Purpose:The rapid development of medical imaging artificial intelligence (AI) models has raised concerns over ensuring that the outputs of these models, when deployed in clinical environments, provide unbiased classifications across different patient populations or operational factors. This scoping review aimed to investigate how the current literature evaluates algorithmic bias in medical imaging AI models using quantitative fairness metrics. Approach:Following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews (PRISMA-ScR), PubMed and Embase were systematically searched for peer-reviewed literature published between January 2015 and May 2025. Studies that evaluated classification medical imaging AI models for algorithmic bias and reported at least one quantitative measure of fairness were included. Articles that investigated fairness in nonimaging AI models used qualitative bias mitigation methods, examined technical bias (e.g., systematic measurement or preprocessing), or were conference proceedings were excluded. Results:After removing duplicates, 2557 studies were included for title and abstract screening. Following initial screening, 143 articles were retrieved for full-text review. Of these, eight studies were eligible for inclusion. Half of the articles focused on bias detection, whereas the other half investigated mitigation techniques. There was considerable heterogeneity in the bias metrics used across the studies. Nearly all studies ( 7 / 8 ) used a chest X-ray radiography classifier for evaluation and examined algorithmic bias with respect to gender, age, and race. Conclusions:The relative paucity of existing literature and variability in fairness evaluation methods underscore the need for more standardized frameworks to evaluate and ensure fairness in imaging AI systems.
Purpose:Virtual staining applies computational methods to transform optical images of biological samples into histology-like representations suitable for interpretation and analysis. Existing methods for evaluating virtual staining often prioritize metrics that do not provide a complete assessment of image quality for a given biomedical or scientific context. Approach:We review existing approaches to conduct image quality assessment (IQA) for virtual staining and identify their limitations. We make the case for context-specific IQA and propose a checklist and guideline for the comprehensive evaluation of image quality of a virtual staining system. Results:We present a context-specific IQA checklist and guideline for virtually stained images that (1) defines a specific context of use (COU), (2) explains the underlying mechanisms as they relate to COU, and (3) provides COU-specific evidence for validation. We build upon existing methods and connect IQA to the underlying imaging methodology and biological truth. Conclusions:The resulting checklist and guideline link engineering evaluation to biomedical utility and aim to enable more reliable development and evaluation of virtual staining systems.
PurposeIn clinical scenarios where only a single labeled reference image is available for the target segmentation task, registration-based one-shot segmentation is a practical solution through anatomical label transfer. However, existing deep learning-based registration methods are often sensitive to appearance variations, which distort anatomical correspondence in the latent feature space and lead to structurally inconsistent deformation fields. This work aims to improve the structural consistency of deformable registration by explicitly aligning inter-subject anatomical representations.ApproachWe propose CoReg-Net, a contrastive registration framework that integrates an inter-subject contrastive learning (ISCL) module to enforce anatomically consistent feature representations by matching spatially corresponding regions across subjects. In parallel, a cascaded skip connection (CSC) mechanism is introduced to bridge the semantic gap between hierarchical encoder and decoder features, enabling more coherent reconstruction of fine-grained deformation details. These two components are jointly optimized to guide the network toward structurally meaningful and topologically regular deformation fields for robust label propagation.ResultsThe proposed method demonstrates robust and consistent one-shot segmentation performance across diverse anatomical structures and imaging modalities on the IXI, OASIS, and BCV datasets. Compared with state-of-the-art registration-based baselines, CoReg-Net generates deformation fields with higher structural similarity to target images, confirming the effectiveness of contrastive feature alignment. Atlas replacement experiments further verify the robustness of the proposed method under atlas variation. Moreover, the framework exhibits adaptive deformability, preserving topological regularity in brain tissues while accurately capturing large non-rigid deformations in abdominal organs.ConclusionsCoReg-Net shows that explicitly enforcing structural consistency through contrastive registration improves structural similarity and maintains segmentation accuracy. The proposed framework provides a reliable and generalizable solution for atlas-guided anatomical analysis across heterogeneous imaging modalities. Code is available at https://github.com/RelMono/CoRegNet.
Purpose:We describe a method for the quantitative analysis of calcification structures in fibrocartilage samples from human hip acetabular labrum, using a lab-based spectral micro-computed tomography ( μ CT ) system. Approach:Tissue samples from 26 patients with osteoarthritis and femoroacetabular impingement were analyzed using a spectral μ CT system featuring photon-counting detector with a 62 μ m pixel size operating at two energy thresholds. Material decomposition was applied to the μ CT images to separate calcium and water-like components, resulting in the 3D spatial distribution and quantification of calcium deposits within the tissue. Results were compared with histological analysis using Alizarin Red™ staining. Two calcium crystal deposition scores on a four-grade scale were assigned based on histological and μ CT analysis, respectively. Results:This method provides 3D images with a voxel size of 20 μ m over a sample size up to 3 cm. The minimum detectable calcium concentration is 0.24 g / cm 3 , and calcified structures as small as 49.6 μ m in diameter were identified. The quantification accuracy was validated using a hydroxyapatite calibration phantom. Across the complete sample set, spectral µCT grading demonstrated significant correlation (Spearman correlation coefficient ρ = 0.67 ) and substantial agreement (weighted Cohen's coefficient κ = 0.70 ) with histological assessment. Moreover, the calcification spatial distribution exhibited very good consistency with histological results, supporting the method validity. Conclusions:This approach offers a nondestructive and reproducible alternative to histology for assessing calcification in joint tissue, providing quantitative 3D mapping of calcium concentration in intact samples without the need for staining or physical sectioning.
Purpose:Medical care in battlefield and austere environments may require minimally trained personnel to perform complex diagnostic procedures such as ultrasound. We present a mixed reality (MR) guidance system that uses predictive anatomical modeling to assist novice operators in acquiring the extended Focused Assessment with Sonography in Trauma (eFAST). Approach:A morphable anatomical atlas of the torso was constructed from 180 CT scans using automated segmentation, deformable registration, and regression-based mapping between internal anatomy and external anthropometry. The atlas was integrated with a portable ultrasound system and an MR headset with probe tracking to provide subject-specific holographic guidance during scanning. System accuracy was validated using imaging phantoms and evaluated in a study involving 16 novice operators performing six-window eFAST examinations using both conventional ultrasound and MR guidance across four healthy volunteers. Ultrasound images were graded by an expert clinician, and operator workload was assessed using the NASA Task Load Index. Results:Atlas performance showed high segmentation accuracy (Dice score 0.93 ± 0.04 ) and in-atlas prediction errors below 10 mm using six anthropometric measures. Phantom testing demonstrated end-to-end image localization accuracy of 2.2 ± 0.6 mm . MR guidance was associated with a nonsignificant increase in correct-window acquisition (48% to 60%) and a significant improvement in diagnostic sufficiency (13% to 25%) while reducing perceived cognitive workload. Conclusions:These results demonstrate the feasibility of anatomically informed MR guidance to support ultrasound acquisition by novice users and highlight its potential for trauma diagnostics in resource-limited or austere environments.
The editorial announces two pilot initiatives linking SPIE Medical Imaging and the Journal of Medical Imaging: a journal-first conference presentation pathway for recently published JMI papers and a streamlined route for selected conference papers to progress to journal publication while preserving rigorous peer review.
Purpose:In this study, we present the refinement of a Mask R-CNN model initially designed for carotid lumen detection to automatically generate bounding boxes (BB) enclosing atherosclerotic plaque. Although the model also produces segmentation masks within these bounding boxes, this study primarily evaluates bounding box detection performance as the key step to support segmentation in our ultrasound elastography workflow. Approach:We utilize a PyTorch torchvision implementation of Mask R-CNN for carotid plaque detection and BB placement. Our dataset consists of 118 severe stenotic carotid plaques from patients clinically indicated for carotid endarterectomy. Due to variability plaque presentation, different R-CNN models showed varying results based on the allowed number of prediction regions. An overview analysis of shared predictions from these models showed slight improvement over individual model results. Results:Using bounding box detection as the primary endpoint, we achieved a maximum Dice similarity coefficient (DSC) of 0.74 and intersection over union (IoU) of 0.61 for the best-performing model, with a filtered multimodel approach improving DSC to 0.76. The corresponding plaque mask segmentation performance was lower ( DSC = 0.64 , IoU = 0.49 ), reflecting the increased difficulty of precise plaque delineation. Conclusion:Due to significant variation in plaque presentation and types among patients, the accuracy of the Plaque Mask R-CNN network would benefit from incorporating additional patient datasets to increase variation in the training data.
Purpose:We are developing a monolithic deep-silicon photon-counting sensor targeting spatial resolution on the order of 1 μ m . This work investigates how pixel pitch, noise level, threshold number, and threshold placement affect the achievable tangential and wafer-thickness resolution to guide the CMOS electronics and sensor design. Approach:Allpix Squared simulations were used to evaluate pixel pitches with two noise levels and two threshold-placement schemes over varying threshold numbers. Interaction position was estimated in both dimensions using lookup tables trained on the thresholded pixel output, and performance was quantified using the MTF. Results:Decreasing pixel pitch strongly improved tangential resolution for Compton interactions, whereas photoelectric interactions showed a weaker dependence on pixel pitch and threshold placement. Equal-counting thresholds approached the achievable resolution with fewer thresholds than equidistant thresholds. For eight equal-counting thresholds, the 10% MTF across the studied pixel sizes ranged from 1284 lp / cm to 2452 lp / cm in the tangential direction and from 15 to 71 lp / cm in the wafer-thickness direction for Compton interactions, compared with 409 to 507 lp / cm and 18 to 79 lp / cm , respectively, for photoelectric interactions. RMSE analysis further showed that some interactions for the 25 μ m pitch reached below 2 μ m tangentially and a few tens of micrometers in the wafer-thickness direction. Conclusions:Spatial resolution improved with a threshold number up to about 8 to 10, beyond which additional programmable thresholds gave only modest gain. The achieved tangential resolution also suggests that analyzer-free phase-contrast imaging is feasible for realistic geometries and expected interference patterns.
Purpose:Unsupervised anomaly detection allows identifying anomalies from unlabeled data, making it useful for neuroimaging analysis and computer-aided diagnosis. Given an individual's scan, we use a generative model to construct a subject-specific image of healthy appearance and compare both images with identify anomalies. Such approach has drawbacks as the reconstructions are imperfect, and model variability is not taken into account. Approach:We study model variability arising from using different random seeds during training and explore strategies to mitigate the effect of unwanted reconstruction errors and variability. The strategies include model ensembling, anomaly map normalization, and anomaly map designs based on single or multiple pseudo-healthy reconstructions. We compare these approaches in the context of dementia-related anomalies on 3D FDG PET from ADNI using variational autoencoder models. Results:Our experiments suggest that variance between models can be reduced by aggregating their reconstructions in a Z -score based anomaly map. This strategy is highly effective, but computationally expensive, as it requires training several instances of the model. An alternative strategy is to normalize the anomaly map using statistics computed from a healthy validation set. We show that this normalization strategy substantially reduces performance variability across models and even increases anomaly detection performance in certain cases. Conclusion:We highlight a largely overlooked source of variability in deep generative models, their random seed, which can lead to substantially different anomaly detection performance and biased evaluation. Our experimental analysis emphasizes that it is crucial to design robust anomaly maps and to mitigate the impact of randomness-induced variability, by accounting for model reconstruction errors for instance using Z -score ensembling, or healthy-control normalization, to support more stable and trustworthy deployment in real-world clinical practice.
Purpose:Assessing treatment response in bone metastases from non-small cell lung cancer (NSCLC) remains a major clinical challenge, particularly for patients receiving immune checkpoint inhibitors (ICIs). The existing response criteria are not optimized for osseous disease, leading to inconsistent evaluation. We aimed to develop and validate a radiomics-based machine learning (ML) framework to non-invasively distinguish immunotherapy response categories-progression, stable disease, and partial response-in NSCLC patients with bone metastases. Approach:Chest computed tomography (CT) scans from 99 NSCLC patients were analyzed before and during ICI therapy. Bone structures were automatically segmented using TotalSegmentator, and 1051 radiomic features were extracted per time point. Clinical variables were incorporated as optional features. Three ML classifiers-random forest, XGBoost, and support vector machine-were trained using fivefold cross-validation. A multistep feature selection pipeline (correlation filtering, mutual information, recursive feature elimination, and ReliefF ranking) was applied. Model performance was evaluated using area under the curve (AUC), F 1 -score, accuracy, sensitivity, and specificity, with additional statistical testing using Kruskal-Wallis, Mann-Whitney U , bootstrapping, and permutation analysis. Results:Inter-rater agreement for radiological response categories was high (Cohen's kappa = 0.91). Post-treatment radiomic features yielded the best performance. The random forest model achieved an AUC of 0.94, an F 1 -score of 0.79, an accuracy of 0.79, a sensitivity of 0.80, and a specificity of 0.83. Clinical features did not meaningfully improve performance. Models based on the largest lesion showed lower accuracy than those using the overall response. Conclusions:Post-treatment CT radiomics captured therapy-induced skeletal changes and enabled differentiation of immunotherapy response categories in NSCLC bone metastases. These findings highlight radiomics as a non-invasive tool for response assessment and guiding personalized treatment strategies.
Purpose:Rigid head motion during interventional C-arm cone-beam CT (CBCT) is a major source of image degradation. Learning-based motion estimation requires realistic training data, but ground-truth motion is scarce, limiting direct validation of compensation trajectories. We address this gap with an open resource consisting of tracked real motion and pregenerated synthetic motion, along with a pretrained variational autoencoder (VAE) to generate larger ground-truth datasets. Approach:Using stereo optical tracking, we recorded rigid 6-DoF head motion trajectories from 25 volunteers lying head-first supine on an examination table, resembling a clinical setting. After data preprocessing, we trained a VAE on 120 sequences of 10 s at 30 Hz. Motion is represented in patient-centered coordinates to support transformation to arbitrary scan geometries. Similarity between measured and generated data is assessed via distributional distances, correlation metrics, low-dimensional embeddings, and a posthoc analysis of the learned latent space. Results:Evaluated based on 120 generated sequences, the trained VAE is capable of producing diverse 6-DoF trajectories that preserve real-world data correlation structure. Distributional and frequency-domain metrics, along with t-SNE embeddings, show overlap between real and synthetic samples without evidence of mode collapse or training data replication. Conclusions:This work provides an openly released resource comprising measured trajectories, a synthetic dataset, and pretrained VAE weights together with full training and evaluation code, combining rigid 6-DoF head motion measured in a realistic C-arm setting with a retrainable generative model. It is intended to support reproducible development, benchmarking, and comparison of head motion estimation methods in medical imaging modalities.
Purpose:Vision-language models (VLMs) are increasingly proposed for radiologic decision support, yet the security implications of deploying optical character recognition (OCR)-capable models in diagnostic workflows remain poorly characterized. When image-embedded text is not treated as untrusted input, the visual channel becomes vulnerable to adversarial manipulation. Approach:Ten VLMs, none validated for clinical diagnosis, were evaluated on 600 brain magnetic resonance imaging studies for binary tumor detection under 5 conditions: clean input, visible radiology report injection, human-imperceptible stealth OCR injection, and multi-stage immune-prompt defense. A total of 27,000 inference calls were analyzed. Results:At baseline, performance was heterogeneous, with a median accuracy of 0.69, a sensitivity of 0.79, and a specificity of 0.59. Visible injection caused universal specificity collapse to 0.00 across all models with a false-positive rate (FPR) of 1.00 and a median attack success rate (ASR) of 0.97. Stealth injection, despite being imperceptible to human reviewers, drove substantial degradation with a median accuracy of 0.43, an ASR of 0.57, and an FPR of 0.84. Immune prompting achieved only partial mitigation: under stealth injection, median ASR decreased to 0.44, and accuracy improved to 0.56, yet residual overcalling persisted with a median FPR of 0.67, and three models maintained an FPR of 1.00. Conclusions:Commercial VLMs exhibit a deployment-critical failure mode in radiology-like scenarios: OCR-readable text embedded in images can override pixel-level evidence, even under stealth conditions that evade human inspection. Prompt-level defenses provide insufficient protection. Any clinical integration of VLMs must be governed by system-level safeguards, including OCR-aware input handling, provenance controls, and enforced human verification, before deployment in safety-sensitive environments.
Purpose:We aim to investigate whether a breast cancer risk model can be trained with transfer learning from a breast cancer detection model. Approach:An existing open-source breast cancer detection model was used to extract the latent space representation of a local dataset composed of images from 17,878 breast cancer screening participants. An autoencoder network was used for dimensionality reduction. A final model training step mapped the reduced representation to 4-year breast cancer risk. Two experimental studies were performed: first, a local dataset was used to investigate the impact of transfer learning to local data. Second, the training was executed for breast cancer risk by excluding cases with obvious signs of cancer in the model training process, to avoid the influence of cancer signs. In addition, the Mirai model was evaluated. Results:The open-source model achieved a baseline performance of 0.72 with 95%CI [0.66, 0.78] for 4-year risk prediction when applied to local test cases of breast cancer risk data. This performance was improved after training on local breast cancer detection data to 0.75 with 95%CI [0.69, 0.81], although not statistically significant ( p = 0.09 ). Further training with local breast cancer risk data achieved an area under the curve (AUC) of 0.77 with 95%CI [0.71, 0.82], which is higher than the original model ( p < 0.01 ). The Mirai model demonstrated a lower performance after fine-tuning ( p < 0.01 ). Conclusions:Transfer learning enabled the creation of a breast cancer risk prediction model from a previously developed breast cancer detection model, demonstrating that a publicly available model can be adapted to local clinical needs.
This editorial revisits the enduring role of literature reviews in scientific training and communication, emphasizing that their value lies in expert synthesis rather than simple aggregation of sources. In an era of AI-assisted discovery and summarization, the authors argue that peer-reviewed reviews must provide deeper perspective on how fields evolve and how evidence interconnects. They outline updated expectations for review and perspective articles, stressing that human judgment remains essential for interpreting and validating the scientific record.
Purpose:We present RadGazeGen, a framework for integrating experts' eye gaze patterns and radiomic feature maps as controls within text-to-image diffusion models to enable high-fidelity medical image generation. Although recent text-to-image diffusion models have achieved impressive success, textual descriptions alone often fail to capture disease-specific details necessary for generating clinically accurate and anatomically faithful images. Approach:To address these limitations, RadGazeGen leverages radiologists' eye gaze trajectories and radiomics feature descriptors as spatial and semantic controls in the diffusion process. Eye gaze patterns encode visuo-cognitive attention and spatial localization of subtle disease cues, whereas radiomics features capture subvisual phenotype characteristics such as texture, intensity, and shape. By combining these multimodal cues, the proposed framework guides the generative model toward anatomically consistent and disease-aware image synthesis. The quality of the generated images were evaluated using a board-certified radiologist. Results:RadGazeGen was evaluated on the REFLACX dataset for image generation quality and diversity. Furthermore, to assess its downstream clinical utility, the generated images were used for disease classification tasks on the CheXpert test set ( n = 500 ) and for long-tailed learning evaluation on the MIMIC-CXR-LT test set ( n = 23,550 ), demonstrating high fidelity and diagnostic relevance of the synthesized images. Conclusions:By jointly conditioning on gaze and radiomic representations, RadGazeGen bridges the gap between human visual cognition and machine perception, improving both realism and clinical validity in medical image generation. This framework underscores the importance of incorporating anatomically grounded and disease-aware controls in diffusion-based medical image synthesis.
Purpose:We aim to overcome the remaining bottlenecks in COVID-19 lesion segmentation from chest CT-namely, blurred lesion boundaries, false-positive responses from vessels or trachea, and the extreme variability of lesion shape, size, and location-by developing a dual-decoder adversarial residual U 2 Net algorithm. Approach:DDARes- U 2 Net , an encoder-dual-decoder architecture, simultaneously performs both lung parenchyma segmentation and lesion segmentation within a unified network. The dual-decoder design incorporates a lung-parenchyma decoder that guides the model to focus specifically on the lung region, eliminating the information loss caused by traditional presegmentation approaches and markedly reducing computational overhead. Furthermore, an edge-loss function is introduced during training to enhance the network's ability to capture fine-grained lesion boundaries. Comparative experiments are conducted under two settings: within-dataset evaluation and cross-dataset evaluation. Ablation studies are performed to validate the contribution of each architectural component. Results:The proposed DDARes- U 2 Net model, with a relatively modest parameter count of 30.74 M, achieves superior generalization performance and segmentation accuracy compared with state-of-the-art methods. Across both within-center and cross-center experiments on five CT datasets (I to V), DDARes- U 2 Net consistently ranks first for Dice and IoU. The largest gain is observed when the model trained on Dataset IV is tested on Dataset II, delivering + 3.41 pp Dice versus ARes- U 2 Net + + and + 5.51 pp versus ARes- U 2 Net , confirming its superior generalizability and clinical applicability. Conclusions:The unified dual-decoder framework with explicit edge-aware supervision provides accurate, robust, and generalizable COVID-19 lesion segmentation, achieving the highest Dice and IoU in both intradataset and cross-center tests on five independent CT cohorts.
Purpose:Automated whole-slide image (WSI) analysis, specifically applications of deep learning (DL)-based algorithms, has been enabling automated detection, classification, segmentation, and prognosis for various diseases. Performance evaluation plays an important role in the success of these complex big-data-based technologies. Our purpose is to conduct a performance evaluation of DL segmentation models applied to a breast cancer WSI dataset provided by the Tumor InfiltratinG lymphocytes in breast cancER challenge and investigate methodological issues in the assessment of WSI segmentation models. Approach:We evaluated the performance of DL models in the segmentation of tumoral and stromal regions and the effect of color normalization on improving the performance of these models when the training and testing data are from different sources. One important issue is the aggregation of image segmentation performance when the reference standard includes annotations only from selected regions of interest (ROIs). We introduced three different methods for aggregating performance based on different units of analysis (pixels, ROIs, and slides) and a bootstrap method to estimate the variance of the performance results at the slide level. Results:We found that using different units of analysis can produce not just different mean performance estimates but also different levels of uncertainty. Our results also showed that color normalization significantly improved DL model performance when the training and testing data are from different sources. Conclusions:Our study demonstrates the importance of image acquisition, study design, and statistical analysis methods used in the performance evaluation of computational pathology applications.
Purpose: Spatial transcriptomics (ST) enables gene expression measurements within the tissue context. However, these measurements are often noisy, low-resolution, and sparsely sampled, which limits the recovery of fine spatial structure. Deep neural networks have become powerful tools for expression imputation from histology, but their performance remains constrained by limited sample sizes and a lack of biologically informed augmentation. Most of the existing augmentation strategies for learning are designed for classification tasks rather than regression, which neglect spatial and transcriptomic relationships, leading to biologically implausible interpolations that hinder prediction performance. Approach: To address these limitations, we propose SNR-ST-Mix, a geometry- and expression-aware data augmentation framework designed specifically for ST data. It constrains mixing to a spot's k-nearest spatial neighbors and adaptively weights interpolation coefficients based on expression similarity, generating augmented samples that preserve local biological structure while ensuring spatial smoothness. This dual conditioning yields synthetic examples that expand the effective training manifold, promote generalization, and enhance prediction stability under sample-specific training. Results: Extensive experiments with various tissue types demonstrate that SNR-ST-Mix consistently outperforms conventional augmentation methods without requiring architectural changes or additional computation. Conclusions: SNR-ST-Mix provides an effective and biologically principled augmentation strategy for spatial transcriptomics regression tasks. By explicitly leveraging spatial geometry and transcriptomic similarity, it expands the effective training manifold and improves predictive performance without increasing model complexity.
Purpose:Computed tomography (CT) scout images are used in CT planning to set the anatomic scan range and optimize the radiation dose. Manual exam planning is highly variable and contributes to excess radiation dose delivered to patients. We present a multi-organ segmentation system for CT scout images with a direct application for automatic scan range prediction, an important component of CT planning. Approach:Variants of a U-Net model with a ResNet-50 encoder were trained to segment 35 distinct anatomic structures from frontal and lateral scout images and evaluated on their performance. Organ boundaries generated by the trained models were used to automatically set the scan range for chest-abdomen-pelvis and abdomen-pelvis CT protocols from frontal and lateral scout images. Results:The best-performing models produced organ segmentations with mean Dice scores of 0.823 ± 0.014 for the frontal scout and 0.784 ± 0.013 for the lateral scout. Predicted scan ranges covered ground-truth anatomy for 87% to 100% of test cases, reduced variability in six out of eight tested scan limits, and produced shorter median scan lengths for three out of four scout-to-protocol prediction tasks. Conclusions:The presented models effectively segment a wide variety of organs from frontal and lateral scout images and demonstrate promising clinical applications to improved scan range prediction and automatic CT planning.
Purpose:Public datasets for training artificial intelligence (AI) models in breast cancer screening are limited in size and quality, making it difficult to develop reliable systems. We introduce OMAMA-DB, an extensive publicly available collection of two-dimensional (2D) mammograms and three-dimensional (3D) tomosynthesis volumes. Approach:Starting from 967,991 images, we created a curated set of 231,080 images using a multi-stage filtering process that removes missing labels, uncommon dimensions, rare scanner types, duplicate studies, and invalid DICOM files. All 2D images then undergo additional outlier detection using histogram filtering and a variational autoencoder to remove low-quality outliers. OMAMA-DB includes pathology-based cancer labels and automated lesion annotations generated using DeepSight. We also provide a web-based annotation tool for expert validation. To demonstrate usability, we fine-tuned MedGemma on a balanced subset of OMAMA-DB. We conducted a preliminary user study comparing human and automated classification of real and synthetic mammograms. Results:OMAMA-DB contains 231,080 images, including 7351 2D and 374 3D cancer cases. Fine-tuned MedGemma achieved 0.989 accuracy, 0.997 sensitivity, and an F 1 score of 0.989 on a balanced validation set of 2942 images. In real-versus-synthetic classification, humans achieved 0.485 accuracy, and logistic regression and convolutional neural network achieved 0.972 and 0.997, respectively. Conclusions:OMAMA-DB provides a large mammography dataset with pathology-based labels and automated lesion annotations to support medical imaging research. Fine-tuned foundation models demonstrate strong cancer classification performance, and the gap between human and automated detection of synthetic images highlights the importance of real clinical data. All data, models, and parameters are openly available for research use.
