The convolutional neural network(CNN)-based models have emerged as the predominant approach for medical image segmentation due to their effective inductive bias. However, their limitation lies in the lack of long-range information. In this study, we propose the Atten-Nonlocal Unet model that integrates CNN and transformer to overcome this limitation and precisely capture global context in 2D features. Specifically, we utilize the BCSM attention module and the Cross Non-local module to enhance feature representation, thereby improving the segmentation accuracy. Experimental results on the Synapse, ACDC, and AVT datasets show that Atten-Nonlocal Unet achieves DSC scores of 84.15%, 91.57%, and 86.94% respectively, and has 95% HD of 15.17, 1.16, and 4.78 correspondingly. Compared to the existing methods for medical image segmentation, the proposed method demonstrates superior segmentation performance, ensuring high accuracy in segmenting large organs while improving segmentation for small organs.

To demonstrate a method of benchmarking the performance of two consecutive software releases of the same commercial artificial intelligence (AI) product to trained human readers using the Personal Performance in Mammographic Screening scheme (PERFORMS) external quality assurance scheme. In this retrospective study, ten PERFORMS test sets, each consisting of 60 challenging cases, were evaluated by human readers between 2012 and 2023 and were evaluated by Version 1 (V1) and Version 2 (V2) of the same AI model in 2022 and 2023 respectively. Both AI and humans considered each breast independently. Both AI and humans considered the highest suspicion of malignancy score per breast for non-malignant cases and per lesion for breasts with malignancy. Sensitivity, specificity, and area under the receiver operating characteristic curve (AUC) were calculated for comparison, with the study powered to detect a medium-sized effect (odds ratio, 3.5 or 0.29) for sensitivity. The study included 1,254 human readers, with a total of 328 malignant lesions, 823 normal, and 55 benign breasts analysed. No significant difference was found between the AUCs for AI V1 (0.93) and V2 (0.94) (p = 0.13). In terms of sensitivity, no difference was observed between human readers and AI V1 (83.2 % vs 87.5 % respectively, p = 0.12), however V2 outperformed humans (88.7 %, p = 0.04). Specificity was higher for AI V1 (87.4 %) and V2 (88.2 %) compared to human readers (79.0 %, p < 0.01 respectively). The upgraded AI model showed no significant difference in diagnostic performance compared to its predecessor when evaluating mammograms from PERFORMS test sets.

Silicosis, a debilitating occupational lung disease caused by inhaling crystalline silica, continues to be a significant global health issue, especially with the increasing use of engineered stone (ES) surfaces containing high silica content. Traditional diagnostic methods, dependent on radiological interpretation, have low sensitivity, especially, in the early stages of the disease, and present variability between evaluators. This study explores the efficacy of deep learning techniques in automating the screening and staging of silicosis using chest X-ray images. Utilizing a comprehensive dataset, obtained from the medical records of a cohort of workers exposed to artificial quartz conglomerates, we implemented a preprocessing stage for rib-cage segmentation, followed by classification using state-of-the-art deep learning models. The segmentation model exhibited high precision, ensuring accurate identification of thoracic structures. In the screening phase, our models achieved near-perfect accuracy, with ROC AUC values reaching 1.0, effectively distinguishing between healthy individuals and those with silicosis. The models demonstrated remarkable precision in the staging of the disease. Nevertheless, differentiating between simple silicosis and progressive massive fibrosis, the evolved and complicated form of the disease, presented certain difficulties, especially during the transitional period, when assessment can be significantly subjective. Notwithstanding these difficulties, the models achieved an accuracy of around 81% and ROC AUC scores nearing 0.93. This study highlights the potential of deep learning to generate clinical decision support tools to increase the accuracy and effectiveness in the diagnosis and staging of silicosis, whose early detection would allow the patient to be moved away from all sources of occupational exposure, therefore constituting a substantial advancement in occupational health diagnostics.

This research presents an innovative neuromorphic method utilizing Spiking Neural Networks (SNNs) to analyze pediatric chest X-rays (PediCXR) to identify prevalent thoracic illnesses. We incorporate spiking-based machine learning models such as Spiking Convolutional Neural Networks (SCNN), Spiking Residual Networks (S-ResNet), and Hierarchical Spiking Neural Networks (HSNN), for pediatric chest radiographic analysis utilizing the publically available benchmark PediCXR dataset. These models employ spatiotemporal feature extraction, residual connections, and event-driven processing to improve diagnostic precision. The HSNN model surpasses benchmark approaches from the literature, with a classification accuracy of 96% across six thoracic illness categories, with an F1-score of 0.95 and a specificity of 1.0 in pneumonia detection. Our research demonstrates that neuromorphic computing is a feasible and biologically inspired approach to real-time medical imaging diagnostics, significantly improving performance.

Multiple sclerosis (MS) is a chronic neurological disease characterized by inflammation, demyelination, and neurodegeneration within the central nervous system. Conventional magnetic resonance imaging (MRI) techniques often struggle to detect small or subtle lesions, particularly in challenging regions such as the cortical gray matter and brainstem. This study introduces a novel deep learning-based approach, combined with a robust preprocessing pipeline and optimized MRI protocols, to improve the precision of MS lesion classification and stratification. We designed a convolutional neural network (CNN) architecture specifically tailored for high-resolution T2-weighted imaging (T2WI), augmented by deep learning-based reconstruction (DLR) techniques. The model incorporates dual attention mechanisms, including spatial and channel attention modules, to enhance feature extraction. A comprehensive preprocessing pipeline was employed, featuring bias field correction, skull stripping, image registration, and intensity normalization. The proposed framework was trained and validated on four publicly available datasets and evaluated using precision, sensitivity, specificity, and area under the curve (AUC) metrics. The model demonstrated exceptional performance, achieving a precision of 96.27 %, sensitivity of 95.54 %, specificity of 94.70 %, and an AUC of 0.975. It outperformed existing state-of-the-art methods, particularly in detecting lesions in underdiagnosed regions such as the cortical gray matter and brainstem. The integration of advanced attention mechanisms enabled the model to focus on critical MRI features, leading to significant improvements in lesion classification and stratification. This study presents a novel and scalable approach for MS lesion detection and classification, offering a practical solution for clinical applications. By integrating advanced deep learning techniques with optimized MRI protocols, the proposed framework achieves superior diagnostic accuracy and generalizability, paving the way for enhanced patient care and more personalized treatment strategies. This work sets a new benchmark for MS diagnosis and management in both research and clinical practice.

Despite the potential of large language models for radiology training, their ability to handle image-based radiological questions remains poorly understood. This study aimed to evaluate the performance of the GPT-4 Turbo and GPT-4o in radiology resident examinations, to analyze differences across question types, and to compare their results with those of residents at different levels. A total of 776 multiple-choice questions from the Korean Society of Radiology In-Training Examinations were used, forming two question sets: one originally written in Korean and the other translated into English. We evaluated the performance of GPT-4 Turbo (gpt-4-turbo-2024-04-09) and GPT-4o (gpt-4o-2024-11-20) on these questions with the temperature set to zero, determining the accuracy based on the majority vote from five independent trials. We analyzed their results using the question type (text-only vs. image-based) and benchmarked them against nationwide radiology residents' performance. The impact of the input language (Korean or English) on model performance was examined. GPT-4o outperformed GPT-4 Turbo for both image-based (48.2% vs. 41.8%, <i>P</i> = 0.002) and text-only questions (77.9% vs. 69.0%, <i>P</i> = 0.031). On image-based questions, GPT-4 Turbo and GPT-4o showed comparable performance to that of 1st-year residents (41.8% and 48.2%, respectively, vs. 43.3%, <i>P</i> = 0.608 and 0.079, respectively) but lower performance than that of 2nd- to 4th-year residents (vs. 56.0%-63.9%, all <i>P</i> ≤ 0.005). For text-only questions, GPT-4 Turbo and GPT-4o performed better than residents across all years (69.0% and 77.9%, respectively, vs. 44.7%-57.5%, all <i>P</i> ≤ 0.039). Performance on the English- and Korean-version questions showed no significant differences for either model (all <i>P</i> ≥ 0.275). GPT-4o outperformed the GPT-4 Turbo in all question types. On image-based questions, both models' performance matched that of 1st-year residents but was lower than that of higher-year residents. Both models demonstrated superior performance compared to residents for text-only questions. The models showed consistent performances across English and Korean inputs.

Proton resonance frequency (PRF)-based magnetic resonance (MR) thermometry plays a critical role in thermal ablation therapies through focused ultrasound (FUS). For clinical applications, accurate and rapid temperature feedback is essential to ensure both the safety and effectiveness of these treatments. This work aims to improve temporal resolution in dynamic MR temperature map reconstructions using an enhanced deep-learning method, thereby supporting the real-time monitoring required for effective FUS treatments. Five classical neural network architectures-cascade net, complex-valued U-Net, shift window transformer for MRI, real-valued U-Net, and U-Net with residual blocks-along with training-optimized methods were applied to reconstruct temperature maps from 2-fold and 4-fold undersampled k-space data. The training enhancements included pre-training/training-phase data augmentations, knowledge distillation, and a novel amplitude-phase decoupling loss function. Phantom and ex vivo tissue heating experiments were conducted using a FUS transducer. Ground truth was the complex MR images with accurate temperature changes, and datasets were manually undersampled to simulate such acceleration here. Separate testing datasets were used to evaluate real-time performance and temperature accuracy. Furthermore, our proposed deep learning-based rapid reconstruction approach was validated on a clinical dataset obtained from patients with uterine fibroids, demonstrating its clinical applicability. Acceleration factors of 1.9 and 3.7 were achieved for 2× and 4× k-space under samplings, respectively. The deep learning-based reconstruction using ResUNet incorporating the four optimizations, showed superior performance. For 2-fold acceleration, the RMSE of temperature map patches were 0.89°C and 1.15°C for the phantom and ex vivo testing datasets, respectively. The DICE coefficient for the 43°C isotherm-enclosed regions was 0.81, and the Bland-Altman analysis indicated a bias of -0.25°C with limits of agreement of ±2.16°C. In the 4-fold under-sampling case, these evaluation metrics showed approximately a 10% reduction in accuracy. Additionally, the DICE coefficient measuring the overlap between the reconstructed temperature maps (using the optimized ResUNet) and the ground truth, specifically in regions where the temperature exceeded the 43°C threshold, were 0.77 and 0.74 for the 2× and 4× under-sampling scenarios, respectively. This study demonstrates that deep learning-based reconstruction significantly enhances the accuracy and efficiency of MR thermometry, particularly in the context of FUS-based clinical treatments for uterine fibroids. This approach could also be extended to other applications such as essential tremor and prostate cancer treatments where MRI-guided FUS plays a critical role.

To develop and validate an automated diabetic macular oedema (DME) classification system based on the images from different three-dimensional optical coherence tomography (3-D OCT) devices. A multicentre, platform-based development study using retrospective and cross-sectional data. Data were subjected to a two-level grading system by trained graders and a retina specialist, and categorised into three types: no DME, non-centre-involved DME and centre-involved DME (CI-DME). The 3-D convolutional neural networks algorithm was used for DME classification system development. The deep learning (DL) performance was compared with the diabetic retinopathy experts. Data were collected from Joint Shantou International Eye Center of Shantou University and the Chinese University of Hong Kong, Chaozhou People's Hospital and The Second Affiliated Hospital of Shantou University Medical College from January 2010 to December 2023. 7790 volumes of 7146 eyes from 4254 patients were annotated, of which 6281 images were used as the development set and 1509 images were used as the external validation set, split based on the centres. Accuracy, F1-score, sensitivity, specificity, area under receiver operating characteristic curve (AUROC) and Cohen's kappa were calculated to evaluate the performance of the DL algorithm. In classifying DME with non-DME, our model achieved an AUROCs of 0.990 (95% CI 0.983 to 0.996) and 0.916 (95% CI 0.902 to 0.930) for hold-out testing dataset and external validation dataset, respectively. To distinguish CI-DME from non-centre-involved-DME, our model achieved AUROCs of 0.859 (95% CI 0.812 to 0.906) and 0.881 (95% CI 0.859 to 0.902), respectively. In addition, our system showed comparable performance (Cohen's κ: 0.85 and 0.75) to the retina experts (Cohen's κ: 0.58-0.92 and 0.70-0.71). Our DL system achieved high accuracy in multiclassification tasks on DME classification with 3-D OCT images, which can be applied to population-based DME screening.

To compare the image quality and pulmonary nodule detectability and measurement accuracy between deep learning reconstruction (DLR) and hybrid iterative reconstruction (HIR) of chest ultra-low-dose CT (ULDCT). Participants who underwent chest standard-dose CT (SDCT) followed by ULDCT from October 2020 to January 2022 were prospectively included. ULDCT images reconstructed with HIR and DLR were compared with SDCT images to evaluate image quality, nodule detection rate, and measurement accuracy using a commercially available deep learning-based nodule evaluation system. Wilcoxon signed-rank test was used to evaluate the percentage errors of nodule size and nodule volume between HIR and DLR images. Eighty-four participants (54 ± 13 years; 26 men) were finally enrolled. The effective radiation doses of ULDCT and SDCT were 0.16 ± 0.02 mSv and 1.77 ± 0.67 mSv, respectively (P < 0.001). The mean ± standard deviation of the lung tissue noises was 61.4 ± 3.0 HU for SDCT, 61.5 ± 2.8 HU and 55.1 ± 3.4 HU for ULDCT reconstructed with HIR-Strong setting (HIR-Str) and DLR-Strong setting (DLR-Str), respectively (P < 0.001). A total of 535 nodules were detected. The nodule detection rates of ULDCT HIR-Str and ULDCT DLR-Str were 74.0% and 83.4%, respectively (P < 0.001). The absolute percentage error in nodule volume from that of SDCT was 19.5% in ULDCT HIR-Str versus 17.9% in ULDCT DLR-Str (P < 0.001). Compared with HIR, DLR reduced image noise, increased nodule detection rate, and improved measurement accuracy of nodule volume at chest ULDCT. Not applicable.

Medical image challenges have played a transformative role in advancing the field, catalyzing algorithmic innovation and establishing new performance standards across diverse clinical applications. Image registration, a foundational task in neuroimaging pipelines, has similarly benefited from the Learn2Reg initiative. Building on this foundation, we introduce the Large-scale Unsupervised Brain MRI Image Registration (LUMIR) challenge, a next-generation benchmark designed to assess and advance unsupervised brain MRI registration. Distinct from prior challenges that leveraged anatomical label maps for supervision, LUMIR removes this dependency by providing over 4,000 preprocessed T1-weighted brain MRIs for training without any label maps, encouraging biologically plausible deformation modeling through self-supervision. In addition to evaluating performance on 590 held-out test subjects, LUMIR introduces a rigorous suite of zero-shot generalization tasks, spanning out-of-domain imaging modalities (e.g., FLAIR, T2-weighted, T2*-weighted), disease populations (e.g., Alzheimer's disease), acquisition protocols (e.g., 9.4T MRI), and species (e.g., macaque brains). A total of 1,158 subjects and over 4,000 image pairs were included for evaluation. Performance was assessed using both segmentation-based metrics (Dice coefficient, 95th percentile Hausdorff distance) and landmark-based registration accuracy (target registration error). Across both in-domain and zero-shot tasks, deep learning-based methods consistently achieved state-of-the-art accuracy while producing anatomically plausible deformation fields. The top-performing deep learning-based models demonstrated diffeomorphic properties and inverse consistency, outperforming several leading optimization-based methods, and showing strong robustness to most domain shifts, the exception being a drop in performance on out-of-domain contrasts.