Journal of Imaging最新文献

Mammographic sensitivity is reduced in women with dense breasts, leading to missed cancers and a higher burden of interval cancers. Automated breast ultrasound (ABUS) and ultrasound tomography (UST) have been introduced as supplemental breast imaging modalities, but primary studies are heterogeneous, and previous reviews have not focused on screening settings or on head-to-head comparisons with handheld ultrasound (HHUS). We systematically searched PubMed, Embase, Web of Science and the Cochrane Library for studies from 1 January 2000 to 31 May 2025 evaluating ABUS or UST as adjuncts to mammographic screening. Two reviewers independently selected studies and assessed risk of bias. When at least two clinically comparable studies were available, we pooled sensitivity and specificity using random-effects bivariate meta-analysis. Eighteen studies (just over 20,000 screening or recall episodes) met the inclusion criteria; 16 evaluated ABUS/ABVS and 2 UST. Adding ABUS to mammography increased sensitivity by 6-35 percentage points and improved cancer detection by 2.4-4.3 per 1000 women with dense breasts, with higher recall rates and modest reductions in specificity. When ABUS was compared directly with HHUS, pooled sensitivity was 0.90 and specificity 0.89, with HHUS showing slightly lower sensitivity and slightly higher specificity. Only two studies had an overall low risk of bias, and heterogeneity (particularly for specificity) was substantial. ABUS is a practical and scalable adjunct to mammography that increases cancer detection in women with dense breasts, with an expected trade-off of higher recall and modest specificity loss. Its comparative diagnostic accuracy appears broadly non-inferior to HHUS. However, the predominance of high-risk-of-bias studies and between-study heterogeneity means that high-quality population-based trials and standardised reporting are still required before widespread implementation in organised screening programmes.

The use of Multispectral (MS) imaging is growing fast across many research fields. However, one of the obstacles researchers face is the limited availability of multispectral image databases. This arises from two factors: multispectral cameras are a relatively recent technology, and they are not widely available. Hence, the development of an image database is crucial for research on multispectral images. This study takes advantage of two high-end MS cameras in visible and near-infrared based on filter array technology developed in the PImRob platform, the University of Burgundy, to provide a freely accessible database. The database includes high-resolution MS images taken from different plants and weeds, along with annotated images and masks. The original raw images and the demosaicked images have been provided. The database has been developed for research on demosaicking techniques, segmentation algorithms, or deep learning for crop/weed discrimination.

In precision agriculture, accurate, non-destructive estimation of fruit volume is crucial for quality grading, yield prediction, and post-harvest management. While vision-based methods provided some usefulness, fruits with complex geometry-such as mangosteen (Garcinia mangostana L.)-are difficult due to their large calyx, which may lead to difficulties in solving using traditional form-modeling methods. Traditional geometric solutions such as ellipsoid approximations, diameter-height estimation, and shape-from-silhouette reconstruction often fail because the irregular calyx generates asymmetric protrusions that violate their basic form assumptions. We offer a novel study framework employing both multi-view instance segmentation and hybrid geometrical feature modeling to quantitatively model mangosteen volume with traditional 2D imaging. A You Only Look Once (YOLO)-based segmentation model was employed to explicitly separate the fruit body from the calyx. Calyx inclusion resulted in dense geometric noise and reduced model performance (R2<0.40). We trained eight regression models on a curated and augmented 900 image dataset (N=720, test N=180). The models used single-view and multi-view geometric regressors (V∝A1.5), polynomial hybrid configurations, ellipsoid-based approximations, as well as hybrid feature formulations. Multi-view models consistently outperformed single-view models, and the average predictive accuracy improved from R2=0.6493 to R2=0.7290. The best model is indeed a hybrid linear regression model with side- and bottom-area features-(As1.5, Ab1.5)-combined with ellipsoid-derived volume estimation-(Vellipsoid)-which resulted in R2=0.7290, a Mean Absolute Percentage Error (MAPE) of 16.04%, and a Root Mean Square Error (RMSE) of 31.9 cm3 on the test set. These results confirm the proposed model as a low-cost, interpretable, and flexible model for real-time fruit volume estimation, ready for incorporation into automated sorting and grading systems integrated in post-harvest processing pipelines.

The development of foundation models for brain MRI depends critically on the scale, diversity, and consistency of available data, yet systematic assessments of these factors remain scarce. In this study, we analyze 54 publicly accessible brain MRI datasets encompassing over 538,031 scans to provide a structured, multi-level overview tailored to foundation model development. At the dataset level, we characterize modality composition, disease coverage, and dataset scale, revealing strong imbalances between large healthy cohorts and smaller clinical populations. At the image level, we quantify voxel spacing, orientation, and intensity distributions across 14 representative datasets, demonstrating substantial heterogeneity that can influence representation learning. We then perform a quantitative evaluation of preprocessing variability, examining how intensity normalization, bias field correction, skull stripping, spatial registration, and interpolation alter voxel statistics and geometry. While these steps improve within-dataset consistency, residual differences persist between datasets. Finally, a feature-space case study using a 3D DenseNet121 shows measurable residual covariate shift after standardized preprocessing, confirming that harmonization alone cannot eliminate inter-dataset bias. Together, these analyses provide a unified characterization of variability in public brain MRI resources and emphasize the need for preprocessing-aware and domain-adaptive strategies in the design of generalizable brain MRI foundation models.

Existing salient object detection methods for optical remote sensing images still face certain limitations due to complex background variations, significant scale discrepancies among targets, severe background interference, and diverse topological structures. On the one hand, the feature transmission process often neglects the constraints and complementary effects of high-level features on low-level features, leading to insufficient feature interaction and weakened model representation. On the other hand, decoder architectures generally rely on simple cascaded structures, which fail to adequately exploit and utilize contextual information. To address these challenges, this study proposes a Hierarchical Semantic Interaction Module to enhance salient object detection performance in optical remote sensing scenarios. The module introduces foreground content modeling and a hierarchical semantic interaction mechanism within a multi-scale feature space, reinforcing the synergy and complementarity among features at different levels. This effectively highlights multi-scale and multi-type salient regions in complex backgrounds. Extensive experiments on multiple optical remote sensing datasets demonstrate the effectiveness of the proposed method. Specifically, on the EORSSD dataset, our full model integrating both CA and PA modules improves the max F-measure from 0.8826 to 0.9100 (↑2.74%), increases maxE from 0.9603 to 0.9727 (↑1.24%), and enhances the S-measure from 0.9026 to 0.9295 (↑2.69%) compared with the baseline. These results clearly demonstrate the effectiveness of the proposed modules and verify the robustness and strong generalization capability of our method in complex remote sensing scenarios.

Precise 3D shape correspondence is a fundamental prerequisite for critical applications ranging from medical anatomical modeling to visual recognition. However, non-isometric 3D shape matching remains a challenging task due to the limited sensitivity of traditional Laplace-Beltrami (LB) bases to local geometric deformations such as stretching and bending. To address these limitations, this paper proposes a Sinkhorn-Regularized Elastic Functional Map framework (SRE-FMaps) that integrates entropy-regularized optimal transport with an elastic thin-shell energy basis. First, a sparse Sinkhorn transport plan is adopted to initialize a bijective correspondence with linear computational complexity. Then, a non-orthogonal elastic basis, derived from the Hessian of thin-shell deformation energy, is introduced to enhance high-frequency feature perception. Finally, correspondence stability is quantified through a cosine-based elastic distance metric, enabling retrieval and classification. Experiments on the SHREC2015, McGill, and Face datasets demonstrate that SRE-FMaps reduces the correspondence error by a maximum of 32% and achieves an average of 92.3% classification accuracy (with a peak of 94.74% on the Face dataset). Moreover, the framework exhibits superior robustness, yielding a recall of up to 91.67% and an F1-score of 0.94, effectively handling bending, stretching, and folding deformations compared with conventional LB-based functional map pipelines. The proposed framework provides a scalable solution for non-isometric shape correspondence in medical modeling, 3D reconstruction, and visual recognition.

This paper introduces a novel framework for sensory representation of brain imaging data, combining deep learning-based segmentation with multimodal visual and auditory outputs. Structural magnetic resonance imaging (MRI) predictions are converted into color-coded maps and stereophonic/MIDI sonifications, enabling intuitive interpretation of cortical activation patterns. High-precision U-Net models efficiently generate these outputs, supporting clinical decision-making, cognitive research, and creative applications. Spatial, intensity, and anomalous features are encoded into perceivable visual and auditory cues, facilitating early detection and introducing the concept of "auditory biomarkers" for potential pathological identification. Despite current limitations, including dataset size, absence of clinical validation, and heuristic-based sonification, the pipeline demonstrates technical feasibility and robustness. Future work will focus on clinical user studies, the application of functional MRI (fMRI) time-series for dynamic sonification, and the integration of real-time emotional feedback in cinematic contexts. This multisensory approach offers a promising avenue for enhancing the interpretability of complex neuroimaging data across medical, research, and artistic domains.

COVID-19 screening is crucial for the early diagnosis and treatment of the disease, with lung ultrasound posing as a cost-effective alternative to other imaging techniques. Given the dependency on medical expertise and experience to accurately identify patterns in ultrasound exams, deep learning techniques have been explored for automatically classifying patients' conditions. However, the limited availability of public medical databases remains a significant obstacle to the development of more advanced models. To address the data scarcity problem, this study proposes a method that leverages generative adversarial networks (GANs) to generate synthetic lung ultrasound images, which are subsequently used to train frame-based classification models. Two types of GANs are considered: Wasserstein GANs (WGAN) and Pix2Pix. Specific tools are used to show that the synthetic data produced present a distribution close to the original data. The classification models trained with synthetic data achieved a peak accuracy of 96.32% ± 4.17%, significantly outperforming the maximum accuracy of 82.69% ± 10.42% obtained when training only with the original data. Furthermore, the best results are comparable to, and in some cases surpass, those reported in recent related studies.

Retroperitoneal sarcomas (RPS) are rare tumours, primarily treated with surgical resection. However, recurrences are frequent. Combining clinical factors with CT-derived radiomic features could enhance treatment stratification and personalization. This study aims to assess whether radiomic features provide additional prognostic value beyond clinicopathological features in patients with high-risk RPS treated with preoperative radiotherapy. This retrospective study included patients aged 18 or older with non-recurrent and non-metastatic RPS treated with preoperative radiotherapy between 2008 and 2016. Hazard ratios (HR) were calculated using Cox proportional hazards regression to assess the impact of clinical and radiomic features on time to event outcomes. Predictive accuracy was assessed with c-statistics. Radiomic analysis was performed on the high-risk group (undifferentiated pleomorphic sarcoma, well-differentiated/de-differentiated liposarcoma or grade 2/3 leiomyosarcoma). Seventy-two patients were included, with a median follow-up of 3.7 years, the 5-year overall survival (OS) was 67%. Multivariable analysis showed older age (HR: 1.3 per 5-year increase, p = 0.04), grade 3 (HR: 180.3, p = 0.02), and larger tumours (HR: 4.0 per 10 cm increase, p = 0.02) predicted worse OS. In the higher-risk group, the c-statistic for the clinical model was 0.59 (time to distant metastasis (TDM)) and 0.56 (OS). Among 27 radiomic features, kurtosis improved OS prediction (c-statistic 0.69, p = 0.013), and Neighbourhood Gray-Tone Difference Matrix (NGTDM) busyness improved it to 0.73 (p = 0.036). Kurtosis also improved TDM prediction (c-statistic 0.72, p = 0.023). Radiomic features may complement clinicopathological factors in predicting overall survival and time to distant metastasis in high-risk retroperitoneal sarcoma. These exploratory findings warrant validation in larger, multi-institutional studies.

Lung and colon cancers remain among the leading causes of cancer-related mortality worldwide, underscoring the need for rapid and accurate histopathological diagnosis. Manual examination of biopsy slides is often time-consuming and prone to inter-observer variability, which highlights the importance of developing reliable and explainable automated diagnostic systems. This study presents DPCSE-Net, a lightweight dual-path convolutional neural network enhanced with a squeeze-and-excitation (SE) attention mechanism for lung and colon cancer classification. The dual-path structure captures both fine-grained cellular textures and global contextual information through multiscale feature extraction, while the SE attention module adaptively recalibrates channel responses to emphasize discriminative features. To enhance transparency and interpretability, Gradient-weighted Class Activation Mapping (Grad-CAM), attention heatmaps, and Integrated Gradients are employed to visualize class-specific activation patterns and verify that the model's focus aligns with diagnostically relevant tissue regions. Evaluated on the publicly available LC25000 dataset, DPCSE-Net achieved state-of-the-art performance with 99.88% accuracy and F1-score, while maintaining low computational complexity. Ablation experiments confirmed the contribution of the dual-path design and SE module, and qualitative analyses demonstrated the model's strong interpretability. These results establish DPCSE-Net as an accurate, efficient, and explainable framework for computer-aided histopathological diagnosis, supporting the broader goals of explainable AI in computer vision.