Breast magnetic resonance imaging (MRI) offers three-dimensional anatomical information for reconstruction-oriented analysis, but annotated public datasets targeting implant-based reconstruction are limited. ImageBR contains MRI volumes from 44 patients (26 bilateral and 18 unilateral reconstruction cases), yielding 70 breast-level samples with expert annotations of planned implant regions, mammary gland tissue, and breast regions. The planned implant region is a surgeon-defined virtual surgical target rather than a directly visible anatomical structure on MRI, and its reference annotations were established without using the final prosthesis specifications. A two-stage benchmark combines breast-region extraction with an nnU-Net 3D cascade incorporating image-derived anatomical priors and weighted loss optimization. Under patient-level three-fold cross-validation, the reported mean breast-level Dice scores were 83.0% and 74.4% for planned implant region and mammary gland segmentation, respectively. The retrospective comparison of predicted-region volume and diameter with final implanted prosthesis specifications yielded relative errors of 14.27% and 8.98%, respectively, and measures consistency rather than independent implant-selection accuracy.
Fire detection in low-resolution surveillance videos at night is challenging under poor illumination. Conventional methods rely on colour information or computationally intensive deep-learning models, potentially limiting their suitability for grayscale video and resource-constrained environments. This study presents a night-time fire detection method that operates on grayscale video frames and jointly models temporal brightness fluctuations and spatial motion cues using a motion-based temporal representation and a brightness-gradient support vector machine (SVM) classifier. A sliding temporal window aggregates frame-level evidence, thereby reducing isolated false detections and providing more stable fire-event decisions over time. Experiments on the UCF-Crime arson and normal subsets (30 FPS) demonstrate promising video-level fire detection, with all evaluated fire-containing videos correctly identified. The method also shows promising classification performance on normal videos. At the frame level, the mean per-video false alarm ratio and miss rate across fire-containing videos are 4.37 ± 13.69% and 14.00 ± 11.24%, respectively. Ablation results for fire-containing videos show that the full system achieves a pooled frame-level miss rate of 9.51% and an FP-to-ground-truth fire-frame ratio of 0.76%. The CPU-only system processes a single video stream at 32–34 FPS while updating all 12 visualisation panels, demonstrating its potential for real-time surveillance.
This study aims to improve quantitative precipitation estimation at high spatial and temporal resolutions using Meteosat Second Generation (MSG) satellite imagery. In this context, two complementary models are exploited for the classification of precipitation intensities into five classes, namely: No rain, Light rain, Moderate rain, Heavy rain and Very heavy rain. To better exploit the information, XGBoost captures the radiometric response at the individual pixel level, while U-Net extracts the texture and spatial context of the cloud neighborhood. The classification architecture merges the probabilistic responses derived from spectral signatures via XGBoost with the extraction of cloud spatio-textural features from the deep multiclass U-Net network. Their output probabilities are combined and recalibrated using a logistic regression meta-classifier into a unified probability vector. To obtain instantaneous estimates, the final value corresponds to the sum of the previously defined precipitation intensities, each multiplied by the probability predicted by the meta-classifier. The models are trained on ground radar observations and validated against temporally independent radar test events from Sétif. Instantaneous precipitation estimates, as well as accumulations, were calculated. The proposed hybrid model significantly outperforms state-of-the-art benchmark architectures (notably, MCM, Ada/Bag-ANN, and DST-ANN/SVM/RF). At the instantaneous scale, the model achieves remarkable accuracy, with a correlation coefficient CC = 0.983, a reduction of RMSE to 0.40 mm/h (representing a decrease of over 45% compared to U-Net alone), and a near-zero bias (BIAS = −0.02 mm/h). During longer-scale temporal integrations, the model maintains strong volumetric consistency and minimal cumulative drift across temporal scales, reaching a CC = 0.993 and an RMSE = 1.38 mm/24 h at the daily scale, as well as a CC = 0.997, an RMSE = 2.76 mm/week, and a BIAS = 0.02 mm/week at the weekly scale. These results confirm that combining spatio-textural and point-wise spectral probabilities provides a reliable solution for high-resolution quantitative precipitation estimation, offering valuable inputs for regional hydrometeorological monitoring in complex geophysical environments. While the trained model is tailored to Mediterranean winter conditions and requires targeted retraining for application to other climatic regions or summer convective regimes, these results confirm that combining spatio-textural and point-wise spectral probabilities provides a highly reliable solution for high-resolution quantitative precipitation estimation.
Corneal confocal microscopy (CCM), also referred to as in vivo confocal microscopy, is a non-invasive ophthalmic imaging technique that allows real-time visualization of the sub-basal nerve plexus and epithelial dendritic cells in the living cornea. Initially used for the diagnosis of anterior segment disorders, it has been increasingly applied to the evaluation of small-fiber involvement in systemic conditions, particularly diabetic peripheral neuropathy, in which CCM-derived parameters correlate with intraepidermal nerve fiber density (IENFD) on skin biopsy. This narrative review summarizes the technical principles of CCM, image-analysis approaches from manual tracing to deep learning, and the clinical evidence in peripheral neuropathies, central neurodegenerative diseases, autoimmune and inflammatory conditions, and long COVID. Outside diabetes, however, most data come from small, single-center, cross-sectional studies. Four barriers currently prevent CCM from becoming a routine clinical biomarker: (i) the absence of validated diagnostic thresholds and age-adjusted normative ranges outside diabetic neuropathy; (ii) the lack of disease specificity, as similar nerve loss occurs in many neurological, metabolic and ocular surface conditions; (iii) limited comparability across devices, acquisition protocols, image selection and analysis software, with automated and deep-learning tools still lacking multicenter external validation; and (iv) cost, contact-based acquisition and operator training. Overcoming these barriers will require standardized acquisition and reporting, open benchmark datasets, externally validated analysis pipelines, and large prospective longitudinal cohorts.
Zero-shot image style transfer has recently attracted increasing attention, as it enables style manipulation without additional model optimization or parameter tuning. However, most existing approaches still rely on latent inversion, where introduced reconstruction errors may accumulate and gradually degrade content structures. Moreover, current style transfer methods still struggle to balance style strength and content preservation due to inadequate guidance. To address these issues, this paper proposes a novel zero-shot image style transfer method based on conditional inversion-free flow. Built upon pre-trained Flow Matching models, our method eliminates the inversion by constructing a style-transfer trajectory ODE from a gradient-injected discrepancy between content-conditioned and style-conditioned velocity fields. To further enhance style representation, we introduce a conditional guidance mechanism driven by a pre-trained CLIP model. Specifically, a style loss is introduced to align the Gram matrix statistics of CLIP features with those of the reference style; a semantic consistency constraint based on CLIP embedding cosine distance preserves high-level semantics. The dual guidance terms are incorporated during the inversion-free editing stage through loss gradients, guiding the generation trajectory without model fine-tuning or access to internal model features. Extensive quantitative and qualitative experiments demonstrate that our method outperforms state-of-the-art approaches, providing a better balance between style consistency and content preservation.
