A recent investigation into deep learning-based denoising for early Monte Carlo (MC) Path Tracing in computed tomography (CT) volume visualization yielded promising quantitative outcomes but inconsistent qualitative assessments. This research probes the underlying causes of this incongruity by deploying a web-based SurveyMonkey questionnaire distributed among healthcare professionals. The survey targeted radiologists, residents, orthopedic surgeons, and veterinarians, leveraging the authors' professional networks for dissemination. To evaluate perceptions, the questionnaire featured randomized sections gauging attitudes towards AI-enhanced image and video quality, confidence in reference images, and clinical applicability. Seventy-four participants took part, encompassing a spectrum of experience levels: <1 year (n=11), 1-3 years (n=27), 3-5 years (n=12), and >5 years (n=24). A substantial majority (77%) expressed a preference for AI-enhanced images over traditional MC estimates, a preference influenced by participant experience (adjusted OR 0.81, 95% CI 0.67-0.98, p=0.033). Experience correlates with confidence in AI-generated images (adjusted OR 0.98, 95% CI 0.95-1, p=0.018-0.047) and satisfaction with video previews, both with and without AI (adjusted OR 0.96-0.98, 95% CI 0.92-1, p = 0.033-0.048). Significant monotonic relationships emerged between experience, confidence (σ= 0.25-0.26, p = 0.025-0.029), and satisfaction (σ= 0.23-0.24, p = 0.037-0.046). The findings underscore the potential of AI post-processing to improve the rendering of biomedical volumes, noting enhanced confidence and satisfaction among experienced participants. The study reveals that participants' preferences may not align perfectly with quality metrics such as peak signal-to-noise ratio and structural similarity index, highlighting nuances in evaluating AI's qualitative impact on CT image denoising.

<i>"Just Accepted" papers have undergone full peer review and have been accepted for publication in <i>Radiology: Artificial Intelligence</i>. This article will undergo copyediting, layout, and proof review before it is published in its final version. Please note that during production of the final copyedited article, errors may be discovered which could affect the content</i>. Purpose To evaluate the technical feasibility of implementing homomorphic encryption in deep learning models for privacy-preserving CT image analysis of renal masses. Materials and Methods A privacy-preserving deep learning system was developed through three sequential technical phases: a reference CNN model (Ref-CNN) based on ResNet architecture, modification for encryption compatibility (Approx-CNN) by replacing ReLU with polynomial approximation and max-pooling with averagepooling, and implementation of fully homomorphic encryption (HE-CNN). The CKKS encryption scheme was used for its capability to perform arithmetic operations on encrypted real numbers. Using 12,446 CT images from a public dataset (3,709 renal cysts, 5,077 normal kidneys, and 2,283 kidney tumors), we evaluated model performance using area under the receiver operating characteristic curve (AUC) and area under the precision-recall curve (AUPRC). Results All models demonstrated high diagnostic accuracy with AUC ranging from 0.89-0.99 and AUPRC from 0.67-0.99. The diagnostic performance trade-off was minimal from Ref-CNN to Approx-CNN (AUC: 0.99 to 0.97 for normal category), with no evidence of differences between models. However, encryption significantly increased storage and computational demands: a 256 × 256-pixel image expanded from 65KB to 32MB, requiring 50 minutes for CPU inference but only 90 seconds with GPU acceleration. Conclusion This technical development demonstrates that privacy-preserving deep learning inference using homomorphic encryption is feasible for renal mass classification on CT images, achieving comparable diagnostic performance while maintaining data privacy through end-to-end encryption. ©RSNA, 2025.

ImportancePredetermined Change Control Plans (PCCPs) are a recent regulatory innovation by the U.S. Food and Drug Administration (FDA) introduced to enable dynamic oversight of artificial intelligence and machine learning (AI/ML)-enabled medical devices. ObjectiveTo characterize FDA program of PCCPs among AI/ML-enabled medical devices, including device characteristics, preapproval testing, planned modifications, and post-clearance update mechanisms. DesignThis cross-sectional study reviewed FDA-cleared or approved AI/ML-enabled medical devices with authorized PCCPs. SettingAI/ML-enabled devices approved or cleared prior to May 30, 2025 were identified from an FDA-maintained public list and their characteristics extracted from FDA approval databases. ParticipantsN/A Main Outcome(s) and Measure(s)Primary outcomes included (1) prevalence and characteristics of devices with authorized PCCPs, (2) types of FDA-authorized modifications, (3) presence and nature of preapproval testing, such as study design and subgroup testing, and (4) postmarket device update mechanisms and transparency. ResultsAmong 26 identified AI/ML-enabled medical devices with authorized PCCPs, 92% were cleared via the 510(k) pathway, and all were classified as moderate risk. Devices were primarily intended for use in diagnosis or clinical assessment, and six had consumer-facing components. Authorized modifications spanned the product lifecycle, most commonly allowing model retraining (69% of devices), logic updates (42% of devices), and expansion of input sources (35% of devices). Preapproval testing was limited with seven devices prospectively evaluated and thirteen undergoing human factors testing. Subgroup analyses were reported for eleven devices and none included patient outcomes data. No postmarket studies or recalls were identified. User manuals could be identified online for 54% of devices, though many lacked performance details or mentioned PCCPs. Conclusions and RelevanceFDA authorization of PCCPs grants manufacturers substantial flexibility to modify AI/ML-enabled devices postmarket, while preapproval testing and postmarket transparency are limited. These findings highlight the need for strengthened oversight mechanisms to ensure ongoing safety and effectiveness of rapidly evolving AI/ML-enabled technologies in clinical care.

Artificial intelligence (AI) can enhance the diagnostic performance of breast cancer imaging and improve workflow optimization, potentially mitigating excessive radiologist workload and suboptimal diagnostic accuracy. AI can also boost imaging capabilities through individual risk prediction, molecular subtyping, and neoadjuvant therapy response predictions. Evidence demonstrates AI's potential across multiple modalities. The most robust data come from mammographic screening, where AI models improve diagnostic accuracy and optimize workflow, but rigorous post-market surveillance is required before any implementation strategy in this field. Commercial tools for digital breast tomosynthesis and ultrasound, potentially able to reduce interpretation time and improve accuracy, are also available, but post-implementation evaluation studies are likewise lacking. Besides basic tools for breast MRI with limited proven clinical benefit, AI applications for other modalities are not yet commercially available. Applications in contrast-enhanced mammography are still in the research stage, especially for radiomics-based molecular subtype classification. Applications of Large Language Models (LLMs) are in their infancy, and there are currently no clinical applications. Consequently, and despite their promise, all commercially available AI tools for breast imaging should currently still be regarded as techniques that, at best, aid radiologists in image evaluation. Their use is therefore optional, and the findings may always be overruled. KEY POINTS: AI systems improve diagnostic accuracy and efficiency of mammography screening, but long-term outcomes data are lacking. Commercial tools for digital breast tomosynthesis and ultrasound are available, but post-implementation evaluation studies are lacking. AI tools for breast imaging should still be regarded as a non-obligatory aid to radiologists for image interpretation.

Recent studies have shown that Machine Learning (ML) models can exhibit bias in real-world scenarios, posing significant challenges in ethically sensitive domains such as healthcare. Such bias can negatively affect model fairness, model generalization abilities and further risks amplifying social discrimination. There is a need to remove biases from trained models. Existing debiasing approaches often necessitate access to original training data and need extensive model retraining; they also typically exhibit trade-offs between model fairness and discriminative performance. To address these challenges, we propose Soft-Mask Weight Fine-Tuning (SWiFT), a debiasing framework that efficiently improves fairness while preserving discriminative performance with much less debiasing costs. Notably, SWiFT requires only a small external dataset and only a few epochs of model fine-tuning. The idea behind SWiFT is to first find the relative, and yet distinct, contributions of model parameters to both bias and predictive performance. Then, a two-step fine-tuning process updates each parameter with different gradient flows defined by its contribution. Extensive experiments with three bias sensitive attributes (gender, skin tone, and age) across four dermatological and two chest X-ray datasets demonstrate that SWiFT can consistently reduce model bias while achieving competitive or even superior diagnostic accuracy under common fairness and accuracy metrics, compared to the state-of-the-art. Specifically, we demonstrate improved model generalization ability as evidenced by superior performance on several out-of-distribution (OOD) datasets.

This narrative review provides a comprehensive and structured overview of recent advances in the application of artificial intelligence (AI) to medical imaging for tumor diagnosis and treatment. By synthesizing evidence from recent literature and clinical reports, we highlight the capabilities, limitations, and translational potential of AI techniques across key imaging modalities such as CT, MRI, and PET. Deep learning (DL) and radiomics have facilitated automated lesion detection, tumour segmentation, and prognostic assessments, improving early cancer detection across various malignancies, including breast, lung, and prostate cancers. AI-driven multi-modal imaging fusion integrates radiomics, genomics, and clinical data, refining precision oncology strategies. Additionally, AI-assisted radiotherapy planning and adaptive dose optimisation have enhanced therapeutic efficacy while minimising toxicity. However, challenges persist regarding data heterogeneity, model generalisability, regulatory constraints, and ethical concerns. The lack of standardised datasets and explainable AI (XAI) frameworks hinders clinical adoption. Future research should focus on improving AI interpretability, fostering multi-centre dataset interoperability, and integrating AI with molecular imaging and real-time clinical decision support. Addressing these challenges will ensure AI's seamless integration into clinical oncology, optimising cancer diagnosis, prognosis, and treatment outcomes.

Artificial intelligence in medical imaging has grown rapidly in the past decade, driven by advances in deep learning and widespread access to computing resources. Applications cover diverse imaging modalities, including those based on electromagnetic radiation (e.g., X-rays), subatomic particles (e.g., nuclear imaging), and acoustic waves (ultrasound). Each modality features and limitations are defined by its underlying physics. However, many artificial intelligence practitioners lack a solid understanding of the physical principles involved in medical image acquisition. This gap hinders leveraging the full potential of deep learning, as incorporating physics knowledge into artificial intelligence systems promotes trustworthiness, especially in limited data scenarios. This work reviews the fundamental physical concepts behind medical imaging and examines their influence on recent developments in artificial intelligence, particularly, generative models and reconstruction algorithms. Finally, we describe physics-informed machine learning approaches to improve feature learning in medical imaging.

To determine whether deep learning (DL)-based image quality (IQ) assessment of T2-weighted images (T2WI) could be biased by the presence of clinically significant prostate cancer (csPCa). In this three-center retrospective study, five abdominal radiologists categorized IQ of 2,105 transverse T2WI series into optimal, mild, moderate, and severe degradation. An IQ classification model was developed using 1,719 series (development set). The agreement between the model and radiologists was assessed using the remaining 386 series with a quadratic weighted kappa. The model was applied to 11,723 examinations that were not included in the development set and without documented prostate cancer at the time of MRI (patient age, 65.5 ± 8.3 years [mean ± standard deviation]). Examinations categorized as mild to severe degradation were used as target groups, whereas those as optimal were used to construct matched control groups. Case-control matching was performed to mitigate the effects of pre-MRI confounding factors, such as age and prostate-specific antigen value. The proportion of patients with csPCa was compared between the target and matched control groups using the chi-squared test. The agreement between the model and radiologists was moderate with a quadratic weighted kappa of 0.53. The mild-moderate IQ-degraded groups had significantly higher csPCa proportions than the matched control groups with optimal IQ: moderate (N = 126) vs. optimal (N = 504), 26.3% vs. 22.7%, respectively, difference = 3.6% [95% confidence interval: 0.4%, 6.8%], p = 0.03; mild (N = 1,399) vs. optimal (N = 1,399), 22.9% vs. 20.2%, respectively, difference = 2.7% [0.7%, 4.7%], p = 0.008. The DL-based IQ tended to be worse in patients with csPCa, raising concerns about its clinical application.

Intracranial hypertension (IH) is a life-threatening complication that may occur after acute brain injury. Early recognition of IH allows prompt interventions that improve outcomes. Even if invasive intracranial monitoring is considered the gold standard for the most severely injured patients, scarce availability of resources, the need for advanced skills, and potential for complications often limit its utilization. On the other hand, different non-invasive methods to evaluate acutely brain-injured patients for elevated intracranial pressure have been investigated. Clinical examination and neuroradiology represent the cornerstone of a patient's evaluation in the intensive care unit (ICU). However, multimodal neuromonitoring, employing widely used different tools, such as brain ultrasound, automated pupillometry, and skull micro-deformation recordings, increase the possibility for continuous or semi-continuous intracranial pressure monitoring. Furthermore, artificial intelligence (AI) has been investigated to as a tool to predict elevated intracranial pressure, shedding light on new diagnostic and treatment horizons with the potential to improve patient outcomes. This narrative review, based on a systematic literature search, summarizes the best available evidence on the use of non-invasive monitoring tools and methods for the assessment of intracranial pressure.

The COVID-19 pandemic emphasised necessity for prompt, precise diagnostics, secure data storage, and robust privacy protection in healthcare. Existing diagnostic systems often suffer from limited transparency, inadequate performance, and challenges in ensuring data security and privacy. The research proposes a novel privacy-preserving diagnostic framework, Heterogeneous Convolutional-recurrent attention Transfer learning based ResNeXt with Modified Greater Cane Rat optimisation (HCTR-MGR), that integrates deep learning, Explainable Artificial Intelligence (XAI), and blockchain technology. The HCTR model combines convolutional layers for spatial feature extraction, recurrent layers for capturing spatial dependencies, and attention mechanisms to highlight diagnostically significant regions. A ResNeXt-based transfer learning backbone enhances performance, while the MGR algorithm improves robustness and convergence. A trust-based permissioned blockchain stores encrypted patient metadata to ensure data security and integrity and eliminates centralised vulnerabilities. The framework also incorporates SHAP and LIME for interpretable predictions. Experimental evaluation on two benchmark chest X-ray datasets demonstrates superior diagnostic performance, achieving 98-99% accuracy, 97-98% precision, 95-97% recall, 99% specificity, and 95-98% F1-score, offering a 2-6% improvement over conventional models such as ResNet, SARS-Net, and PneuNet. These results underscore the framework's potential for scalable, secure, and clinically trustworthy deployment in real-world healthcare systems.