Artificial intelligence has become an impressive force manifesting itself in the radiology field, improving workflows, and influencing clinical decision-making. With this increasing presence, a closer look at how residents can be properly exposed to this technology is needed. Within this paper, we aim to discuss the three pillars central to a trainee's experience including education on AI, AI-Education tools, and clinical implementation of AI. An already overcrowded clinical residency curricula makes little room for a thorough AI education; the challenge of which may be overcome through longitudinal distinct educational tracks during residency or external courses offered through a variety of societies. In addition to teaching the fundamentals of AI, programs which offer education tools utilizing AI will improve on antiquated clinical curricula. These education tools are a growing field in research and industry offering a variety of unique opportunities to promote active inquiry, improved comprehension and overall clinical competence. The near 700 FDA-approved AI clinical tools almost guarantees that residents will be exposed to this technology which may have mixed effects on education, although more research needs to be done to further elucidate this challenge. Ethical considerations, including algorithmic bias, liability, and post-deployment monitoring, highlight the need for structured instruction and mentorship. As AI continues to evolve, residency programs must prioritize evidence-based, adaptable curricula to prepare future radiologists to critically assess, utilize, and contribute to AI advancements, ensuring that these tools complement rather than undermine clinical expertise.

Radiology education is challenged by increasing clinical workloads, limiting trainee supervision time and hindering real-time feedback. Large language models (LLMs) can enhance radiology education by providing real-time guidance, feedback, and educational resources while supporting efficient clinical workflows. We present an interpretation-centric framework for integrating LLMs into radiology education subdivided into distinct phases spanning pre-dictation preparation, active dictation support, and post-dictation analysis. In the pre-dictation phase, LLMs can analyze clinical data and provide context-aware summaries of each case, suggest relevant educational resources, and triage cases based on their educational value. In the active dictation phase, LLMs can provide real-time educational support through processes such as differential diagnosis support, completeness guidance, classification schema assistance, structured follow-up guidance, and embedded educational resources. In the post-dictation phase, LLMs can be used to analyze discrepancies between trainee and attending reports, identify areas for improvement, provide targeted educational recommendations, track trainee performance over time, and analyze the radiologic entities that trainees encounter. This framework offers a comprehensive approach to integrating LLMs into radiology education, with the potential to enhance trainee learning while preserving clinical efficiency.

There has been increasing interest in the generation of high-quality, realistic synthetic medical images in recent years. Such synthetic datasets can mitigate the scarcity of public datasets for artificial intelligence research, and can also be used for educational purposes. In this paper, we propose a combination of diffusion-based generation (PanoDiff) and Super-Resolution (SR) for generating synthetic dental panoramic radiographs (PRs). The former generates a low-resolution (LR) seed of a PR (256 X 128) which is then processed by the SR model to yield a high-resolution (HR) PR of size 1024 X 512. For SR, we propose a state-of-the-art transformer that learns local-global relationships, resulting in sharper edges and textures. Experimental results demonstrate a Frechet inception distance score of 40.69 between 7243 real and synthetic images (in HR). Inception scores were 2.55, 2.30, 2.90 and 2.98 for real HR, synthetic HR, real LR and synthetic LR images, respectively. Among a diverse group of six clinical experts, all evaluating a mixture of 100 synthetic and 100 real PRs in a time-limited observation, the average accuracy in distinguishing real from synthetic images was 68.5% (with 50% corresponding to random guessing).

Foundation models like CLIP and SAM have transformed computer vision and medical imaging via low-shot transfer learning. However, deployment of these models hindered by two key challenges: \textit{distribution shift} between training and test data, and \textit{confidence misalignment} that leads to overconfident incorrect predictions. These issues manifest differently in vision-language classification and medical segmentation tasks, yet existing solutions remain domain-specific. We propose \textit{StaRFM}, a unified framework addressing both challenges. It introduces a Fisher information penalty (FIP), extended to 3D medical data via patch-wise regularization, to reduce covariate shift in CLIP and SAM embeddings. Additionally, a confidence misalignment penalty (CMP), reformulated for voxel-level predictions, calibrates uncertainty in segmentation tasks. We theoretically derive PAC-Bayes bounds showing FIP controls generalization via the Fisher-Rao norm, while CMP minimizes calibration error through Brier score optimization. StaRFM shows consistent performance like \texttt{+}3.5\% accuracy and 28\% lower ECE on 19 vision datasets (e.g., ImageNet, Office-Home), 84.7\% DSC and 4.8mm HD95 in medical segmentation (e.g., BraTS, ATLAS), and 40\% lower cross-domain performance gap compared to prior benchmarking methods. The framework is plug-and-play, requiring minimal architectural changes for seamless integration with foundation models. Code and models will be released at https://anonymous.4open.science/r/StaRFM-C0CD/README.md

Large language models (LLMs) can generate outputs understandable by humans, such as answers to medical questions and radiology reports. With the rapid development of LLMs, clinicians face a growing challenge in determining the most suitable algorithms to support their work. We aimed to provide clinicians and other health care practitioners with systematic guidance in selecting an LLM that is relevant and appropriate to their needs and facilitate the integration process of LLMs in health care. We conducted a literature search of full-text publications in English on clinical applications of LLMs published between January 1, 2022, and March 31, 2025, on PubMed, ScienceDirect, Scopus, and IEEE Xplore. We excluded papers from journals below a set citation threshold, as well as papers that did not focus on LLMs, were not research based, or did not involve clinical applications. We also conducted a literature search on arXiv within the same investigated period and included papers on the clinical applications of innovative multimodal LLMs. This led to a total of 270 studies. We collected 330 LLMs and recorded their application frequency in clinical tasks and frequency of best performance in their context. On the basis of a 5-stage clinical workflow, we found that stages 2, 3, and 4 are key stages in the clinical workflow, involving numerous clinical subtasks and LLMs. However, the diversity of LLMs that may perform optimally in each context remains limited. GPT-3.5 and GPT-4 were the most versatile models in the 5-stage clinical workflow, applied to 52% (29/56) and 71% (40/56) of the clinical subtasks, respectively, and they performed best in 29% (16/56) and 54% (30/56) of the clinical subtasks, respectively. General-purpose LLMs may not perform well in specialized areas as they often require lightweight prompt engineering methods or fine-tuning techniques based on specific datasets to improve model performance. Most LLMs with multimodal abilities are closed-source models and, therefore, lack of transparency, model customization, and fine-tuning for specific clinical tasks and may also pose challenges regarding data protection and privacy, which are common requirements in clinical settings. In this review, we found that LLMs may help clinicians in a variety of clinical tasks. However, we did not find evidence of generalist clinical LLMs successfully applicable to a wide range of clinical tasks. Therefore, their clinical deployment remains challenging. On the basis of this review, we propose an interactive online guideline for clinicians to select suitable LLMs by clinical task. With a clinical perspective and free of unnecessary technical jargon, this guideline may be used as a reference to successfully apply LLMs in clinical settings.

Cone beam computed tomography (CBCT) has revolutionized dental imaging due to its high spatial resolution and ability to provide detailed three-dimensional reconstructions of dental structures. This study introduces an innovative CBCT image processing method using an oriented object detection approach integrated with a Region of Interest (RoI) Transformer. This study addresses the challenge of accurate tooth detection and classification in PAN derived from CBCT, introducing an innovative oriented object detection approach, which has not been previously applied in dental imaging. This method better aligns with the natural growth patterns of teeth, allowing for more accurate detection and classification of molars, premolars, canines, and incisors. By integrating RoI transformer, the model demonstrates relatively acceptable performance metrics compared to conventional horizontal detection methods, while also offering enhanced visualization capabilities. Furthermore, post-processing techniques, including distance and grayscale value constraints, are employed to correct classification errors and reduce false positives, especially in areas with missing teeth. The experimental results indicate that the proposed method achieves an accuracy of 98.48%, a recall of 97.21%, an F1 score of 97.21%, and an mAP of 98.12% in tooth detection. The proposed method enhances the accuracy of tooth detection in CBCT-derived PAN by reducing background interference and improving the visualization of tooth orientation.

In medical image segmentation, convolutional neural networks (CNNs) and transformers are dominant. For CNNs, given the local receptive fields of convolutional layers, long-range spatial correlations are captured through consecutive convolutions and pooling. However, as the computational cost and memory footprint can be prohibitively large, 3D models can only afford fewer layers than 2D models with reduced receptive fields and abstract levels. For transformers, although long-range correlations can be captured by multi-head attention, its quadratic complexity with respect to input size is computationally demanding. Therefore, either model may require input size reduction to allow more filters and layers for better segmentation. Nevertheless, given their discrete nature, models trained with patch-wise training or image downsampling may produce suboptimal results when applied on higher resolutions. To address this issue, here we propose the resolution-robust HNOSeg-XS architecture. We model image segmentation by learnable partial differential equations through the Fourier neural operator which has the zero-shot super-resolution property. By replacing the Fourier transform by the Hartley transform and reformulating the problem in the frequency domain, we created the HNOSeg-XS model, which is resolution robust, fast, memory efficient, and extremely parameter efficient. When tested on the BraTS'23, KiTS'23, and MVSeg'23 datasets with a Tesla V100 GPU, HNOSeg-XS showed its superior resolution robustness with fewer than 34.7k model parameters. It also achieved the overall best inference time (< 0.24 s) and memory efficiency (< 1.8 GiB) compared to the tested CNN and transformer models¹.

This study investigates the automation of MRI protocoling, a routine task in radiology, using large language models (LLMs), comparing an open-source (LLama 3.1 405B) and a proprietary model (GPT-4o) with and without retrieval-augmented generation (RAG), a method for incorporating domain-specific knowledge. This retrospective study included MRI studies conducted between January and December 2023, along with institution-specific protocol assignment guidelines. Clinical questions were extracted, and a neuroradiologist established the gold standard protocol. LLMs were tasked with assigning MRI protocols and contrast medium administration with and without RAG. The results were compared to protocols selected by four radiologists. Token-based symmetric accuracy, the Wilcoxon signed-rank test, and the McNemar test were used for evaluation. Data from 100 neuroradiology reports (mean age = 54.2 years ± 18.41, women 50%) were included. RAG integration significantly improved accuracy in sequence and contrast media prediction for LLama 3.1 (Sequences: 38% vs. 70%, P < .001, Contrast Media: 77% vs. 94%, P < .001), and GPT-4o (Sequences: 43% vs. 81%, P < .001, Contrast Media: 79% vs. 92%, P = .006). GPT-4o outperformed LLama 3.1 in MRI sequence prediction (81% vs. 70%, P < .001), with comparable accuracies to the radiologists (81% ± 0.21, P = .43). Both models equaled radiologists in predicting contrast media administration (LLama 3.1 RAG: 94% vs. 91% ± 0.2, P = .37, GPT-4o RAG: 92% vs. 91% ± 0.24, P = .48). Large language models show great potential as decision-support tools for MRI protocoling, with performance similar to radiologists. RAG enhances the ability of LLMs to provide accurate, institution-specific protocol recommendations.

Semi-supervised medical image segmentation techniques have demonstrated significant potential and effectiveness in clinical diagnosis. The prevailing approaches using the mean-teacher (MT) framework achieve promising image segmentation results. However, due to the unreliability of the pseudo labels generated by the teacher model, existing methods still have some inherent limitations that must be considered and addressed. In this paper, we propose an innovative semi-supervised method for medical image segmentation by combining the heterogeneous complementary correction network and confidence contrastive learning (HC-CCL). Specifically, we develop a triple-branch framework by integrating a heterogeneous complementary correction (HCC) network into the MT framework. HCC serves as an auxiliary branch that corrects prediction errors in the student model and provides complementary information. To improve the capacity for feature learning in our proposed model, we introduce a confidence contrastive learning (CCL) approach with a novel sampling strategy. Furthermore, we develop a momentum style transfer (MST) method to narrow the gap between labeled and unlabeled data distributions. In addition, we introduce a Cutout-style augmentation for unsupervised learning to enhance performance. Three medical image datasets (including left atrial (LA) dataset, NIH pancreas dataset, Brats-2019 dataset) were employed to rigorously evaluate HC-CCL. Quantitative results demonstrate significant performance advantages over existing approaches, achieving state-of-the-art performance across all metrics. The implementation will be released at https://github.com/xxmmss/HC-CCL .

To develop an artificial intelligence (AI)-driven model for automatic Lenke classification of adolescent idiopathic scoliosis (AIS) and assess its performance. This retrospective study utilized 860 spinal radiographs from 215 AIS patients with four views, including 161 training sets and 54 testing sets. Additionally, 1220 spinal radiographs from 610 patients with only anterior-posterior (AP) and lateral (LAT) views were collected for training. The model was designed to perform keypoint detection, pedicle segmentation, and AIS classification based on a custom classification strategy. Its performance was evaluated against the gold standard using metrics such as mean absolute difference (MAD), intraclass correlation coefficient (ICC), Bland-Altman plots, Cohen's Kappa, and the confusion matrix. In comparison to the gold standard, the MAD for all predicted angles was 2.29°, with an excellent ICC. Bland-Altman analysis revealed minimal differences between the methods. For Lenke classification, the model exhibited exceptional consistency in curve type, lumbar modifier, and thoracic sagittal profile, with average Kappa values of 0.866, 0.845, and 0.827, respectively, and corresponding accuracy rates of 87.07%, 92.59%, and 92.59%. Subgroup analysis further confirmed the model's high consistency, with Kappa values ranging from 0.635 to 0.930, 0.672 to 0.926, and 0.815 to 0.847, and accuracy rates between 90.7 and 98.1%, 92.6-98.3%, and 92.6-98.1%, respectively. This novel AI system facilitates the rapid and accurate automatic Lenke classification, offering potential assistance to spinal surgeons.