cross-attention maps to assess interpretability and conduct an ablation study to evaluate the contributions of encoder-only and decoder-only fine-tuning. Our findings highlight the importance of targeted adaptation for medical applications and suggest that decoder-only fine-tuning (encoder-frozen) offers a strong performance baseline with 5% lower training time than full fine-tuning, while full model fine-tuning still yields the best results overall.

Neuroimaging screening and surveillance is one of the first frontline diagnostic tools leveraged in the acute assessment (first 24 h postinjury) of patients suspected to have traumatic brain injury (TBI). While imaging, in particular computed tomography, is used almost universally in emergency departments worldwide to evaluate possible features of TBI, there is no currently agreed-upon reporting system, standard terminology, or framework to contextualize brain imaging findings with other available medical, psychosocial, and environmental data. In 2023, the NIH-National Institute of Neurological Disorders and Stroke convened six working groups of international experts in TBI to develop a new framework for nomenclature and classification. The goal of this effort was to propose a more granular system of injury classification that incorporates recent progress in imaging biomarkers, blood-based biomarkers, and injury and recovery modifiers to replace the commonly used Glasgow Coma Scale-based diagnosis groups of mild, moderate, and severe TBI, which have shown relatively poor diagnostic, prognostic, and therapeutic utility. Motivated by prior efforts to standardize the nomenclature for pathoanatomic imaging findings of TBI for research and clinical trials, along with more recent studies supporting the refinement of the originally proposed definitions, the Imaging Working Group sought to update and expand this application specifically for consideration of use in clinical practice. Here we report the recommendations of this working group to enable the translation of structured imaging common data elements to the standard of care. These leverage recent advances in imaging technology, electronic medical record (EMR) systems, and artificial intelligence (AI), along with input from key stakeholders, including patients with lived experience, caretakers, providers across medical disciplines, radiology industry partners, and policymakers. It was recommended that (1) there would be updates to the definitions of key imaging features used for this system of classification and that these should be further refined as new evidence of the underlying pathology driving the signal change is identified; (2) there would be an efficient, integrated tool embedded in the EMR imaging reporting system developed in collaboration with industry partners; (3) this would include AI-generated evidence-based feature clusters with diagnostic, prognostic, and therapeutic implications; and (4) a "patient translator" would be developed in parallel to assist patients and families in understanding these imaging features. In addition, important disclaimers would be provided regarding known limitations of current technology until such time as they are overcome, such as resolution and sequence parameter considerations. The end goal is a multifaceted TBI characterization model incorporating clinical, imaging, blood biomarker, and psychosocial and environmental modifiers to better serve patients not only acutely but also through the postinjury continuum in the days, months, and years that follow TBI.

Neuropsychiatric disorders have complex pathological mechanism, pronounced clinical heterogeneity, and a prolonged preclinical phase, which presents a challenge for early diagnosis and development of precise intervention strategies. With the development of large-scale multimodal neuroimaging datasets and advancement of artificial intelligence (AI) algorithms, the integration of multimodal imaging with AI techniques has emerged as a pivotal avenue for early detection and tailoring individualized treatment for neuropsychiatric disorders. To support these advances, in this review, we outline multimodal neuroimaging techniques, AI methods, and strategies for multimodal data fusion. We highlight applications of multimodal AI based on neuroimaging data in precision medicine for neuropsychiatric disorders, discussing challenges in clinical adoption, their emerging solutions, and future directions.

The progress of artificial intelligence (AI) research in dental medicine is hindered by data acquisition challenges and imbalanced distributions. These problems are especially apparent when planning to develop AI-based diagnostic or analytic tools for various lesions, such as maxillary sinus lesions (MSL) including mucosal thickening and polypoid lesions. Traditional unsupervised generative models struggle to simultaneously control the image realism, diversity, and lesion-type specificity. This study establishes an expert-guided framework to overcome these limitations to elevate AI-based diagnostic accuracy. A StyleGAN2 framework was developed for generating clinically relevant MSL images (such as mucosal thickening and polypoid lesion) under expert control. The generated images were then integrated into training datasets to evaluate their effect on ResNet50's diagnostic performance. Here we show: 1) Both lesion subtypes achieve satisfactory fidelity metrics, with structural similarity indices (SSIM > 0.996) and maximum mean discrepancy values (MMD < 0.032), and clinical validation scores close to those of real images; 2) Integrating baseline datasets with synthetic images significantly enhances diagnostic accuracy for both internal and external test sets, particularly improving area under the precision-recall curve (AUPRC) by approximately 8% and 14% for mucosal thickening and polypoid lesions in the internal test set, respectively. The StyleGAN2-based image generation tool effectively addressed data scarcity and imbalance through high-quality MSL image synthesis, consequently boosting diagnostic model performance. This work not only facilitates AI-assisted preoperative assessment for maxillary sinus lift procedures but also establishes a methodological framework for overcoming data limitations in medical image analysis.

Brain-to-image decoding has been recently propelled by the progress in generative AI models and the availability of large ultra-high field functional Magnetic Resonance Imaging (fMRI). However, current approaches depend on complicated multi-stage pipelines and preprocessing steps that typically collapse the temporal dimension of brain recordings, thereby limiting time-resolved brain decoders. Here, we introduce Dynadiff (Dynamic Neural Activity Diffusion for Image Reconstruction), a new single-stage diffusion model designed for reconstructing images from dynamically evolving fMRI recordings. Our approach offers three main contributions. First, Dynadiff simplifies training as compared to existing approaches. Second, our model outperforms state-of-the-art models on time-resolved fMRI signals, especially on high-level semantic image reconstruction metrics, while remaining competitive on preprocessed fMRI data that collapse time. Third, this approach allows a precise characterization of the evolution of image representations in brain activity. Overall, this work lays the foundation for time-resolved brain-to-image decoding.

Large language models (LLMs) have demonstrated remarkable capabilities in various domains, including radiology report generation. Previous approaches have attempted to utilize multimodal LLMs for this task, enhancing their performance through the integration of domain-specific knowledge retrieval. However, these approaches often overlook the knowledge already embedded within the LLMs, leading to redundant information integration and inefficient utilization of learned representations. To address this limitation, we propose RADAR, a framework for enhancing radiology report generation with supplementary knowledge injection. RADAR improves report generation by systematically leveraging both the internal knowledge of an LLM and externally retrieved information. Specifically, it first extracts the model's acquired knowledge that aligns with expert image-based classification outputs. It then retrieves relevant supplementary knowledge to further enrich this information. Finally, by aggregating both sources, RADAR generates more accurate and informative radiology reports. Extensive experiments on MIMIC-CXR, CheXpert-Plus, and IU X-ray demonstrate that our model outperforms state-of-the-art LLMs in both language quality and clinical accuracy