Various deep learning models have been developed and employed for medical image classification. This study conducted comprehensive experiments on 12 models, aiming to establish reliable benchmarks for research on breast dynamic contrast-enhanced magnetic resonance imaging image classification. Twelve deep learning models were systematically compared by analyzing variations in 4 key hyperparameters: optimizer (Op), learning rate, batch size (BS), and data augmentation. The evaluation criteria encompassed a comprehensive set of metrics including accuracy (Ac), loss value, precision, recall rate, F1-score, and area under the receiver operating characteristic curve. Furthermore, the training times and model parameter counts were assessed for holistic performance comparison. Adjustments in the BS within Adam Op had a minimal impact on Ac in the convolutional neural network models. However, altering the Op and learning rate while maintaining the same BS significantly affected the Ac. The ResNet152 network model exhibited the lowest Ac. Both the recall rate and area under the receiver operating characteristic curve for the ResNet152 and Vision transformer-base (ViT) models were inferior compared to the others. Data augmentation unexpectedly reduced the Ac of ResNet50, ResNet152, VGG16, VGG19, and ViT models. The VGG16 model boasted the shortest training duration, whereas the ViT model, before data augmentation, had the longest training time and smallest model weight. The ResNet152 and ViT models were not well suited for image classification tasks involving small breast dynamic contrast-enhanced magnetic resonance imaging datasets. Although data augmentation is typically beneficial, its application should be approached cautiously. These findings provide important insights to inform and refine future research in this domain.

Glaucoma is increasingly recognized as a neurodegenerative condition involving both retinal and central nervous system structures. Here, we present an integrated framework that combines MK-Curve-corrected diffusion kurtosis imaging (DKI), tractometry, and deep autoencoder-based normative modeling to detect localized white matter abnormalities associated with glaucoma. Using UK Biobank diffusion MRI data, we show that MK-Curve approach corrects anatomically implausible values and improves the reliability of DKI metrics - particularly mean (MK), radial (RK), and axial kurtosis (AK) - in regions of complex fiber architecture. Tractometry revealed reduced MK in glaucoma patients along the optic radiation, inferior longitudinal fasciculus, and inferior fronto-occipital fasciculus, but not in a non-visual control tract, supporting disease specificity. These abnormalities were spatially localized, with significant changes observed at multiple points along the tracts. MK demonstrated greater sensitivity than MD and exhibited altered distributional features, reflecting microstructural heterogeneity not captured by standard metrics. Node-wise MK values in the right optic radiation showed weak but significant correlations with retinal OCT measures (ganglion cell layer and retinal nerve fiber layer thickness), reinforcing the biological relevance of these findings. Deep autoencoder-based modeling further enabled subject-level anomaly detection that aligned spatially with group-level changes and outperformed traditional approaches. Together, our results highlight the potential of advanced diffusion modeling and deep learning for sensitive, individualized detection of glaucomatous neurodegeneration and support their integration into future multimodal imaging pipelines in neuro-ophthalmology.

Medical segmentation plays an important role in clinical applications like radiation therapy and surgical guidance, but acquiring clinically acceptable results is difficult. In recent years, progress has been witnessed with the success of utilizing transformer-like models, such as combining the attention mechanism with CNN. In particular, transformer-based segmentation models can extract global information more effectively, compensating for the drawbacks of CNN modules that focus on local features. However, utilizing transformer architecture is not easy, because training transformer-based models can be resource-demanding. Moreover, due to the distinct characteristics in the medical field, especially when encountering mid-sized and small organs with compact regions, their results often seem unsatisfactory. For example, using ViT to segment medical images directly only gives a DSC of less than 50\%, which is far lower than the clinically acceptable score of 80\%. In this paper, we used Mask2Former with deformable attention to reduce computation and proposed offset adjustment strategies to encourage sampling points within the same organs during attention weights computation, thereby integrating compact foreground information better. Additionally, we utilized the 4th feature map in Mask2Former to provide a coarse location of organs, and employed an FCN-based auxiliary head to help train Mask2Former more quickly using Dice loss. We show that our model achieves SOTA (State-of-the-Art) performance on the HaNSeg and SegRap2023 datasets, especially on mid-sized and small organs.Our code is available at link https://github.com/earis/Offsetadjustment\_Background-location\_Decoder\_Mask2former.

To facilitate implementation of plan-of-the-day (POTD) selection for treating locally advanced cervical cancer (LACC), we developed a POTD assessment tool for CBCT-guided radiotherapy (RT). A female pelvis segmentation model (U-Seg3) is combined with a quantitative standard operating procedure (qSOP) to identify optimal and acceptable plans. 

Approach: The planning CT[i], corresponding structure set[ii], and manually contoured CBCTs[iii] (n=226) from 39 LACC patients treated with POTD (n=11) or non-adaptive RT (n=28) were used to develop U-Seg3, an algorithm incorporating deep-learning and deformable image registration techniques to segment the low-risk clinical target volume (LR-CTV), high-risk CTV (HR-CTV), bladder, rectum, and bowel bag. A single-channel input model (iii only, U-Seg1) was also developed. Contoured CBCTs from the POTD patients were (a) reserved for U-Seg3 validation/testing, (b) audited to determine optimal and acceptable plans, and (c) used to empirically derive a qSOP that maximised classification accuracy. 

Main Results: The median [interquartile range] DSC between manual and U-Seg3 contours was 0.83 [0.80], 0.78 [0.13], 0.94 [0.05], 0.86[0.09], and 0.90 [0.05] for the LR-CTV, HR-CTV, bladder, rectum, and bowel bag. These were significantly higher than U-Seg1 in all structures but bladder. The qSOP classified plans as acceptable if they met target coverage thresholds (LR-CTV≧99%, HR-CTV≧99.8%), with lower LR-CTV coverage (≧95%) sometimes allowed. The acceptable plan minimising bowel irradiation was considered optimal unless substantial bladder sparing could be achieved. With U-Seg3 embedded in the qSOP, optimal and acceptable plans were identified in 46/60 and 57/60 cases. 

Significance: U-Seg3 outperforms U-Seg1 and all known CBCT-based female pelvis segmentation models. The tool combining U-Seg3 and the qSOP identifies optimal plans with equivalent accuracy as two observers. In an implementation strategy whereby this tool serves as the second observer, plan selection confidence and decision-making time could be improved whilst simultaneously reducing the required number of POTD-trained radiographers by 50%.

&#xD.

To investigate the effects of the deep learning model RaDynPET on fourfold reduced-count whole-body PET examinations. A total of 120 patients (84 internal cohorts and 36 external cohorts) undergoing ¹⁸F-FDG PET/CT examinations were enrolled. PET images were reconstructed using OSEM algorithm with 120-s (G120) and 30-s (G30) list-mode data. RaDynPET was developed to generate enhanced images (R30) from G30. Two experienced nuclear medicine physicians independently evaluated subjective image quality using a 5-point Likert scale. Standardized uptake values (SUV), standard deviations, liver signal-to-noise ratio (SNR), lesion tumor-to-background ratio (TBR), and contrast-to-noise ratio (CNR) were compared. Subgroup analyses evaluated performance across demographics, and lesion detectability were evaluated using external datasets. RaDynPET was also compared to other deep learning methods. In internal cohorts, R30 demonstrated significantly higher image quality scores than G30 and G120. R30 showed excellent agreement with G120 for liver and lesion SUV values and surpassed G120 in liver SNR and CNR. Liver SNR and CNR of R30 were comparable to G120 in thin group, and the CNR of R30 was comparable to G120 in young age group. In external cohorts, R30 maintained strong SUV agreement with G120, with lesion-level sensitivity and specificity of 95.45% and 98.41%, respectively. There was no statistical difference in lesion detection between R30 and G120. RaDynPET achieved the highest PSNR and SSIM among deep learning methods. The RaDynPET model effectively restored high image quality while maintaining SUV agreement for ¹⁸F-FDG PET scans acquired in 25% of the standard acquisition time.

To assess the suitability of Transformer-based architectures for medical image segmentation and investigate the potential advantages of Graph Neural Networks (GNNs) in this domain. We analyze the limitations of the Transformer, which models medical images as sequences of image patches, limiting its flexibility in capturing complex and irregular tumor structures. To address it, we propose U-GNN, a pure GNN-based U-shaped architecture designed for medical image segmentation. U-GNN retains the U-Net-inspired inductive bias while leveraging GNNs' topological modeling capabilities. The architecture consists of Vision GNN blocks stacked into a U-shaped structure. Additionally, we introduce the concept of multi-order similarity and propose a zero-computation-cost approach to incorporate higher-order similarity in graph construction. Each Vision GNN block segments the image into patch nodes, constructs multi-order similarity graphs, and aggregates node features via multi-order node information aggregation. Experimental evaluations on multi-organ and cardiac segmentation datasets demonstrate that U-GNN significantly outperforms existing CNN- and Transformer-based models. U-GNN achieves a 6% improvement in Dice Similarity Coefficient (DSC) and an 18% reduction in Hausdorff Distance (HD) compared to state-of-the-art methods. The source code will be released upon paper acceptance.

Self-supervised methods for magnetic resonance imaging (MRI) reconstruction have garnered significant interest due to their ability to address the challenges of slow data acquisition and scarcity of fully sampled labels. Current regularization-based self-supervised techniques merge the theoretical foundations of regularization with the representational strengths of deep learning and enable effective reconstruction under higher acceleration rates, yet often fall short in interpretability, leaving their theoretical underpinnings lacking. In this paper, we introduce a novel self-supervised approach that provides stringent theoretical guarantees and interpretable networks while circumventing the need for fully sampled labels. Our method exploits the intrinsic relationship between convolutional neural networks and the null space within structural low-rank models, effectively integrating network parameters into an iterative reconstruction process. Our network learns gradient descent steps of the projected gradient descent algorithm without changing its convergence property, which implements a fully interpretable unfolded model. We design a non-expansive mapping for the network architecture, ensuring convergence to a fixed point. This well-defined framework enables complete reconstruction of missing <math xmlns="http://www.w3.org/1998/Math/MathML"><semantics><mi>k</mi> <annotation>$k$</annotation></semantics> </math> -space data grounded in matrix completion theory, independent of fully sampled labels. Qualitative and quantitative experimental results on multi-coil MRI reconstruction demonstrate the efficacy of our self-supervised approach, showing marked improvements over existing self-supervised and traditional regularization methods, achieving results comparable to supervised learning in selected scenarios. Our method surpasses existing self-supervised approaches in reconstruction quality and also delivers competitive performance under supervised settings. This work not only advances the state-of-the-art in MRI reconstruction but also enhances interpretability in deep learning applications for medical imaging.

Autism and attention deficit hyperactivity disorder (ADHD) are two highly heterogeneous neurodevelopmental conditions with variable underlying neurobiology. Imaging studies have yielded varied results, and it is now clear that there is unlikely to be one characteristic neuroanatomical profile of either condition. Parsing this heterogeneity could allow us to identify more homogeneous subgroups, either within or across conditions, which may be more clinically informative. This has been a pivotal goal for neurodevelopmental research using both clinical and neuroanatomical features, though results thus far have again been inconsistent with regards to the number and characteristics of subgroups. Here, we use population modelling to cluster a multi-site dataset based on global and regional centile scores of cortical thickness, surface area and grey matter volume. We use HYDRA, a novel semi-supervised machine learning algorithm which clusters based on differences to controls and compare its performance to a traditional clustering approach. We identified distinct subgroups within autism and ADHD, as well as across diagnosis, often with opposite neuroanatomical alterations relatively to controls. These subgroups were characterised by different combinations of increased or decreased patterns of morphometrics. We did not find significant clinical differences across subgroups. Crucially, however, the number of subgroups and their membership differed vastly depending on chosen features and the algorithm used, highlighting the impact and importance of careful method selection. We highlight the importance of examining heterogeneity in autism and ADHD and demonstrate that population modelling is a useful tool to study subgrouping in autism and ADHD. We identified subgroups with distinct patterns of alterations relative to controls but note that these results rely heavily on the algorithm used and encourage detailed reporting of methods and features used in future studies.

<b>BACKGROUND.</b> The importance of including the thoracic aortic calcification (TAC), in addition to coronary artery calcification (CAC), in prognostic assessments has been difficult to determine, partly due to greater challenge in performing standardized TAC assessments. <b>OBJECTIVE.</b> The purpose of this study was to evaluate long-term prognostic implications of TAC assessed using artificial intelligence (AI)-based quantification on routine chest CT in a screening population. <b>METHODS.</b> This retrospective study included 7404 asymptomatic individuals (median age, 53.9 years; 5875 men, 1529 women) who underwent nongated noncontrast chest CT as part of a national general health screening program at one of two centers from January 2007 to December 2014. A commercial AI program quantified TAC and CAC using Agatston scores, which were stratified into categories. Radiologists manually quantified TAC and CAC in 2567 examinations. The role of AI-based TAC categories in predicting major adverse cardiovascular events (MACE) and all-cause mortality (ACM), independent of AI-based CAC categories as well as clinical and laboratory variables, was assessed by multivariable Cox proportional hazards models using data from both centers and concordance statistics from prognostic models developed and tested using center 1 and center 2 data, respectively. <b>RESULTS.</b> AI-based and manual quantification showed excellent agreement for TAC and CAC (concordance correlation coefficient: 0.967 and 0.895, respectively). The median observation periods were 7.5 years for MACE (383 events in 5342 individuals) and 11.0 years for ACM (292 events in 7404 individuals). When adjusted for AI-based CAC categories along with clinical and laboratory variables, the risk for MACE was not independently associated with any AI-based TAC category; risk of ACM was independently associated with AI-based TAC score of 1001-3000 (HR = 2.14, <i>p</i> = .02) but not with other AI-based TAC categories. When prognostic models were tested, the addition of AI-based TAC categories did not improve model fit relative to models containing clinical variables, laboratory variables, and AI-based CAC categories for MACE (concordance index [C-index] = 0.760-0.760, <i>p</i> = .81) or ACM (C-index = 0.823-0.830, <i>p</i> = .32). <b>CONCLUSION.</b> The addition of TAC to models containing CAC provided limited improvement in risk prediction in an asymptomatic screening population undergoing CT. <b>CLINICAL IMPACT.</b> AI-based quantification provides a standardized approach for better understanding the potential role of TAC as a predictive imaging biomarker.

<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> Large language models (LLMs) hold substantial promise in addressing the growing workload in radiology, but recent studies also reveal limitations, such as hallucinations and opacity in sources for LLM responses. Retrieval-augmented Generation (RAG) based LLMs offer a promising approach to streamline radiology workflows by integrating reliable, verifiable, and customizable information. Ongoing refinement is critical to enable RAG models to manage large amounts of input data and to engage in complex multiagent dialogues. This report provides an overview of recent advances in LLM architecture, including few-shot and zero-shot learning, RAG integration, multistep reasoning, and agentic RAG, and identifies future research directions. Exemplary cases demonstrate the practical application of these techniques in radiology practice. ©RSNA, 2025.