International Journal of Imaging Systems and Technology最新文献

Macular edema is a retinal disorder that can lead to significant vision loss, underscoring the need for accurate and intelligent automated diagnosis. However, its subtle manifestations in color fundus photography (CFP) pose considerable challenges for conventional deep learning models. In this work, we propose a novel diagnostic framework that integrates Dempster–Shafer (D–S) evidence theory—a principled approach for uncertainty quantification and multi-source information fusion—with the advanced Mamba architecture. The proposed method employs a dual-branch network to selectively enhance and extract discriminative features from both bright and dark regions of fundus images. These features are dynamically aligned and fused via an Adaptive Multi-Branch Feature Synthesis (AMFS) module. To model long-range dependencies and aggregate complementary information from multiple scanning views, we introduce a multi-scan Mamba module, whose outputs are further fused using a principled D–S evidence mechanism. This synergistic integration of mathematical theory and deep learning not only reduces information redundancy but also enables confidence-aware automated decision-making. Extensive experiments on three retinal image datasets—IDRiD, Messidor, and a proprietary clinical cohort—demonstrate that our framework consistently outperforms state-of-the-art methods in terms of accuracy, F1-score, and robustness. These results highlight the promise of combining evidence theory with modern deep learning for challenging medical image analysis tasks.

Electroencephalogram (EEG) plays a vital role in identifying the neural activities and human behaviors. EEG is a high-temporal-resolution signal that is contaminated and influenced by various external factors and artifacts. It is very difficult to identify and analyze the EEG signal when it is contaminated by artifacts. The impact of artifacts in EEG signals became an open challenge in EEG signal processing. Even though many techniques and methods are available to remove the artifacts in EEG, analyzing the contaminated EEG signals remains a significant challenge. The signal is polluted by different artifacts like eye blink, muscle activities, environmental interfaces, and so on. Compared with other artifacts and noises, task-specific artifacts such as motor movement and motor imagery have a greater impact on EEG signals due to their direct inference on cognitive signals, and isolating task-specific artifacts from the EEG signal is difficult. To overcome these challenges, a novel hybrid task-specific Dynamic Multistage k-Means Clustering algorithm (TS-DMKC) has been proposed that detects and extracts the processed EEG signal, which will be helpful for all the EEG applications. Initially, the motor movement and motor imagery artifacts are distinguished using the Fast ICA by applying the filtering threshold. Then, a multistage k-means clustering algorithm is employed to isolate the artifacts in different clusters dynamically, and the cluster effectiveness is calculated by using the silhouette score. This proposed novel hybrid algorithm will be useful in various EEG applications like neuroscience, medical diagnosis, brain–computer interface, authentication, brain signal analysis, and gaming. The experimental results demonstrate that the proposed algorithm TS-DMKC achieved a classification accuracy of 94.2% using the Physionet motor movement/imagery dataset, reducing task-dependent variability and offering more robust and reliable EEG systems.

Accurate diagnosis of diseases relies on the subjective experience of physicians. Automatic image segmentation technology can assist doctors in quickly locating lesion areas, thereby improving diagnostic efficiency. In the field of medical image segmentation, the current mainstream models are mostly variants of Swin-Unet, which perform well on large-sample datasets. However, due to the window constraints of the Swin-Transformer architecture, these models often face performance limitations when processing small-sample datasets. In contrast, convolutional-based Unet variant models exhibit better segmentation performance under small-sample conditions but demonstrate poorer segmentation results for medical images with blurred boundaries. To tackle these challenges, this paper introduces a novel boundary information-enhanced architecture, the Hybrid-Parallel-Attention and Pyramid-Edge-Extraction Enhanced UNet (HPPE-Unet). Based on an encoder-decoder architecture of convolutional neural networks, the model incorporates a multi-scale feature extraction strategy and an optimized skip-connection mechanism. It integrates two key modules: the Pyramid Edge Extraction (PEE) module and the Hybrid Parallel Attention (HP) module. The PEE module is applied at each stage of the encoder and decoder, significantly enhancing the model's ability to capture subtle boundary structures through multi-scale feature fusion and boundary reinforcement mechanisms. The HP employs a residual parallel structure combining spatial (SA) and channel attention (CA) blocks, effectively bridging the semantic gap between features in the encoding and decoding stages and addressing the issue of information loss in skip connections between the encoder and decoder. The proposed method was evaluated on an aortic dissection dataset provided by a tertiary hospital. The results show that the method achieved a Dice similarity coefficient (DSC) of 97.65% and a Mean intersection over union (Miou) of 97.92% in the segmentation task. On the public Automated Cardiac Diagnostic Challenge (ACDC) dataset, the DSC reached 90.39%. These results demonstrate that the proposed method holds significant practical value for clinical disease diagnosis.

Classical self-training methods for graph convolutional networks (GCNs) assume that both labeled and unlabeled data follow the identical distribution. However, these works do not work well when they are used in medical applications such as classifying autism spectrum disorder (ASD) in multiple centers, where the unlabeled samples from different imaging centers have varying distribution shifts from the labeled samples. To this end, we propose uncertainty-aware graph self-training (UA-GST) by extending graph self-training to a new situation, in which the labeled data come from one imaging center and the unlabeled data come from several other imaging centers. Specifically, an uncertainty-aware mechanism is proposed to select unlabeled centers and adversarial domain adaptation is introduced into graph self-training to reduce domain shift between centers. With the progression of self-training, more pseudo-labeled test samples are gradually included in the training set, and a final model is finally trained on all the labeled and pseudo-labeled samples. Considering the over-confidence issue of the classifier, an evidential GCN is further proposed to estimate the uncertainty of the pseudo-labels using Dempster–Shafer (D–S) evidence theory. It is evaluated on the Autism Brain Imaging Data Exchange (ABIDE). Experimental results verified the effectiveness of the proposed method in classifying ASD in multiple centers, outperforming existing state-of-the-art methods.

The accurate survival prediction in glioblastoma patients remains a major challenge, largely owing to the considerable variability among brain tumor subtypes and their clinical profiles. This research introduces an innovative SimCLR-Twin Critic DDPG framework that integrates self-supervised representation learning with deep reinforcement learning for enhanced prognostic prediction. Tumor subregion segmentations were independently obtained using four advanced models namely nnU-Net, VNet, SwinUNETR, and UNet. Based on the performance of these segmentation models, deep features were extracted from the nnU-Net segmented tumor subregions and combined with handcrafted radiomics features. The combined features were subjected to the feature selection process using BorutaShap, followed by self-supervised pretraining through SimCLR to improve feature representation and generalization. The optimized features were then fed into a Twin Critic DDPG agent designed for regression-based survival time predictions. The proposed method outperformed existing techniques on the BraTS 2020 dataset, achieving the lowest MSE of 31 062.36 and the highest C-index of 0.87 in overall survival prediction. To further confirm robustness and generalizability, the framework was externally validated on the BraTS 2019 dataset, where it achieved a comparable MSE of 31 156.56 and a C-index of 0.86 using fused features. Additionally, LIME-based local interpretability provided clinically relevant explanations for individual predictions, thereby enhancing trust in the AI-driven system. This study highlights the SimCLR-Twin Critic DDPG framework as a robust and interpretable solution for accurate survival prediction in glioblastoma patients.

This research dives into creating optimized neural network architectures tailored for image classification and medical image segmentation, with a particular emphasis on analyzing lung nodules using the LIDC-IDRI dataset. We introduce a hybrid approach to neural architecture search (NAS) that fuses convolutional neural networks (CNNs), transformer-based models, and custom-built graph architectures. Our evolutionary strategies employ genetic operations like crossover and mutation to gradually enhance these architectures, while the NSGA-III algorithm helps us navigate the tricky balance of multiple conflicting objectives. The goal of our method is to boost performance metrics such as the F1 score and classification accuracy, all while keeping an eye on minimizing computational complexity, which we measure in terms of FLOPS, parameter count, and inference time. Our experiments demonstrate competitive performance on the CIFAR-10 and CIFAR-100 datasets, with promising results on the LIDC-IDRI segmentation task. This work aims to push the boundaries of automated model design, contributing to more efficient and effective deep learning architectures in both general and medical imaging contexts.

This research work aims to improve brain tumor segmentation using a transformer architecture known as SegFormer. It is essential to segment brain tumors correctly for surgical planning and tumor progression assessment. Many existing segmentation techniques have problems in visualizing the shapes and structures of a brain tumor; they include issues of overfitting and underfitting. The SegFormer model is developed to change transformers employed in natural language processing to overcome such difficulties. Moreover, the SegFormer model improves accuracy by adequately integrating the image patches, hence improving the structures of the tumors. Accurate segmentation of brain tumors is crucial in clinical neuro-oncology, diagnosis, treatment planning, and follow-up care. Here, we present a novel deep-learning-based SegFormer framework for brain tumor segmentation that includes hierarchical patch embedding and transformer-based self-attention to consider both the global and local tumor features simultaneously. In comparison to CNN-based and hybrid models, our model is much more accurate in segmentation and has a lower computational cost. Our initial attempt to apply the proposed model on the BraTS2020 benchmark dataset, with a Dice score of 0.9425, exceeded the state-of-the-art models, including RMTF-Net, HMNet, and CBAM TransUnet. To additionally confirm generalization, we applied the model to the BraTS2021 dataset (Dice: 0.92123), the dataset of ischemic stroke lesion segmentation in ISLES (Dice: 0.874), and the TCGA dataset (Dice: 0.860). These findings indicate the soundness and applicability of our structure in a variety of clinical imaging conditions. We make the following contributions: (1) we combine transformer-based self-attention and lightweight decoding to segment the image accurately, (2) we achieve higher results on BraTS2020 than state-of-the-art models, and (3) we also validate the obtained results on a large set of datasets, which confirms the good generalization. These results establish SegFormer as a promising foundation for the next generation of brain tumor and lesion segmentation tasks in clinical imaging.

To achieve precise and efficient skin cancer segmentation, an innovative SkinSegNet architecture is proposed. Inspired by U-Net, SkinSegNet uses an encoder–decoder architecture incorporating advanced feature extraction and attention mechanisms. The encoder utilizes convolutional blocks and pooling attention (PA) layers to downsample feature maps and focus on significant regions. At the bottleneck, the proposed feature aggregation block integrates channel-aware multihead attention (CAMA) and Adaptive Spatial Cross-Scale Attention (ASCA) modules. These modules let the model capture channel relationships and complex spatial dependencies for accurate segmentation. The decoder reconstructs the segmentation mask through continuous upsampling and skips connections, ensuring fine-grained spatial detail is preserved. Experiments are carried out on benchmark datasets namely ISIC 2016, ISIC 2017, and ISIC 2018 demonstrating SkinSegNet's superior performance in skin lesion segmentation, achieving state-of-the-art accuracy of 95.12%, 94.01%, and 95.04%, respectively. Furthermore, cross-dataset experiments show that SkinSegNet performs well on ISIC datasets, demonstrating its strong generalization ability.

Convolutional neural networks (CNNs) have augmented conventional approaches in medical imaging by improving tumor detection and classification efficacy. To enable oncologists to diagnose abnormalities promptly, this research proposes an innovative classification framework for breast cancer diagnosis. It integrates an improved optimization method with a hybridized CNN architecture. In this article, a custom CNN, feed-forward and backpropagation have been implemented. The scaled conjugate algorithm is employed in the feed-forward paradigm, yielding a formidable accuracy of 99.1%. On the other hand, backpropagation implements stochastic gradient descent and exhibits a remarkable accuracy rate of 97.3%. Additionally, by integrating the grey wolf optimization (GWO) algorithm with the Backpropagation Neural Network (BPNN), model performance is enhanced by optimizing parameters and accuracy to 100%. Furthermore, the custom CNN achieves an incredible 98% accuracy by utilizing the Adam optimizer in conjunction with the ReduceLROnPlateau approach. Statistical analysis utilizing Analysis of Variance (ANOVA) and Honestly Significant Difference (HSD) tests has demonstrated that the suggested hybrid model improves detection accuracy and reliability. These results highlight the adaptability and effectiveness of various optimization techniques in enhancing the performance of neural network models on a range of demanding tasks related to machine learning and pattern recognition.