The availability of large multi-view mammographic datasets is limited by privacy concerns and acquisition variability, restricting the development of robust artificial intelligence models. This study investigates the impact of RGB encoding strategies on diffusion-based simultaneous dual-view mammogram synthesis. A Denoising Diffusion Probabilistic Model was fine-tuned using RGB representations that encode craniocaudal and mediolateral oblique views with different third-channel formulations. Three encoding strategies were evaluated using image quality metrics, distribution-based comparisons, and anatomical consistency measurements. Results show that zero-channel and difference representations achieved comparable performance, while sum-based encoding provided lower performance. These findings highlight the importance of input representation design for generating consistent synthetic multi-view mammograms.
In this work, we investigate a new deep learning method for the reconstruction of blood velocity from 4D phase contrast MRI measurements based on variational autoencoders. The principle of this approach is to leverage lower resolution velocity measurements to train a variational autoencoder. The approach identifies optimal latent variables within the latent space and uses the generative process of the variational autoencoder to enhance reconstruction accuracy. Realistic simulations demonstrate that the proposed deep learning approach outperforms classical regularisation methods under undersampled and noisy conditions. The nonlinear encoding of the velocity field achieves better reconstruction results and compensates for missing information.
Biomechanical movement analysis supports clinical, rehabilitation, and sports applications, but marker-based motion capture is costly, complex, and restricted to specialised laboratory settings. Markerless 2D human pose estimation (HPE) provides a scalable alternative, although general-purpose models use sparse keypoint annotations that limit biomechanical relevance. We fine-tuned HRNet, ViTPose, and YOLO-Pose on the SUMediPose dataset using 17-, 34-, and 59-keypoint configurations, establishing dense 2D localisation as a foundation for 3D reconstruction and direct joint-angle estimation without anatomical model fitting or inverse kinematics. Nine model variants were evaluated using conventional positional metrics alongside joint-angle error and temporal similarity. HRNet-W48-768 and ViTPose-B-Simple were selected for full-scale evaluation. HRNet-W48-768 achieved the lowest positional error (23.24 mm), while ViTPose-B-Simple provided competitive accuracy with faster inference (34.97 mm). Angular errors were 4.93° sagittal, 3.58° frontal, and 6.14° transverse, with high temporal similarity. These findings demonstrate the feasibility of low-cost markerless biomechanical assessment using adapted 2D HPE models.
Introduction Delineating infiltrative peritumoral edema (PTE) in gliomas remains challenging due to acquisition variability and the lack of histopathological validation. We systematically assessed artificial intelligence (AI) and radiomics models for peritumoral characterization.Methods Following PRISMA, we searched five databases on July 18, 2025, for peritumoral AI/radiomics studies predicting tumor infiltration, recurrence, or molecular status. Quality was evaluated using Prediction model Risk Of Bias Assessment Tool (PROBAST).Results Nineteen studies (6,851 patients) were included. Deep learning (DL) (nnU-Net, CNNs) identified PTE infiltration with high performance, achieving >80% sensitivity for recurrence (OR = 9.97), and outperforming radiomics (AUC = 0.61–0.98). Multimodal MRI (DTI/DSC+FLAIR) showed higher accuracy (0.90–0.98) than standard sequences. Synthetic MRI enabled standardized IDH/1p/19q prediction (AUC = 0.95/0.86). Voxel-wise histopathological validation was extremely rare (1 study).Discussion AI-based peritumoral phenotyping shows strong non-invasive potential. However, prospective clinical trials are essential to validate these predictive maps before their clinical use in surgical or radiotherapy planning.
Automated segmentation of unstained live cells in bright-field microscopy remains challenging for resource-constrained laboratories due to limited annotated data, computational resources, low contrast, motion blur, and overlapping cellular structures. We present a resource-efficient CNN pipeline comparing six frozen encoder backbones within enhanced U-Net architectures incorporating attention mechanisms, instance-aware modules, composite loss functions, hard-example mining, and 14-model ensemble voting. Trained on 857 bright-field and phase-contrast images expanded to 17,997 samples through augmentation, the framework required 6.5 hours on a Google Colab T4 GPU. External validation on LIVECell achieved Dice and F1 scores of 0.89, demonstrating strong cross-modality generalisation. On challenging bright-field images, the framework outperformed CellPose-SAM and StarDist while requiring substantially lower computational resources than foundation-model-based approaches. The proposed pipeline provides a reproducible and cost-effective solution for robust live-cell segmentation under constrained computational environments. 1
