The increasing volume of radiological images and the associated workload of report generation necessitate efficient solutions, making artificial intelligence (AI) a crucial tool to streamline this process for radiologists. Recent years have seen a surge in research exploring AI-driven radiological report generation directly from images, particularly with the emergence of large vision language models. However, a comprehensive understanding of the current landscape, including specific limitations and the extent to which efforts move beyond abnormality detection to full textual report generation, remains unclear. This scoping review aims to systematically map the existing literature to provide an overview of the current state of AI in generating radiological reports from medical images, including the scope and limitations of existing research. To our knowledge, no prior scoping review has comprehensively mapped this landscape, especially considering recent advancements in foundation models in medicine and related AI architectures. Considering the explosive growth of related studies in recent years, a comprehensive scoping review will be significant in mapping the current research status and understanding relevant limitations. This scoping review will follow the Preferred Reporting Items for Systematic Reviews and Meta-Analysis extension for Scoping Reviews guidelines to map the literature on AI generating radiological reports from medical images. We will search PubMed, Scopus and Web of Science for peer-reviewed articles (January 2016 to March 2025) using keywords related to AI, radiological reports and medical images. Original research in English focusing on AI-driven report generation from images will be included and studies without report generation or not using medical images as input will be excluded. Two independent reviewers will perform a two-stage screening. Data extraction, guided by the Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modelling Studies checklist and focusing on study characteristics, AI methods, image modalities, report features, limitations and key findings, will be analysed using narrative and descriptive synthesis, with results presented in tables, figures and a narrative summary. This protocol describes a scoping literature review methodology that does not involve research on humans, animals or their data; therefore, no ethical approval is required. Following the review, the results will be considered for publication in a relevant peer-reviewed journal and may be shared with stakeholders through reports or summaries.

This study aims to develop and validate an interpretable deep learning (DL) model and a nomogram based on endoscopic ultrasound (EUS) images for the prediction of pathological grading in pancreatic neuroendocrine tumors (PNETs). This multicenter retrospective study included 108 patients with PNETs, who were divided into train (<i>n</i> = 81, internal center) and test cohorts (<i>n</i> = 27, external centers). Univariate and multivariate logistic regression were used for screening demographic characteristics and EUS semantic features. Deep transfer learning was employed using a pre-trained ResNet18 model to extract features from EUS images. Feature selection was conducted using the least absolute shrinkage and selection operator (LASSO), and various machine learning algorithms were utilized to construct DL models. The optimal model was then integrated with clinical features to develop a nomogram. The performance of the model was assessed using the area under the curve (AUC), calibration curves, decision curve analysis (DCA), and clinical impact curves (CIC). The nomogram, which integrates the optimal DL model (Naive Bayes) with clinical features, achieved AUC values of 0.928 (95% CI 0.849–0.981) in the train cohort and 0.882 (95% CI 0.778–0.954) in the test cohort. Calibration curves revealed minimal discrepancies between predicted and actual probabilities, with mean absolute errors of 4.5% and 6.6% in the train and test cohorts, respectively. DCA and CIC demonstrated substantial net benefit and clinical utility. The SHapley Additive exPlanations (SHAP) method provided insights into the contribution of each DL feature to the model’s predictions. This study developed and validated a novel interpretable DL model and nomogram using EUS images and machine learning, which holds promise for enhancing the clinical application of EUS in identifying PNETs’ pathological grading. The online version contains supplementary material available at 10.1186/s12911-025-03193-3.

Non-muscle invasive bladder cancer (NMIBC) has a high rate of postoperative recurrence and the efficacy of existing clinical prediction models is limited. This study aimed to combine multiparametric magnetic resonance imaging (mp-MRI) radiomic features with clinical characteristics to construct a machine learning model for accurately predicting the risk of recurrence within 2 years postoperatively in NMIBC patients. Retrospectively including 183 NMIBC patients (57 in the recurrence group, 126 in the non-recurrence group), radiomic features from mp-MRI imaging (T2W, ADC, and enhancement sequences) were extracted. Through LASSO selection, 4 key imaging features (MajorAxisLength, SZNN, S/V, Skewness) and 6 clinical features based on the EAU 2021 risk stratification were identified to constitute the clinical-imaging dataset. Through comparison with 10 machine learning models, Support Vector Machine (SVM) performed the best (training set AUC = 0.973, validation set AUC = 0.891), with external independent validation (108 cases) showing AUCs of 0.88 and 0.87, demonstrating good generalization ability. A bar chart integrating radiomics score (Rad-Score) with clinical features provides an intuitive prognostic tool. The study indicates that the clinical-imaging radiomics model based on SVM significantly enhances the efficacy of NMIBC recurrence prediction, addressing the shortcomings of traditional risk assessment and offering a reliable basis for personalized postoperative management. Study limitations include the retrospective design and the absence of molecular biomarkers, necessitating future multicenter prospective validation.

Dyslexia is a neurobiological disorder characterized by reading difficulties, yet its causes remain unclear. Neuroimaging and behavioral studies found anomalous responses in tasks requiring phonological processing, motion perception, and implicit learning, and showed gray and white matter abnormalities in dyslexics compared to controls, indicating that dyslexia is highly heterogeneous and promoting a multifactorial approach. To evaluate whether combining behavioral and multimodal MRI improves sensitivity in identifying dyslexia neurocognitive traits compared to monocomponential approaches, 19 dyslexic and 19 control subjects underwent cognitive assessments, multiple (phonological, visual motion, rhythmic) mismatch-response functional MRI tasks, structural diffusion-weighted imaging (DWI) and T1-weighted imaging. Between group differences in the neurocognitive measures were tested with univariate and multivariate approaches. Results showed that dyslexics performed worse than controls in phonological tasks and presented reduced cerebellar responses to mismatching rhythmic stimuli, as well as structural disorganization in white matter tracts and cortical regions. Most importantly, a machine learning model trained with features from all three MRI modalities discriminated between dyslexics and controls with greater accuracy than single-modality models. The individual classification scores in the multimodal machine learning model correlated with behavioral reading accuracy. These results characterize dyslexia as a composite condition with multiple distinctive cognitive and brain traits.

Cardiac magnetic resonance (CMR) is a complex imaging modality requiring a broad variety of image processing tasks for comprehensive assessment of the study. Recently, foundation models (FM) have shown promise for automated image analyses in natural images (NI). In this study, a CMR-specific vision FM was developed and then finetuned in a supervised manner for 9 different imaging tasks typical to a CMR workflow, including classification, segmentation, landmark localization, and pathology detection. A ViT-S/8 model was trained in a self-supervised manner using DINO on 36 million CMR images from 27,524 subjects from three sources (UK Biobank and two clinical centers). The model was then finetuned for 9 tasks: classification (sequence, cine view), segmentation (cine SAX, cine LAX, LGE SAX, Mapping SAX), landmark localization, pathology detection (LGE, cardiac disease), on data from various sources (both public and 3 clinical datasets). The results were compared against metrics from state-of-the-art methods on the same tasks. A comparable baseline model was also trained on the same datasets for direct comparison. Additionally, the effect of pretraining strategy, as well as generalization and few-shot performance (training on few labeled samples) were explored for the pretrained model, compared to the baseline. The proposed model obtained similar performance or moderate improvements to results reported in the literature in most tasks (except disease detection), without any task-specific optimization of methodology. The proposed model outperformed the baseline in most cases, with an average increase of 6.8 percentage points (pp) for cine view classification, and 0.1 to 1.8 pp for segmentation tasks. The proposed method also obtained generally lower standard deviations in the metrics. Improvements of 3.7 and 6.6 pp for hyperenhancement detection from LGE and 14 pp for disease detection were observed. Ablation studies highlighted the importance of pretraining strategy, architecture and the impact of domain shifts from pretraining to finetuning. Moreover, CMR-pretrained model achieved better generalization and few-shot performance compared to the baseline. Vision FM specialized for medical imaging can improve accuracy and robustness over NI-FM. Self-supervised pretraining offers a resource-efficient, unified framework for CMR assessment, with the potential to accelerate the development of deep learning-based solutions for image analysis tasks, even with few annotated data available.

Artificial intelligence has significantly enhanced disease diagnosis in healthcare, particularly through Deep Learning (DL) and Federated Learning (FL) approaches. These technologies have shown promise in detecting ocular diseases using medical imaging while addressing challenges related to data privacy and security. FL enables collaborative learning without sharing sensitive medical data, making it an attractive solution for healthcare applications. This systematic review aims to analyze the advancements in AI-driven ocular disease detection, with a particular focus on FL-based approaches. The article evaluates the evolution, methodologies, challenges, and effectiveness of FL in enhancing diagnostic accuracy while ensuring data confidentiality. The systematic review followed the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) framework to ensure transparency and reliability. Research articles published between 2017 and 2024 were identified using academic databases, including Web of Science, Scopus, IEEE Xplore, and PubMed. Studies focusing on DL and FL models for detecting ocular diseases were selected based on predefined inclusion and exclusion criteria. A comparative analysis of the methodologies, architectures, datasets, and performance metrics of different FL models has been presented. The findings indicated that FL preserves data privacy while achieving diagnostic performance comparable to traditional centralized AI models. Various FL models, including FedAvg and FedProx, have been implemented for ocular disease detection, with high accuracy and efficiency. However, challenges, such as data heterogeneity, communication efficiency, and model convergence, persist. FL represents a promising approach for ocular disease detection, balancing diagnostic accuracy with data privacy. Future research may focus on optimizing FL frameworks for improving scalability, communication efficiency, and integrating advanced privacy-preserving techniques.

Deep learning methods were employed to perform harmonization analysis on whole-brain scans obtained from 1.5-T and 3.0-T scanners, aiming to increase comparability between different magnetic resonance imaging (MRI) scanners. Thirty patients evaluated in Beijing Tsinghua Changgung Hospital between August 2020 and March 2023 were included in this retrospective study. Three MRI scanners were used to scan patients, and automated brain image segmentation was performed to obtain volumes of different brain regions. Differences in regional volumes across scanners were analyzed using repeated-measures analysis of variance. For regions showing significant differences, super-resolution deep learning was applied to enhance consistency, with subsequent comparison of results. For regions still exhibiting differences, the Intraclass Correlation Coefficient (ICC) was calculated and the consistency was evaluated using Cicchetti's criteria. Average whole-brain volumes for different scanners among patients were 1152.36mm³ (SD = 95.34), 1136.92mm³ (SD = 108.21), and 1184.00mm³ (SD = 102.78), respectively. Analysis revealed significant variations in all 12 brain regions (p<0.05), indicating a lack of comparability among imaging results obtained from different magnetic field strengths. After deep learning-based consistency optimization, most brain regions showed no significant differences, except for six regions where differences remained significant. Among these, three regions demonstrated ICC values of 0.868 (95%CI 0.771-0.931), 0.776 (95%CI 0.634-0.877), and 0.893 (95%CI 0.790-0.947), indicating high reproducibility and comparability. This study employed a novel machine learning approach that significantly improved the comparability of imaging results from patients using different magnetic field strengths and various models of MRI scanners. Furthermore, it enhanced the consistency of central nervous system image segmentation.

Nearly one in five adolescents currently live with a diagnosed mental or behavioral health condition, such as anxiety, depression, or conduct disorder, underscoring the urgency of developing accurate and interpretable diagnostic tools. Resting-state functional magnetic resonance imaging (rs-fMRI) provides a powerful lens into large-scale functional connectivity, where brain regions are modeled as nodes and inter-regional synchrony as edges, offering clinically relevant biomarkers for psychiatric disorders. While prior works use graph neural network (GNN) approaches for disorder prediction, they remain complex black-boxes, limiting their reliability and clinical translation. In this work, we propose CONCEPTNEURO, a concept-based diagnosis framework that leverages large language models (LLMs) and neurobiological domain knowledge to automatically generate, filter, and encode interpretable functional connectivity concepts. Each concept is represented as a structured subgraph linking specific brain regions, which are then passed through a concept classifier. Our design ensures predictions through clinically meaningful connectivity patterns, enabling both interpretability and strong predictive performance. Extensive experiments across multiple psychiatric disorder datasets demonstrate that CONCEPTNEURO-augmented GNNs consistently outperform their vanilla counterparts, improving accuracy while providing transparent, clinically aligned explanations. Furthermore, concept analyses highlight disorder-specific connectivity patterns that align with expert knowledge and suggest new hypotheses for future investigation, establishing CONCEPTNEURO as an interpretable, domain-informed framework for psychiatric disorder diagnosis.

In this study, we utilized the multimodal capabilities of OpenAI GPT-4o to automatically generate jaw cyst findings on dental panoramic radiographs. To improve accuracy, we constructed a Self-correction Loop with Structured Output (SLSO) framework and verified its effectiveness. A 10-step process was implemented for 22 cases of jaw cysts, including image input and analysis, structured data generation, tooth number extraction and consistency checking, iterative regeneration when inconsistencies were detected, and finding generation with subsequent restructuring and consistency verification. A comparative experiment was conducted using the conventional Chain-of-Thought (CoT) method across seven evaluation items: transparency, internal structure, borders, root resorption, tooth movement, relationships with other structures, and tooth number. The results showed that the proposed SLSO framework improved output accuracy for many items, with 66.9%, 33.3%, and 28.6% improvement rates for tooth number, tooth movement, and root resorption, respectively. In the successful cases, a consistently structured output was achieved after up to five regenerations. Although statistical significance was not reached because of the small size of the dataset, the overall SLSO framework enforced negative finding descriptions, suppressed hallucinations, and improved tooth number identification accuracy. However, the accurate identification of extensive lesions spanning multiple teeth is limited. Nevertheless, further refinement is required to enhance overall performance and move toward a practical finding generation system.

BackgroundDeep learning has achieved rapid development in recent years and has been applied to various fields in dentistry. While cross-disciplinary research between artificial intelligence and dentistry is growing exponentially, most studies rely on off-the-shelf machine learning models, with only a small portion introducing technological novelty. Furthermore, tasks such as dental disease diagnosis are inherently complex, with high intra- and interobserver variability where dentists often interpret radiographs differently and offer varying subsequent treatments. However, many studies overlooked this variability, assuming no data and model uncertainty in dental tasks. Additionally, many evaluated their methods using private and small-scale datasets, making fair comparisons of their outcome metrics challenging and introducing significant predictive bias in artificial intelligence models. The goal of the current study was to examine and critically assess recent novel advances in artificial intelligence in dentistry across a wide range of dental applications. MethodsWe begin by presenting foundational concepts in artificial intelligence and adopt a unique approach by focusing on the novelty of deep learning methods. Following that, we conducted a systematic review by searching online databases (PubMed, IEEE Xplore, arXiv, and Google Scholar) for publications related to artificial intelligence, machine learning, and deep learning applications in dentistry. ResultsA total of 91 articles met the inclusion criteria, and we presented a comprehensive analysis of the studies. Moreover, we discuss the limitations of recent studies on artificial intelligence in dentistry and identify key research opportunities for progress and innovation. These include integrating dental domain knowledge, quantifying uncertainty, leveraging large models and multiple sources of datasets, developing efficient deep learning pipelines, and conducting thorough evaluations in both simulated and real-world experimental settings. ConclusionRecent advancements in deep learning demonstrate great potential in dentistry applications. However, future research to address the limitations in recent studies is needed to fully realize its potential for enhancing dental professionals to utilize AI effectively and improve clinical and patient outcome in dentistry.