The identification of breast cancer using mammography and ultrasonography is influenced by inter-observer variability. Artificial intelligence (AI) models for breast classification have been investigated, but limited reports exist on Breast Imaging Reporting and Data System (BI-RADS)-output AI systems. This study attempted to assess a BI-RADS-output AI system by comparing breast cancer detection rates among radiologist-only, AI-only, and combined interpretations, while also focusing on the AI’s ability to identify cancers missed by radiologists and analyzing diagnostic concordance concerning imaging features and BI-RADS categories. A retrospective analysis was performed on 331 lesions from 321 consecutive breast cancer patients who underwent preoperative digital mammography and ultrasonography. McNemar’s test was used to find the detection rates, and chi-square/Fisher’s test, Receiver Operating Characteristic Curve (ROC) analysis, and Kappa statistics were used to verify the diagnostic performance. The combined radiologist-AI interpretation achieved the highest cancer detection rate of 90.6
Precise segmentation of cardiac magnetic resonance images (CMR) is essential to quantitative measurements of the heart, yet it is often affected by respiratory motion misregistration’s and this is particularly common in free-breathing patient environments. These artifacts reduce the quality of the image and present a big challenge to traditional deep learning models such as the baseline U-Net. This paper carries out a systematic evaluation of the capabilities of four alternate U-Net architectures, Attention U-Net, Dense U-Net, Residual U-Net and Sharp U-Net, on CMR data with synthetically added motion artifacts of various intensities. To add synthetic artifacts, the deformation fields were employed that mimic the realistic breathing-induced deformations and obtained the clean, mild, moderate, and severe moving conditions. Hybrid loss function that combines Dice, Focal, and Tversky losses was used to train each model to create robustness to class imbalance and boundary degradation. The segmentation quality was assessed by the following evaluation metrics: Dice Similarity Coefficient (DSC), Hausdorff Distance (HD) and Jaccard Index (JI). Their results show that Sharp U-Net had the best overall performance with the highest average DSC (0.927) and lowest HD (3.62 mm) in severe motion conditions. The region centered attention mechanisms were also capable of withstanding these in Attention U-Net. The dense and Residual U-Nets did not perform much better than the baseline, and did not perform as well when there were severe artifacts. These results show the significance of architectural improvements and motion consciousness instruction strategies in guaranteeing the dependable segmentation. The article provides a scalable and reproducible methodology to assess the performance of motion-robust deep learning models for CMR and adaptable for use in other clinical diagnostic applications where respiratory gating is not available. The proposed framework provides a reproducible and computationally practical approach to motion-robust cardiac MRI segmentation. Sharp U-Net achieved the highest segmentation performance under severe respiratory motion, with a DSC of 0.784, compared with 0.778 for Attention U-Net, while requiring a slightly higher inference latency of approximately 28–30 ms per 256 × 256 slice and GPU memory consumption of approximately 1.8 GB. This modest computational overhead represents a trade-off between improved boundary preservation and computational efficiency and should be considered when deploying the model in resource-constrained clinical environments. Across the clean, mild, moderate, and severe respiratory motion conditions, Sharp U-Net achieved an average DSC of 0.927. Under the most challenging severe-motion condition, the DSC decreased to 0.784, while maintaining an ASSD of 1.69 mm, sensitivity of 0.823, precision of 0.774, and RVE of 8.9
Smoking is associated with structural and functional alterations in the lungs, some of which may be reflected in quantitative imaging features. Radiomics enables the extraction of high-dimensional image-derived descriptors and may provide information beyond routinely available clinical variables. We investigated whether bilateral lung radiomics, both independently and integrated with clinical variables, could classify patient-level smoking status. An exploratory secondary analysis was also performed to assess whether radiomic and clinical feature representations could discriminate smoking exposure intensity among smokers. A total of 506 patients were included, comprising 253 smokers and 253 nonsmokers. Right and left lung radiomic features were extracted separately from visually normal-appearing lung parenchyma on thoracic CT images. Multiple feature representations were evaluated, including right lung radiomics, left lung radiomics, concatenated right–left radiomics, summary/asymmetry radiomics, clinical variables alone, and radiomics combined with clinical variables. Clinical variables included age, sex, height, weight, and body mass index. Pack-years were excluded from the binary smoking status classification models to avoid target-related information leakage. For each binary classification task, a balanced internal hold-out set containing 25 smokers and 25 nonsmokers was reserved, while the remaining patients were evaluated using stratified tenfold cross-validation. Four feature selection or dimensionality reduction strategies and five classifier families were assessed, resulting in 435 model configurations across the main cohort and sex-specific subgroup analyses. Threshold selection was performed using the Youden index derived from cross-validated out-of-fold predictions. For the exploratory smoking exposure analysis, 253 smokers with available pack-year information were categorized into < 10, 10–20, and > 20 pack-year groups and evaluated using pooled stratified tenfold cross-validation. The best-performing radiomics-based configurations showed higher discrimination than the clinical-only configuration. The clinical-only random forest model achieved a cross-validation ROC-AUC of 0.673 and a hold-out ROC-AUC of 0.610. Among radiomics-only configurations, concatenated bilateral radiomics achieved a cross-validation ROC-AUC of 0.734 and a hold-out ROC-AUC of 0.712. The strongest hold-out discrimination was obtained using concatenated bilateral radiomics combined with clinical variables, ReliefF feature selection, and stacking classification, achieving a hold-out ROC-AUC of 0.803, balanced accuracy of 0.700, sensitivity of 0.720, and specificity of 0.680. Feature selection stability analysis showed that the final concatenated bilateral radiomics + clinical configuration retained five clinical variables and 15 radiomic features, several of which were repeatedly selected across cross-validation folds. Sex-specific subgroup analyses showed moderate but variable performance. In the exploratory three-class smoking exposure analysis, integrating left lung radiomics with clinical variables yielded the most balanced class-wise performance, achieving the highest balanced accuracy (0.658) and the highest sensitivity for the intermediate 10–20 pack-year group (0.417), although the clinical-only model achieved the highest macro-AUC. Bilateral lung radiomics showed potential for patient-level smoking status classification, particularly when combined with clinical variables. These findings suggest that quantitative imaging features could provide complementary information beyond routinely available patient-level variables. The exploratory smoking exposure analysis further suggested that radiomics + clinical configurations may contain limited information related to smoking exposure intensity, although intermediate exposure remained challenging to discriminate. However, the moderate performance, subgroup variability, and lack of external validation indicate that this approach should be considered investigational until confirmed in larger independent cohorts.
Accurate diagnosis of muscle-invasive bladder cancer (MIBC) is critical for effective treatment planning, yet remains challenging owing to the scarcity of large-scale multicenter datasets and stringent data privacy requirements. To address these issues, we propose a federated learning (FL)-based framework for privacy-preserving MIBC diagnosis. The framework enables multi-institutional collaboration by training local models on site-specific data and aggregating them into a global model through secure parameter exchange, without transferring raw data. Furthermore, we introduce a multi-task learning architecture that simultaneously performs tumor segmentation and invasion classification, supported by a split-attention convolutional module designed to improve feature discrimination. Experimental results on MRI datasets from four medical centers show that our proposed multi‑task federated learning framework yields an AUC of 0.844, outperforming both single‑center models and the single‑task federated baseline (AUC = 0.829), and exhibits robustness under heterogeneous data distributions. Ablation experiments showed that incorporating the split-attention convolutional module increased the AUC of the multi-task FL model from 0.818 to 0.844.
Spectral geometric methods provide mathematically interpretable representations of tissue morphology, yet their standalone performance in histopathology artificial intelligence has not been systematically evaluated. This study benchmarks purely spectral pipelines across representative digital pathology and spatial omics tasks relative to convolutional baselines. Spectral features derived from graph Laplacians, heat kernel signatures, spectral wavelet energy, Slepian concentrations, fractal descriptors, and persistent homology were extracted from histopathology patches using fixed-grid graphs. Features were processed using spectral neural operators and gradient boosting. Four tasks were evaluated: binary breast cancer classification (BreaKHis), multi-class glioblastoma region-patch classification, spatial transcriptomics prediction (10× Visium), and spot-wise denoising. Cross-validation and leave-one-slide-out strategies were applied. Metrics included accuracy, sensitivity, specificity, area under the receiver operating characteristic curve (AUC), mean squared error, and Pearson correlation. Equivalence between spectral and CNN models was not established within the prespecified ±0.03 AUC margin. Purely spectral models achieved lower performance than convolutional baselines, with AUC differences of 0.12–0.23. The CNN→LGBM model achieved an accuracy of 0.83 ± 0.06, balanced accuracy of 0.73 ± 0.10, and AUC of 0.80 ± 0.13. Equivalence was not established within the prespecified margin. Fusion achieved balanced accuracy of 0.73 ± 0.08, equal to the CNN baseline mean balanced accuracy, without surpassing CNN AUC. Spatial transcriptomics prediction showed limited cross-slide generalization (Pearson correlation 0.06), whereas spot-wise denoising under sparse conditions showed competitive performance. Spectral geometric models do not outperform convolutional methods for classification or segmentation but provide complementary utility in denoising and sparse spatial settings.
