International Journal of Imaging Systems and Technology最新文献

Explainable AI (XAI) frameworks are becoming essential in many areas, including the medical field, as they help us to understand AI decisions, increasing clinical trust and improving patient care. This research presents a robust and comprehensive Explainable AI framework. To classify images from the BloodMNIST and Raabin-WBC datasets, various pre-trained convolutional neural network (CNN) architectures: the VGG, the ResNet, the DenseNet, the EfficientNet, the MobileNet variants, the SqueezeNet, and the Xception are implemented both individually and in combination with SpinalNet. For parameter analysis, four models, VGG16, VGG19, ResNet50, and ResNet101, were combined with SpinalNet. Notably, these SpinalNet hybrid models significantly reduced the model parameters while maintaining or even improving the model accuracy. For example, the VGG 16 + SpinalNet shows a 40.74% parameter reduction and accuracy of 98.92% (BloodMnist) and 98.32% (Raabin-WBC). Similarly, the combinations of VGG19, ResNet50, and ResNet101 with SpinalNet resulted in weight parameter reductions by 36.36%, 65.33%, and 52.13%, respectively, with improved accuracy for both datasets. These hybrid SpinalNet models are highly efficient and well-suited for resource-limited environments. The authors have developed a dynamic model selection framework. This framework optimally selects the best models based on prediction scores, prioritizing lightweight models in cases of ties. This method guarantees that for every input, the most effective model is used, which results in higher accuracy as well as better outcomes. Explainable AI (XAI) techniques: Local Interpretable Model-agnostic Explanations (LIME), SHapley Additive ExPlanations (SHAP), and Gradient-weighted Class Activation Mapping (Grad-CAM) are implemented. These help us to understand the key features that influence the model predictions. By combining these XAI methods with dynamic model selection, this research not only achieves excellent accuracy but also provides useful insights into the elements that influence model predictions.

Lung infections such as tuberculosis (TB), COVID-19, and pneumonia share similar symptoms, making early differentiation challenging with x-ray imaging. This can delay correct treatment and increase disease transmission. The study focuses on extracting hybrid features using multiple techniques to effectively distinguish between TB and other lung infections, proposing several methods for early detection and differentiation. To better diagnose TB, the paper presented an ensemble DenseNet with a Vision Transformer (ViT) network (EDenseNetViT). The proposed EDenseNetViT is an ensemble model of Densenet201 and a ViT network that will enhance the detection performance of TB with other lung infections such as pneumonia and COVID-19. Additionally, the EDenseNetViT is extended to predict the severity level of TB. This severity score approach is based on combined weighted low-level features and high-level features to show the severity level of TB as mild, moderate, severe, and fatal. The result evaluation was conducted using chest image datasets, that is Montgomery Dataset, Shenzhen Dataset, Chest x-ray Dataset, and COVID-19 Radiography Database. All data are merged and approx. Seven thousand images were selected for experimental design. The study tested seven baseline models for lung infection differentiation. Initially, DenseNet transfer learning models, including DenseNet121, DenseNet169, and DenseNet201, were assessed, with DenseNet201 performing the best. Subsequently, DenseNet201 was combined with Principal component analysis (PCA) and various classifiers, with the combination of PCA and random forest classifier proving the most effective. However, the EDenseNetViT model surpassed all and achieved approximately 99% accuracy in detecting TB and distinguishing it from other lung infections like pneumonia and COVID-19. The proposed EdenseNetViT model was used for classifying TB, Pneumonia, and COVID-19 and achieved an average accuracy of 99%, 98%, and 96% respectively. Compared to other existing models, EDenseNetViT outperformed the best.

Precise segmentation of pneumonia lesions using deep learning has been a research focus in medical image segmentation, in which convolutional neural networks (CNNs) excel at capturing local features through convolutional layers but struggle with global information, while Transformers handle global features and long-range dependencies well but require substantial computational resources and data. Motivated by the recently introduced Mamba that effectively models long-range dependencies with less complexity, we develop a novel network architecture in order to simultaneously enhance the handling of both global and local features. It integrates an enhanced Mamba module SE3DMamba to improve the extraction of three-dimensional global features and a medical version of deep residual convolution MDRConv to enhance the extraction of local features with a self-configuring mechanism. Experiments conducted on two pneumonia CT datasets, including the pneumonia multilesion segmentation dataset (PMLSegData) with three lesion types—consolidations, nodules, and cavities—and MosMedData of ground-glass opacifications demonstrate that our network surpasses state-of-the-art CNN and Transformer-based segmentation models across all tasks, advancing the clinical feasibility of deep learning for pneumonia multilesion segmentation.

Magnetic resonance spectroscopy (MRS) provides a non-invasive method for examining metabolic alterations associated with diseases. While ¹H-based MRS is commonly employed, its effectiveness is often limited by signal interference from water, reducing the accuracy of metabolite differentiation. In contrast, X-nuclei MRS leverages the broader chemical shift dispersion of non-hydrogen nuclei to enhance the ability to distinguish between metabolites. This article presents the design and analysis of a dual-resonant meandered coil for 7 Tesla magnetic resonance imaging (MRI), to simultaneously help in image hydrogen protons (¹H) and detect Phosphorus (³¹P) atomic nuclei at 298 MHz and 120.6 MHz, respectively. Both single-channel and four-channel configurations were designed and analyzed. The single-channel coil integrates an LC network for dual resonance, achieving excellent impedance matching (S₁₁ < −10 dB) and a homogeneous magnetic field distribution within the region of interest. A transmission-line-based matching network was implemented to optimize performance at both frequencies. The four-channel coil was simulated using CST Microwave Studio and experimentally validated. Simulations demonstrated impedance matching and minimal mutual coupling of −38 dB at 298 MHz and −24 dB at 120.6 MHz. The measured S-parameters confirmed these results, showing high decoupling and robust performance across all channels. The prototype featured integrated LC networks and optimized meander structures, ensuring efficient power transmission and uniform field distribution. This work highlights the effectiveness of the proposed dual-resonant coil designs for MRS applications, offering promising potential for advanced clinical diagnostics.

For patients with brain tumors, effectively utilizing the complementary information between multimodal medical images is crucial for accurate lesion segmentation. However, effectively utilizing the complementary features across different modalities remains a challenging task. To address these challenges, we propose a modal feature supplement network (MFSNet), which extracts modality features simultaneously using both a main and an auxiliary network. During this process, the auxiliary network supplements the modality features of the main network, enabling accurate brain tumor segmentation. We also design a modal feature enhancement module (MFEM), a cross-layer feature fusion module (CFFM), and an edge feature supplement module (EFSM). MFEM enhances the network performance by fusing the modality features from the main and auxiliary networks. CFFM supplements additional contextual information by fusing features from adjacent encoding layers at different scales, which are then passed into the corresponding decoding layers. This aids the network in preserving more details during upsampling. EFSM improves network performance by using deformable convolution to extract challenging boundary lesion features, which are then used to supplement the final output of the decoding layer. We evaluated MFSNet on the BraTS2018 and BraTS2021 datasets. The Dice scores for the whole tumor, tumor core, and enhancing tumor regions were 90.86%, 90.59%, 84.72%, and 92.28%, 92.47%, 86.07%, respectively. This validates the accuracy of MFSNet in brain tumor segmentation, demonstrating its superiority over other networks of similar type.

Medical images typically contain complex structures and abundant detail, exhibiting variations in texture, contrast, and noise across different imaging modalities. Different types of images contain both local and global features with varying expressions and importance, making accurate classification highly challenging. Convolutional neural network (CNN)-based approaches are limited by the size of the convolutional kernel, which restricts their ability to capture global contextual information effectively. In addition, while transformer-based models can compensate for the limitations of convolutional neural networks by modeling long-range dependencies, they are difficult to extract fine-grained local features from images. To address these issues, we propose a novel architecture, the Interactive CNN and Transformer for Cross Attention Fusion Network (IFC-Net). This model leverages the strengths of CNNs for efficient local feature extraction and transformers for capturing global dependencies, enabling it to preserve local features and global contextual relationships. Additionally, we introduce a cross-attention fusion module that adaptively adjusts the feature fusion strategy, facilitating efficient integration of local and global features and enabling dynamic information exchange between the CNN and transformer components. Experimental results on four benchmark datasets, ISIC2018, COVID-19, and liver cirrhosis (line array, convex array), demonstrate that the proposed model achieves superior classification performance, outperforming both CNN and transformer-only architectures.

This study aims to develop an algorithm for selecting the most informative and diverse patches from breast histopathology images while excluding irrelevant areas to enhance cancer detection. A key contribution of the method is the creation of a new metric called patch score that integrates SHAP values with Haralick features, improving both explainability and diagnostic accuracy. The algorithm begins by calculating Haralick features and measuring cosine similarity between patches to construct a graph, which is then used to train a graph neural network (GNN). To assess each patch's contribution to the analysis, we employ a SHAP explainer on the GNN model. The SHAP values and the features from each patch are then used to calculate a score called the patch score, which determines the importance of each patch. Additionally, to incorporate diversity in the selected patches, all patches are clustered based on local binary patterns, and the patch with the highest patch score from each cluster is selected to obtain the final patches for image classification. Features extracted from these patches using a ResNeXt 50 model, fused with 3-norm pooling, are used to classify the images as benign or malignant. The proposed framework was evaluated on the BreakHis dataset and demonstrated superior accuracy and precision compared to existing methods. By integrating both explainability and diversity into patch selection, the algorithm delivers a robust, interpretable model, offering dependable diagnostic support for pathologists.

Medical image segmentation is a crucial process in medical image analysis, with convolutional neural network (CNN)-based methods achieving notable success in recent years. Among these, U-Net has gained widespread use due to its simple yet effective architecture. However, CNNs still struggle to capture global, long-range semantic information. To address this limitation, we present CS U-NET, a novel method built upon Swin-U-Net, which integrates spatial and contextual attention mechanisms. This hybrid approach combines the strengths of both transformers and U-Net architectures to enhance segmentation performance. In this framework, tokenized image patches are processed through a transformer-based U-shaped encoder-decoder, enabling the learning of both local and global semantic features via skip connections. Our method achieves a Dice Similarity Coefficient of 78.64% and a 95% Hausdorff distance of 21.25 on the Synapse multiorgan segmentation dataset, outperforming Trans-U-Net and other state-of-the-art U-Net variants by 4% and 6%, respectively. The experimental results highlight the significant improvements in prediction accuracy and edge detail preservation provided by our approach.