Masked AutoEncoder (MAE) has demonstrated significant potential in medical image analysis by reducing the cost of manual annotations. However, MAE and its recent variants are not well-developed for ultrasound images in breast cancer diagnosis, as they struggle to generalize to the task of distinguishing ultrasound breast tumors of varying sizes. This limitation hinders the model's ability to adapt to the diverse morphological characteristics of breast tumors. In this paper, we propose a novel Breast UltraSound Multi-scale Masked AutoEncoder (BUS-M2AE) model to address the limitations of the general MAE. BUS-M2AE incorporates multi-scale masking methods at both the token level during the image patching stage and the feature level during the feature learning stage. These two multi-scale masking methods enable flexible strategies to match the explicit masked patches and the implicit features with varying tumor scales. By introducing these multi-scale masking methods in the image patching and feature learning phases, BUS-M2AE allows the pre-trained vision transformer to adaptively perceive and accurately distinguish breast tumors of different sizes, thereby improving the model's overall performance in handling diverse tumor morphologies. Comprehensive experiments demonstrate that BUS-M2AE outperforms recent MAE variants and commonly used supervised learning methods in breast cancer classification and tumor segmentation tasks.

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.

Artificial Intelligence (AI) has been shown to enhance fracture-detection-accuracy, but the most effective AI-implementation in clinical practice is less well understood. In the current study, four approaches to AI-implementation are evaluated for their impact on AI-effectiveness. Retrospective single-center study based on all consecutive, around-the-clock radiographic examinations for suspected fractures, and accompanying clinical-practice radiologist-diagnoses, between January and March 2023. These image-sets were independently analysed by a dedicated bone-fracture-detection-AI. Findings were combined with radiologist clinical-practice diagnoses to simulate the four AI-implementation methods deemed most relevant to clinical workflows: AI-standalone (radiologist-findings not consulted); AI-problem-solving (AI-findings consulted when radiologist in doubt); AI-triage (radiologist-findings consulted when AI in doubt); and AI-safety net (AI-findings consulted when radiologist diagnosis negative). Reference-standard diagnoses were established by two senior musculoskeletal-radiologists (by consensus in cases of disagreement). Radiologist- and radiologist + AI diagnoses were compared for false negatives (FN), false positives (FP) and their clinical consequences. Experience-level-subgroups radiologists-in-training-, non-musculoskeletal-radiologists, and dedicated musculoskeletal-radiologists were analysed separately. 1508 image-sets were included (1227 unique patients; 40 radiologist-readers). Radiologist results were: 2.7 % FN (40/1508), 28 with clinical consequences; 1.2 % FP (18/1508), 2 received full-fracture treatments (11.1 %). All AI-implementation methods changed overall FN and FP with statistical significance (p < 0.001): AI-standalone 1.5 % FN (23/1508; 11 consequences), 6.8 % FP (103/1508); AI-problem-solving 3.2 % FN (48/1508; 31 consequences), 0.6 % FP (9/1508); AI-triage 2.1 % FN (32/1508; 18 consequences), 1.7 % FP (26/1508); AI-safety net 0.07 % FN (1/1508; 1 consequence), 7.6 % FP (115/1508). Subgroups show similar trends, except AI-triage increased FN for all subgroups except radiologists-in-training. Implementation methods have a large impact on AI-effectiveness. These results suggest AI should not be considered for problem-solving or triage at this time; AI standalone performs better than either and may be a source of assistance where radiologists are unavailable. Best results were obtained implementing AI as safety net, which eliminates missed fractures with serious clinical consequences; even though false positives are increased, unnecessary treatments are limited.

Artificial intelligence in radiology critically depends on vast amounts of quality data, and there are controversies surrounding the topic of data ownership. In the current clinical framework, the secondary use of clinical data should be treated as a form of public good to benefit future patients. In this article, we propose that the physicians' input in data curation and interpretation adds value to the data and is crucial for building clinically relevant artificial intelligence models.

Early detection of brain tumors in MRI images is vital for improving treatment results. However, deep learning models face challenges like limited dataset diversity, class imbalance, and insufficient interpretability. Most studies rely on small, single-source datasets and do not combine different feature extraction techniques for better classification. To address these challenges, we propose a robust and explainable stacking ensemble model for multiclass brain tumor classification. To address these challenges, we propose a stacking ensemble model that combines EfficientNetB0, MobileNetV2, GoogleNet, and Multi-level CapsuleNet, using CatBoost as the meta-learner for improved feature aggregation and classification accuracy. This ensemble approach captures complex tumor characteristics while enhancing robustness and interpretability. The proposed model integrates EfficientNetB0, MobileNetV2, GoogleNet, and a Multi-level CapsuleNet within a stacking framework, utilizing CatBoost as the meta-learner to improve feature aggregation and classification accuracy. We created two large MRI datasets by merging data from four sources: BraTS, Msoud, Br35H, and SARTAJ. To tackle class imbalance, we applied Borderline-SMOTE and data augmentation. We also utilized feature extraction methods, along with PCA and Gray Wolf Optimization (GWO). Our model was validated through confidence interval analysis and statistical tests, demonstrating superior performance. Error analysis revealed misclassification trends, and we assessed computational efficiency regarding inference speed and resource usage. The proposed ensemble achieved 97.81% F1 score and 98.75% PR AUC on M1, and 98.32% F1 score with 99.34% PR AUC on M2. Moreover, the model consistently surpassed state-of-the-art CNNs, Vision Transformers, and other ensemble methods in classifying brain tumors across individual four datasets. Finally, we developed a web-based diagnostic tool that enables clinicians to interact with the proposed model and visualize decision-critical regions in MRI scans using Explainable Artificial Intelligence (XAI). This study connects high-performing AI models with real clinical applications, providing a reliable, scalable, and efficient diagnostic solution for brain tumor classification.

In this work, we developed deep neural networks for the fast and comprehensive estimation of the most salient features of aortic blood flow. These features include velocity magnitude and direction, 3D pressure, and wall shear stress. Starting from 40 subject-specific aortic geometries obtained from 4D Flow MRI, we applied statistical shape modeling to generate 1,000 synthetic aorta geometries. Complete computational fluid dynamics (CFD) simulations of these geometries were performed to obtain ground-truth values. We then trained deep neural networks for each characteristic flow feature using 900 randomly selected aorta geometries. Testing on remaining 100 geometries resulted in average errors of 3.11% for velocity and 4.48% for pressure. For wall shear stress predictions, we applied two approaches: (i) directly derived from the neural network-predicted velocity, and, (ii) predicted from a separate neural network. Both approaches yielded similar accuracy, with average error of 4.8 and 4.7% compared to complete 3D CFD results, respectively. We recommend the second approach for potential clinical use due to its significantly simplified workflow. In conclusion, this proof-of-concept analysis demonstrates the numerical robustness, rapid calculation speed (less than seconds), and good accuracy of the CFD-based machine learning approach in predicting velocity, pressure, and wall shear stress distributions in subject-specific aortic flows.

The pathological changes in deep medullary veins (DMVs) have been reported in various diseases. However, accurate modeling and quantification of DMVs remain challenging. We aim to propose and assess an automated approach for modeling and quantifying DMVs at 7 Tesla (7 T) MRI. A multi-echo-input Res-Net was developed for vascular segmentation, and a minimum path loss function was used for modeling and quantifying the geometric parameter of DMVs. Twenty-one patients diagnosed as subcortical vascular dementia (SVaD) and 20 condition matched controls were included in this study. The amplitude and phase images of gradient echo with five echoes were acquired at 7 T. Ten GRE images were manually labeled by two neurologists and compared with the results obtained by our proposed method. Independent samples t test and Pearson correlation were used for statistical analysis in our study, and p value < 0.05 was considered significant. No significant offset was found in centerlines obtained by human labeling and our algorithm (p = 0.734). The length difference between the proposed method and manual labeling was smaller than the error between different clinicians (p < 0.001). Patients with SVaD exhibited fewer DMVs (mean difference = -60.710 ± 21.810, p = 0.011) and higher curvature (mean difference = 0.12 ± 0.022, p < 0.0001), corresponding to their higher Vascular Dementia Assessment Scale-Cog (VaDAS-Cog) scores (mean difference = 4.332 ± 1.992, p = 0.036) and lower Mini-Mental State Examination (MMSE) (mean difference = -3.071 ± 1.443, p = 0.047). The MMSE scores were positively correlated with the numbers of DMVs (r = 0.437, p = 0.037) and were negatively correlated with the curvature (r = -0.426, p = 0.042). In summary, we proposed a novel framework for automated quantifying the morphologic parameters of DMVs. These characteristics of DMVs are expected to help the research and diagnosis of cerebral small vessel diseases with DMV lesions.

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 develop and evaluate a deep learning technique for the differentiation of hepatocellular carcinoma (HCC) using "simplified intravoxel incoherent motion (IVIM) parameters" derived from only 3 b-value images. Ninety-eight retrospective magnetic resonance imaging data were collected (68 men, 30 women; mean age 59 ± 14 years), including T2-weighted imaging with fat suppression, in-phase, out-of-phase, and diffusion-weighted imaging (b = 0, 100, 800 s/mm2). Ninety percent of data were used for stratified 10-fold cross-validation. After data preprocessing, diffusion-weighted imaging images were used to compute simplified IVIM and apparent diffusion coefficient (ADC) maps. A 17-layer 3D convolutional neural network (3D-CNN) was implemented, and the input channels were modified for different strategies of input images. The 3D-CNN with IVIM maps (ADC, f, and D*) demonstrated superior performance compared with other strategies, achieving an accuracy of 83.25 ± 6.24% and area under the receiver-operating characteristic curve of 92.70 ± 8.24%, significantly surpassing the baseline of 50% (P < 0.05) and outperforming other strategies in all evaluation metrics. This success underscores the effectiveness of simplified IVIM parameters in combination with a 3D-CNN architecture for enhancing HCC differentiation accuracy. Simplified IVIM parameters derived from 3 b-values, when integrated with a 3D-CNN architecture, offer a robust framework for HCC differentiation.

Preoperative identification of genetic mutations is conducive to individualized treatment and management of papillary thyroid carcinoma (PTC) patients. <i>Purpose</i>: To investigate the predictive value of the machine learning (ML)-based ultrasound (US) radiomics approaches for BRAF V600E and TERT promoter status (individually and coexistence) in PTC. This multicenter study retrospectively collected data of 1076 PTC patients underwent genetic testing detection for BRAF V600E and TERT promoter between March 2016 and December 2021. Radiomics features were extracted from routine grayscale ultrasound images, and gene status-related features were selected. Then these features were included to nine different ML models to predicting different mutations, and optimal models plus statistically significant clinical information were also conducted. The models underwent training and testing, and comparisons were performed. The Decision Tree-based US radiomics approach had superior prediction performance for the BRAF V600E mutation compared to the other eight ML models, with an area under the curve (AUC) of 0.767 versus 0.547-0.675 (p < 0.05). The US radiomics methodology employing Logistic Regression exhibited the highest accuracy in predicting TERT promoter mutations (AUC, 0.802 vs. 0.525-0.701, p < 0.001) and coexisting BRAF V600E and TERT promoter mutations (0.805 vs. 0.678-0.743, p < 0.001) within the test set. The incorporation of clinical factors enhanced predictive performances to 0.810 for BRAF V600E mutant, 0.897 for TERT promoter mutations, and 0.900 for dual mutations in PTCs. The machine learning-based US radiomics methods, integrated with clinical characteristics, demonstrated effectiveness in predicting the BRAF V600E and TERT promoter mutations in PTCs.