International Journal of Imaging Systems and Technology最新文献

Electroencephalography (EEG) signals, due to their non-invasiveness, high temporal resolution, and low cost, have demonstrated broad application prospects in fields such as Brain-Computer Interface (BCI), motor rehabilitation, and emotion recognition. However, the inherently low signal-to-noise ratio, non-stationarity, and high dimensionality of EEG signals pose substantial challenges for signal decoding. To address these issues and improve the accuracy and robustness of EEG classification, this study proposes a novel EEG classification architecture for motor imagery data—Parallel Hybrid CNN-Transformer (PHCT). The PHCT model consists of a shallow feature extraction module and parallel fine-grained and coarse-grained feature extractors. The former efficiently captures temporal and spatial local features using separable convolutions, while the latter integrates multi-head self-attention mechanisms with convolutional structures to model global and local dependencies. Additionally, data augmentation and Gaussian noise injection are introduced to enhance the model's generalization ability. Empirical studies conducted on the BCI Competition IV-2b dataset show that the proposed model outperforms existing methods across nine subjects, achieving at least a 4.2% improvement in average classification accuracy compared to the current best model, resulting in a significant model performance boost. The experimental results are visualized using the t-SNE method. This study provides a new effective pathway for EEG decoding in complex environments.

Accurate diagnosis of breast cancer is critical for improving patient outcomes. Yet breast lesions are small and mammograms are high-resolution, patch-based methods that often ignore peritumoral context, undermining diagnostic accuracy. We therefore develop an interpretable lesion-aware diagnostic model (LADM), which directly identifies tumor regions from full-field mammograms to improve classification accuracy and clinical trust. LADM incorporates a hyper lesion-aware module that combines spatial- and channel-guided attention. LADM processes craniocaudal (CC) and mediolateral oblique (MLO) mammograms via three branches—CC, MLO, and dual-view. The single-view branches independently learn diagnostic features and perform classification, while the dual-view branch concatenates CC and MLO features for late-fusion prediction to exploit cross-view complementarity. Model interpretability and lesion localization are assessed with Grad-CAM++ heatmaps. On INBreast and CBIS-DDSM, LADM consistently outperforms single-view baselines. In the dual-view setting, it achieves AUC 0.950 and accuracy 0.903 on INBreast, and AUC 0.911 and accuracy 0.848 on CBIS-DDSM, respectively. Visualizations show that the model focuses on diagnostically relevant regions, supporting clinical interpretability. LADM learns lesion-aware features directly on full-field images, fuses CC-MLO while staying single-view robust, and offers transparent predictions with lesion maps. Together these yield accurate, interpretable classification of breast cancer.

Axillary lymph node (ALN) metastasis is a critical prognostic factor in breast cancer; accurate preoperative evaluation is crucial for guiding treatment decisions. This study aims to develop and validate a combined model integrating clinical, radiomics, and deep learning (DL) features derived from preoperative ultrasound images to non-invasively predict ALN metastasis in breast cancer patients, while systematically comparing the predictive value of features derived from different peritumoral regions and multiple feature sources. A total of 431 breast cancer patients, with axillary lymph node dissection pathology serving as the gold standard, were retrospectively enrolled and randomly assigned to training (n = 301) and test (n = 130) sets. Clinical features, radiomics features, and deep learning features (extracted from the penultimate layer of a pre-trained ResNet50 convolutional neural network using global average pooling) were obtained from intratumoral regions, combined intratumoral and peritumoral regions (1–3 mm margins), and whole ultrasound images. Machine learning models were constructed separately for clinical, radiomics, DL, and combined feature sets. Models were evaluated using the area under the curve (AUC), sensitivity (SEN), specificity (SPE), positive predictive value (PPV), and negative predictive value (NPV). Clinical models demonstrated limited predictive performance (test set AUC: 0.591–0.611). Radiomics and DL models showed improved performance, particularly when including peritumoral information. A random forest-based combined model integrating clinical features, radiomics features (intratumoral+3-mm peritumoral region), and DL features yielded the highest performance, achieving a test set AUC of 0.869, with SEN/SPE/PPV/NPV values of 0.886/0.856/0.856/0.839, respectively. Additionally, the model demonstrated robust predictive accuracy across molecular subtypes. The combined model integrating clinical, radiomics, and deep learning features from intratumoral and peritumoral ultrasound images effectively and non-invasively predicts ALN metastasis in breast cancer patients. This approach shows potential for clinical decision support, may help reduce unnecessary sentinel lymph node biopsies, and shows promise for generalizability across different molecular subtypes.

This study investigates the potential of the Mistral large language model (LLM) to classify radiological reports as normal or abnormal using three techniques: Zero-Shot Learning (ZSL), Few-Shot Learning (FSL), and Fine-Tuning (FT), aiming to optimize radiology workflows and improve clinical decision-making. The dataset consisted of 124 807 radiology reports from MRI and CT scans conducted between 1 May 2024, and 1 November 2024, at our institution. After applying inclusion and exclusion criteria, 123 296 reports were selected for analysis. The Mistral LLM was tested with ZSL, FSL, and FT techniques. Quantitative metrics, including precision, recall, F1 score, and accuracy, were calculated for each technique. Confusion matrices and qualitative analyses of misclassified cases were also performed. ZSL yielded the lowest performance, with an F1 score of 0.191 for the normal class and an overall accuracy of 0.438, due to a high false-positive rate. FSL improved accuracy to 0.806 but still showed limitations in classifying normal reports (F1 = 0.404). FT achieved the best results, with F1 scores above 0.98 for both classes and an overall accuracy of 0.998, minimizing false positives and false negatives. Classifying radiological reports as normal or abnormal is crucial for prioritizing urgent cases and optimizing workflows. The Mistral LLM, particularly with Fine-Tuning, demonstrated strong potential for automating this task, outperforming ZSL and FSL.

Parallel magnetic resonance imaging (pMRI) reduces MRI acquisition time, with Sensitivity Encoding (SENSE) being a widely used method that exploits coil sensitivity maps for efficient image reconstruction. However, SENSE can introduce aliasing and noise artifacts, especially at high acceleration factors. To address this limitation, we propose a deep learning-based postprocessing framework that enhances SENSE-reconstructed images using a Convolutional Autoencoder (CAE). The CAE is applied after the SENSE reconstruction to reduce artifacts and improve image quality, without modifying the original reconstruction pipeline. A dataset of 842 fully sampled anatomical images is used to simulate 8-channel coil data, with both uniform and variable-density (VD) undersampling applied at different acceleration factors. The CAE is trained on paired inputs (SENSE-reconstructed images) and targets (fully sampled references) to learn the mapping from degraded to high-quality images. Quantitative evaluation using Peak Signal-to-Noise Ratio (PSNR) and Normalized Mean Squared Error (NMSE), along with qualitative visual assessment, shows that the proposed CAE-SENSE framework significantly improves image fidelity, particularly with variable-density undersampling. These results demonstrate the potential of deep learning as a complementary tool to enhance conventional parallel imaging methods in accelerated MRI.

Recent advancements in deep learning and the utilization of pre-trained convolutional neural network (CNN) architectures have led to enhancements in classification tasks. However, these architectures often entail millions of training parameters, posing challenges for real-world deployment. In this work, we propose an iterative Gaussian feature extractor with a custom 3-layer CNN network (IGF-CNN) coupled with a feedforward artificial neural network (ANN) classifier. The input images undergo pre-processing before being fed to the proposed IGF-CNN and then ANN classifies the input into Covid-19, non-Covid-19 and pneumonia classes. The suggested model demands considerably fewer parameters and reduces training time substantially and achieves accuracies of 99.80%, 98.78%, 99.0%, respectively, across three different benchmark datasets. We have also performed cross-dataset validation and obtained consistently good results, further demonstrating the robustness of the proposed approach. The proposed architecture is accurate and efficient and can be integrated with real-time systems.

Medical image analysis typically employs modality-specific approaches, limiting comprehensive diagnostic capabilities. We introduce MediFusionNet, a deep learning architecture unifying brain MRI and chest X-ray analysis through specialized encoding, cross-modality attention, and uncertainty-aware predictions. Our architecture preserves modality-specific features while enabling knowledge transfer between anatomically distinct regions. Experiments demonstrate 97.73% overall accuracy, significantly outperforming specialized single-modality networks and existing multimodal approaches. MediFusionNet demonstrates positive cross-modal knowledge transfer (CMKTB, +2.61%) while providing calibrated uncertainty estimates (Expected Calibration Error, ECE = 0.027). This uncertainty quantification facilitates clinically meaningful workflow optimization, automatically processing 82.5% of cases with 99.3% accuracy. Ablation studies quantify each architectural component's contribution, providing insights for robust, uncertainty-aware, multimodal medical image analysis systems.

In modern healthcare, deep learning methods have gained considerable attention in analyzing medical imaging over the last few years, achieving reliable outcomes in various segmentation tasks. However, these models heavily rely on high-quality and large annotated samples, which remain a scarce and hard resource to acquire in several medical fields, particularly in obstetrics and gynecology, limiting the deep learning models' capability to generalize effectively on unseen datasets. Therefore, this study proposes a two-stage framework that enhances the model segmentation performance by incorporating synthetic ultrasound image generation for fetal head segmentation and thus measures its circumference. In this paper, we propose a novel two-stage pipeline designed to segment the fetal head circumference. Initially, an improved Deep Convolutional Generative Adversarial Network is employed to generate synthetic fetal head ultrasound images, producing high-quality and high structural similarity to real ones. Preliminary annotations were obtained through automated segmentation using a lightweight U-Net, followed by a refined enhancement phase. These annotations were incorporated progressively into the U-Net training to enhance the model's performance and effectiveness. The proposed framework was evaluated on the HC18 Grand Challenge dataset. The GAN-based synthetic images achieved a Peak Signal-to-Noise Ratio of 56 and a Structural Similarity Index Measure of 0.99, demonstrating the diversity and similarity of the images and thus significantly improving the U-Net segmentation Dice Score with an increase of 1.62%. Compared to prior works using the same dataset, our model achieved the highest Dice Coefficient of 98%, a Jaccard Index of 96.11%, and a Hausdorff Distance of 0.329 mm on the test set. Our lightweight U-Net, combined with GAN-based data augmentation, effectively addresses the challenge of data scarcity and enhances the segmentation of the fetal head with precise delineation, providing a robust solution in clinical application for early fetal anomaly detection and prenatal diagnosis.

Breast cancer remains a major universal health issue among women, and early detection is essential for effective treatment and better survival rates. This study presents a non-invasive method for early breast cancer detection using thermal imaging and Explainable Artificial Intelligence (XAI) techniques. Thermal imaging is a radiation-free, safe and comfortable substitute for mammography. Thermal imaging influences the heat generated by cancerous tissues due to their higher metabolic activity. Deep learning methods train using the images by detecting the significant features between the classes. The best models before and after segmentation were VGG19 and ResNet, with accuracies of 95.2% and 96.8%, respectively. XAI techniques, specifically DeepSHAP and Local Interpretable Model-agnostic Explanations (LIME), were applied to improve model interpretability and confidence by detecting regions in images revealing abnormalities. The model outcome and explanations were consequently used to validate the reliability of the predictions made by YOLOv8 for finding an abnormal contour. To address dataset scalability, a bigger segmented dataset of 760 images was applied to a novel method, PenFeatNet, and it achieved 97.6% accuracy. This approach improved accuracy by 1.3% by isolating feature extraction from classification, reducing architectural complexity. These findings provide new avenues for further research and validation, potentially revolutionizing breast cancer screening.

Diabetic retinopathy (DR) is a primary reason for visual impairment and blindness in individuals with diabetes worldwide. Timely detection of DR is essential to prevent vision loss in diabetics. However, noise and limited model transparency often compromise the accuracy of diagnosing retinal fundus images. Noise and interpretability are the two main challenges occurring in imaging datasets, overshadowing concerns such as class imbalance or device variability. These distortions are present in all datasets and devices, reducing the clarity of diagnostic signals at the pixel level and often obscuring early lesions within background noise. Addressing these challenges, this research introduces an innovative model called Explainable MINet-ViT, which combines advanced noise reduction techniques with explainable deep learning for more reliable identification of DR. The model incorporates a multi-level denoising network (MINet), modified by a noise-specific pre-processing module using a Variance-Stabilizing Transform (VST) and deep residual feature mapping. A hybrid deep learning architecture that combines Convolutional Neural Networks (CNNs) with Vision Transformers (ViTs) is employed to extract both local and global spatial information. We apply explainability strategies, such as Grad-CAM and SHAP, to ensure clinical interpretability by identifying the crucial retinal regions that influence model predictions. Quantitative and qualitative results show improved performance, robustness, and clinical applicability, achieving an accuracy of 97.6%, a sensitivity of 0.96, a specificity of 0.97, a Kappa of 0.92, and an AUC of 96.7%. Analyses of standard datasets reveal that our proposed model outperforms prior models in accuracy, noise robustness, and interpretability, rendering it exceptionally suitable for real-world clinical applications.