Deep learning (DL), a branch of artificial intelligence (AI), has been applied to diagnose developmental dysplasia of the hip (DDH) on pelvic radiographs and ultrasound (US) images. This technology can potentially assist in early screening, enable timely intervention and improve cost-effectiveness. We conducted a systematic review to evaluate the diagnostic accuracy of the DL algorithm in detecting DDH. PubMed, Medline, EMBASE, EMCARE, the clinicaltrials.gov (clinical trial registry), IEEE Xplore and Cochrane Library databases were searched in October 2024. Prospective and retrospective cohort studies that included children (< 16 years) at risk of or suspected to have DDH and reported hip ultrasonography (US) or X-ray images using AI were included. A review was conducted using the guidelines of the Cochrane Collaboration Diagnostic Test Accuracy Working Group. Risk of bias was assessed using the QUADAS-2 tool. Twenty-three studies met inclusion criteria, with 15 (n = 8315) evaluating DDH on US images and eight (n = 7091) on pelvic radiographs. The area under the curve of the included studies ranged from 0.80 to 0.99 for pelvic radiographs and 0.90-0.99 for US images. Sensitivity and specificity for detecting DDH on radiographs ranged from 92.86% to 100% and 95.65% to 99.82%, respectively. For US images, sensitivity ranged from 86.54% to 100% and specificity from 62.5% to 100%. AI demonstrated comparable effectiveness to physicians in detecting DDH. However, limited evaluation on external datasets restricts its generalisability. Further research incorporating diverse datasets and real-world applications is needed to assess its broader clinical impact on DDH diagnosis.

Medical imaging foundation models must adapt over time, yet full retraining is often blocked by privacy constraints and cost. We present a continual learning framework that avoids storing patient exemplars by pairing class conditional diffusion replay with Elastic Weight Consolidation. Using a compact Vision Transformer backbone, we evaluate across eight MedMNIST v2 tasks and CheXpert. On CheXpert our approach attains 0.851 AUROC, reduces forgetting by more than 30\% relative to DER\texttt{++}, and approaches joint training at 0.869 AUROC, while remaining efficient and privacy preserving. Analyses connect forgetting to two measurable factors: fidelity of replay and Fisher weighted parameter drift, highlighting the complementary roles of replay diffusion and synaptic stability. The results indicate a practical route for scalable, privacy aware continual adaptation of clinical imaging models.

Current models based on deep learning for low-dose CT denoising rely heavily on paired data and generalize poorly. Even the more concerned diffusion models need to learn the distribution of clean data for reconstruction, which is difficult to satisfy in medical clinical applications. At the same time, self-supervised-based methods face the challenge of significant degradation of generalizability of models pre-trained for the current dose to expand to other doses. To address these issues, this paper proposes a novel method of tunable-generalization diffusion powered by self-supervised contextual sub-data for low-dose CT reconstruction, named SuperDiff. Firstly, a contextual subdata similarity adaptive sensing strategy is designed for denoising centered on the LDCT projection domain, which provides an initial prior for the subsequent progress. Subsequently, the initial prior is used to combine knowledge distillation with a deep combination of latent diffusion models for optimizing image details. The pre-trained model is used for inference reconstruction, and the pixel-level self-correcting fusion technique is proposed for fine-grained reconstruction of the image domain to enhance the image fidelity, using the initial prior and the LDCT image as a guide. In addition, the technique is flexibly applied to the generalization of upper and lower doses or even unseen doses. Dual-domain strategy cascade for self-supervised LDCT denoising, SuperDiff requires only LDCT projection domain data for training and testing. Full qualitative and quantitative evaluations on both datasets and real data show that SuperDiff consistently outperforms existing state-of-the-art methods in terms of reconstruction and generalization performance.

Solving tough clinical questions that require both image and text understanding is still a major challenge in healthcare AI. In this work, we propose Q-FSRU, a new model that combines Frequency Spectrum Representation and Fusion (FSRU) with a method called Quantum Retrieval-Augmented Generation (Quantum RAG) for medical Visual Question Answering (VQA). The model takes in features from medical images and related text, then shifts them into the frequency domain using Fast Fourier Transform (FFT). This helps it focus on more meaningful data and filter out noise or less useful information. To improve accuracy and ensure that answers are based on real knowledge, we add a quantum inspired retrieval system. It fetches useful medical facts from external sources using quantum-based similarity techniques. These details are then merged with the frequency-based features for stronger reasoning. We evaluated our model using the VQA-RAD dataset, which includes real radiology images and questions. The results showed that Q-FSRU outperforms earlier models, especially on complex cases needing image text reasoning. The mix of frequency and quantum information improves both performance and explainability. Overall, this approach offers a promising way to build smart, clear, and helpful AI tools for doctors.

Structural MRI of the brain is routinely performed on patients referred to memory clinics; however, resulting radiology reports, including volumetric assessments, are conventionally stored as unstructured free text. We sought to use natural language processing (NLP) to extract text relating to intracranial volumetric assessment from brain MRI text reports to enhance routine data availability for research purposes. Electronic records from a large mental healthcare provider serving a geographic catchment of 1.3 million residents in four boroughs of south London, UK. A corpus of 4007 de-identified brain MRI reports from patients referred to memory assessment services. An NLP algorithm was developed, using a span categorisation approach, to extract six binary (presence/absence) categories from the text reports: (i) global volume loss, (ii) hippocampal/medial temporal lobe volume loss and (iii) other lobar/regional volume loss. Distributions of these categories were evaluated. The overall F1 score for the six categories was 0.89 (precision 0.92, recall 0.86), with the following precision/recall for each category: presence of global volume loss 0.95/0.95, absence of global volume loss 0.94/0.77, presence of regional volume loss 0.80/0.58, absence of regional volume loss 0.91/0.93, presence of hippocampal volume loss 0.90/0.88, and absence of hippocampal volume loss 0.94/0.92. These results support the feasibility and accuracy of using NLP techniques to extract volumetric assessments from radiology reports, and the potential for automated generation of novel meta-data from dementia assessments in electronic health records.

Deep learning models are increasingly used for radiographic analysis, but their reliability is challenged by the stochastic noise inherent in clinical imaging. A systematic, cross-task understanding of how different noise types impact these models is lacking. Here, we evaluate the robustness of state-of-the-art convolutional neural networks (CNNs) to simulated quantum (Poisson) and electronic (Gaussian) noise in two key chest X-ray tasks: semantic segmentation and pulmonary disease classification. Using a novel, scalable noise injection framework, we applied controlled, clinically-motivated noise severities to common architectures (UNet, DeepLabV3, FPN; ResNet, DenseNet, EfficientNet) on public datasets (Landmark, ChestX-ray14). Our results reveal a stark dichotomy in task robustness. Semantic segmentation models proved highly vulnerable, with lung segmentation performance collapsing under severe electronic noise (Dice Similarity Coefficient drop of 0.843), signifying a near-total model failure. In contrast, classification tasks demonstrated greater overall resilience, but this robustness was not uniform. We discovered a differential vulnerability: certain tasks, such as distinguishing Pneumothorax from Atelectasis, failed catastrophically under quantum noise (AUROC drop of 0.355), while others were more susceptible to electronic noise. These findings demonstrate that while classification models possess a degree of inherent robustness, pixel-level segmentation tasks are far more brittle. The task- and noise-specific nature of model failure underscores the critical need for targeted validation and mitigation strategies before the safe clinical deployment of diagnostic AI.

Federated learning (FL) allows collaborative model training across healthcare sites without sharing sensitive patient data. However, real-world FL deployment is often hindered by complex operational challenges that demand substantial human efforts. This includes: (a) selecting appropriate clients (hospitals), (b) coordinating between the central server and clients, (c) client-level data pre-processing, (d) harmonizing non-standardized data and labels across clients, and (e) selecting FL algorithms based on user instructions and cross-client data characteristics. However, the existing FL works overlook these practical orchestration challenges. These operational bottlenecks motivate the need for autonomous, agent-driven FL systems, where intelligent agents at each hospital client and the central server agent collaboratively manage FL setup and model training with minimal human intervention. To this end, we first introduce an agent-driven FL framework that captures key phases of real-world FL workflows from client selection to training completion and a benchmark dubbed FedAgentBench that evaluates the ability of LLM agents to autonomously coordinate healthcare FL. Our framework incorporates 40 FL algorithms, each tailored to address diverse task-specific requirements and cross-client characteristics. Furthermore, we introduce a diverse set of complex tasks across 201 carefully curated datasets, simulating 6 modality-specific real-world healthcare environments, viz., Dermatoscopy, Ultrasound, Fundus, Histopathology, MRI, and X-Ray. We assess the agentic performance of 14 open-source and 10 proprietary LLMs spanning small, medium, and large model scales. While some agent cores such as GPT-4.1 and DeepSeek V3 can automate various stages of the FL pipeline, our results reveal that more complex, interdependent tasks based on implicit goals remain challenging for even the strongest models.

To evaluate the predictive performance of a clinical-CT-radiomics nomogram based on radiomics signature and independent clinical-CT predictors for predicting endoscopic red color sign (RC) in cirrhotic patients with esophageal varices (EV). We retrospectively evaluated 215 cirrhotic patients. Among them, 108 and 107 cases were positive and negative for endoscopic RC, respectively. Patients were assigned to a training cohort (n = 150) and a validation cohort (n = 65) at a 7:3 ratio. In the training cohort, univariate and multivariate logistic regression analyses were performed on clinical and CT features to develop a clinical-CT model. Radiomic features were extracted from portal venous phase CT images to generate a Radiomic score (Rad-score) and to construct five machine learning models. A combined model was built using clinical-CT predictors and Rad-score through logistic regression. The performance of different models was evaluated using the receiver operating characteristic (ROC) curves and the area under the curve (AUC). The spleen-to-platelet ratio, liver volume, splenic vein diameter, and superior mesenteric vein diameter were independent predictors. Six radiomics features were selected to construct five machine learning models. The adaptive boosting model showed excellent predictive performance, achieving an AUC of 0.964 in the validation cohort, while the combined model achieved the highest predictive accuracy with an AUC of 0.985 in the validation cohort. The clinical-CT-radiomics nomogram demonstrates high predictive accuracy for endoscopic RC in cirrhotic patients with EV, which provides a novel tool for non-invasive prediction of esophageal varices bleeding.

Inferior alveolar nerve (IAN) injury is a critical complication of mandibular third molar extraction. This study aimed to construct and evaluate a deep learning framework that integrates contrastive learning and Bayesian optimization to enhance predictive performance on cone-beam computed tomography (CBCT) and panoramic radiographs. A retrospective dataset of 902 panoramic radiographs and 1,500 CBCT images was used. Five deep learning architectures (MobileNetV2, ResNet101D, Vision Transformer, Twins-SVT, and SSL-ResNet50) were trained with and without contrastive learning and Bayesian optimization. Model performance was evaluated using accuracy, F1-score, and comparison with oral and maxillofacial surgeons (OMFSs). Contrastive learning significantly improved the F1-scores across all models (e.g., MobileNetV2: 0.302 to 0.740; ResNet101D: 0.188 to 0.689; Vision Transformer: 0.275 to 0.704; Twins-SVT: 0.370 to 0.719; SSL-ResNet50: 0.109 to 0.576). Bayesian optimization further enhanced the F1-scores for MobileNetV2 (from 0.740 to 0.923), ResNet101D (from 0.689 to 0.857), Vision Transformer (from 0.704 to 0.871), Twins-SVT (from 0.719 to 0.857), and SSL-ResNet50 (from 0.576 to 0.875). The AI model outperformed OMFSs on CBCT cross-sectional images (F1-score: 0.923 vs. 0.667) but underperformed on panoramic radiographs (0.666 vs. 0.730). The proposed single-step deep learning approach effectively predicts IAN injury, with contrastive learning addressing data imbalance and Bayesian optimization optimizing model performance. While artificial intelligence surpasses human performance in CBCT images, panoramic radiographs analysis still benefits from expert interpretation. Future work should focus on multi-center validation and explainable artificial intelligence for broader clinical adoption.

Accurate image registration is essential for downstream applications, yet current deep registration networks provide limited indications of whether and when their predictions are reliable. Existing uncertainty estimation strategies, such as Bayesian methods, ensembles, or MC dropout, require architectural changes or retraining, limiting their applicability to pretrained registration networks. Instead, we propose a test-time uncertainty estimation framework that is compatible with any pretrained networks. Our framework is grounded in the transformation equivariance property of registration, which states that the true mapping between two images should remain consistent under spatial perturbations of the input. By analyzing the variance of network predictions under such perturbations, we derive a theoretical decomposition of perturbation-based uncertainty in registration. This decomposition separates into two terms: (i) an intrinsic spread, reflecting epistemic noise, and (ii) a bias jitter, capturing how systematic error drifts under perturbations. Across four anatomical structures (brain, cardiac, abdominal, and lung) and multiple registration models (uniGradICON, SynthMorph), the uncertainty maps correlate consistently with registration errors and highlight regions requiring caution. Our framework turns any pretrained registration network into a risk-aware tool at test time, placing medical image registration one step closer to safe deployment in clinical and large-scale research settings.