International Journal of Imaging Systems and Technology最新文献

This study explores the effectiveness of deep learning methodologies in the detection and classification of lumbar disc intensity using MRI scans. Initially, region-based deep learning frameworks, including Faster R-CNN and Mask R-CNN with different backbones such as ResNet50 and ResNet101 are evaluated. Results demonstrated that backbone selection significantly impacts model performance, with Mask R-CNN combined with ResNet101 achieving a remarkable [email protected] (AP50) of 99.83%. In addition to object detection models, Transformer-based classification architectures, including MaxViT, Vision Transformer (ViT), a Hybrid CNN-ViT model, and Fine-Tuned Enhanced Pyramid Network (FT-EPN), are implemented. Among these, the Hybrid model achieved the highest classification accuracy (83.1%), while MaxViT yielded the highest precision (0.804). Comparative analyses highlighted that while Mask R-CNN models excelled in segmentation and detection tasks, Transformer-based models provided effective solutions for direct severity classification of lumbar discs. These findings emphasize the critical role of both backbone architecture and model type in optimizing diagnostic performance. The study demonstrates the potential of integrating region-based and Transformer-based models in advancing automated lumbar spine assessment, paving the way for more accurate and reliable medical diagnostic systems.

Deep learning-based methods have achieved great progress in CT image segmentation in recent years. However, the lack of unified large-scale datasets, unbalanced categories of segmented images, blurred boundaries between infected and healthy regions, and different sizes and shapes of lesions have led to the fact that the existing methods still have challenges in further improving the segmentation accuracy in medical applications. In this work, a novel network with spectrum Transformer and triplet attention for CT image segmentation is proposed. The computed spectrum Transformer module and the triplet attention module (TAM) are fused in a parallel way utilizing the parallel hybrid fusion module (PHFM), which extracts global and local contextual information from sequence features and cross-dimensional features. A spectrum Transformer block (STB) is proposed, which utilizes the fast Fourier transform (FFT) to learn the weights of each frequency component in the spectral space. Extensive comparison experiments are conducted on both the COVID-19-Seg dataset and Mosmeddata, showing that the proposed network has better accuracy than most existing methods for CT image segmentation tasks. In particular, the proposed model achieves improvement by 1.29% and 2.28% in terms of DSC, respectively, by 1.15% and 1.35% in terms of mIoU. SEN metrics increase by 1.45% and 1.48%, respectively. PRE also achieves the best results, showing its significant advantage in accurately segmenting medical images. SPE also shows quite good results in both datasets. Ablation studies show that the proposed STB and TAM modules improve the segmentation performance significantly.

Low-Intensity Transcranial Focused Ultrasound (LITFU) is a highly anticipated non-invasive neuromodulation technique in terms of targeting accuracy, spatial resolution, stimulation depth, and reversibility. However, LITFU still faces challenges, most notably is and the reduced precision in deep tissue targeting and limited stimulation depth due to the distortion of the acoustic field by the skull. This study proposes a method for using a rectangular concave ultrasonic transducer to achieve deep brain stimulation with LITFU under Magnetic Resonance Imaging Acoustic Radiation Force Imaging guidance. First, k-Wave simulation modeling is performed by the skull magnetic resonance imaging (MRI) with ultrashort echo time (UTE) sequence to calculate the acoustic pressure distribution and the maximum value within the tissue. Then, the time required for maximum tissue displacement is calculated using a biological tissue elastic deformation model. Based on the displacement and time, the motion encoding gradient is optimized to obtain a more distinct phase contrast. We validated the feasibility of this method on water-filled balloons and ex vivo porcine brains, provided the optimized ultrasonic delay time coefficient, and achieved precise LITFU focusing on the amygdala located deep within the human volunteers. The results demonstrate that the rectangular transducer achieves a maximum stimulation depth of 80 mm with a maximum peak acoustic pressure of 3 MPa. The simulated and ARFI measured acoustic fields were compared and their consistency and differences were demonstrated. The improved EPI-ARFI significantly enhances measurement sensitivity and reduces scanning time to 24 s and three-dimensional imaging can be achieved. This work provides a localization and verification method for LITFU guided by MRI, deep brain stimulation of LITFU in future clinical applications.

Lung cancer significantly contributes to cancer mortality worldwide, and prompt diagnosis is essential for enhancing patient outcomes. Identifying the condition at an early stage remains challenging, particularly in regions with limited medical facilities and experienced radiologists. The paper aims to introduce a fully automated DL system capable of identifying, segmenting, and classifying lung cancer at an early phase. This approach aims to enhance the precision and efficacy of lung cancer screening in resource-constrained environments. The recommended architecture has three stages: (1) lung preprocessing, employing a bilateral filter to enhance image quality; (2) lung segmentation, utilizing Otsu's thresholding to delineate lung areas; and (3) lung cancer classification, implementing a modified CNN referred to as the P-Model. The framework was assessed utilizing the publicly available IQ-OTH/NCCD dataset. The proposed framework exhibited high performance metrics, with lung segmentation accuracy at 96%, classification precision at 97%, sensitivity at 96%, and an F1 score of 96%. Additionally, the Matthews correlation coefficient was recorded at 0.9406. The findings indicate that our paradigm surpasses previous studies across all pertinent evaluation metrics. The proposed deep learning approach shows considerable potential for precise detection and classification of lung cancer, particularly in areas with limited resources. This study significantly advances the field of lung cancer detection and has the capacity to enhance healthcare outcomes and patient care standards globally.

The increasing global burden of lung diseases necessitates the development of improved diagnostic tools. According to the WHO, hundreds of millions of individuals worldwide are currently affected by various forms of lung disease. The rapid advancement of artificial neural networks has revolutionized lung disease diagnosis, enabling the development of highly effective detection and classification systems. This article presents dual channel neural networks in image feature extraction based on classical CNN and vision transformers for multi-label lung disease diagnosis. Two separate subnetworks are employed to capture both global and local feature representations, thereby facilitating the extraction of more informative and discriminative image features. The global network analyzes all-organ regions, while the local network simultaneously focuses on multiple single-organ regions. We then apply a novel feature fusion operation, leveraging a multi-head attention mechanism to weight global features according to the significance of localized features. Through this multi-channel approach, the framework is designed to identify complicated and subtle features within images, which often go unnoticed by the human eye. Evaluation on the ChestX-ray14 benchmark dataset demonstrates that our hybrid model consistently outperforms established state-of-the-art architectures, including ResNet-50, DenseNet-121, and CheXNet, by achieving significantly higher AUC scores across multiple thoracic disease classification tasks. By incorporating test-time augmentation, the model achieved an average accuracy of 95.7% and a specificity of 99%. The experimental findings indicated that our model attained an average testing AUC of 87%. In addition, our method tackles a more practical clinical problem, and preliminary results suggest its feasibility and effectiveness. It could assist clinicians in making timely decisions about lung diseases.

Minimally invasive surgery can benefit significantly from automated surgical tool detection, enabling advanced analysis and assistance. However, the limited availability of annotated data in surgical settings poses a challenge for training robust deep learning models. This paper introduces a novel staged adaptive fine-tuning approach consisting of two steps: a linear probing stage to condition additional classification layers on a pre-trained CNN-based architecture and a gradual freezing stage to dynamically reduce the fine-tunable layers, aiming to regulate adaptation to the surgical domain. This strategy reduces network complexity and improves efficiency, requiring only a single training loop and eliminating the need for multiple iterations. We validated our method on the Cholec80 dataset, employing CNN architectures (ResNet-50 and DenseNet-121) pre-trained on ImageNet for detecting surgical tools in cholecystectomy endoscopic videos. Our results demonstrate that our method improves detection performance compared to existing approaches and established fine-tuning techniques, achieving a mean average precision (mAP) of 96.4%. To assess its broader applicability, the generalizability of the fine-tuning strategy was further confirmed on the CATARACTS dataset, a distinct domain of minimally invasive ophthalmic surgery. These findings suggest that gradual freezing fine-tuning is a promising technique for improving tool presence detection in diverse surgical procedures and may have broader applications in general image classification tasks.

Medical image segmentation plays a crucial role in biomedical engineering and computer-aided medical systems. Fully supervised medical image segmentation algorithms have achieved significant performance improvements. However, these improvements often rely on large amounts of finely annotated data, while semi-supervised medical image segmentation can better address this issue. In this work, we propose the MaFMatch model based on the principle of weak-to-strong consistency. It effectively addresses the limitations of current semi-supervised medical image segmentation, such as noise generated by perturbation leading the model to learn in unfavorable directions, and simple feature perturbations being insufficient to explore a broader perturbation space. On the one hand, this approach introduces a mixed data perturbation flow to utilize all pixel information, making the model inclined to consider global semantic information rather than simply discarding unreliable pixels. On the other hand, to fully utilize feature perturbation information flow, we propose a feature augmentation perturbation scheme that simultaneously supplements and discards information from the original feature flow, enabling the model to effectively overcome the diminishing marginal returns brought about by multi-branch perturbations. MaFMatch achieved an mDsc of 90.8 on the automatic cardiac diagnosis challenge (ACDC) dataset. It outperforms most methods across major metrics on both ACDC and LA datasets. The code is available at https://github.com/HandsomeRed/MaFMatch-main.

Accurate segmentation of brain tumors in magnetic resonance imaging (MRI) is critical for diagnosis, treatment planning, and longitudinal monitoring. The development of robust and generalizable deep learning models for tumor segmentation is hindered by challenges such as data privacy, limited annotations, and domain variability across clinical institutions. To address these issues, we propose a dual-model federated learning model for brain tumor segmentation that enables collaborative model training without sharing patient data. The model employs two specialized architectures: a Multi-Scale Encoder U-Net (MSE-UNet) for fine-grained, multi-resolution feature extraction and a Residual Attention Transpose U-Net (ART-UNet) that leverages residual learning and dual attention mechanisms to enhance contextual sensitivity and robustness under non-IID conditions. To ensure effective learning across distributed, heterogeneous clients, we introduce a Dual-Model Architecture-Aware Aggregation (DAAA) strategy, which performs independent, performance-weighted aggregation of each architecture's updates. The proposed method is evaluated on two benchmark datasets, BraTS 2018 and TCGA-LGG, demonstrating superior performance compared to several state-of-the-art baselines. The model achieves Dice scores of 91.30% and 90.10% on BraTS and TCGA-LGG, respectively, with improved IoU, sensitivity, and boundary precision. Ablation studies confirm that each component, including auxiliary supervision, architectural duality, and adaptive aggregation, contributes significantly to overall performance. Clinically, this framework offers a scalable and privacy-preserving solution that can be integrated into real-world healthcare systems without compromising patient data security. By enabling cross-institutional collaboration and ensuring robust performance across diverse imaging protocols, the proposed model facilitates early and accurate tumor delineation, supporting radiologists in critical decision-making processes such as surgical planning, radiotherapy guidance, and follow-up assessment. This study establishes a foundation for practical deployment in federated clinical environments, particularly within resource-constrained or privacy-sensitive institutions.

Skin cancer is a prevalent and potentially life-threatening condition that requires accurate and timely diagnosis. Dermoscopic imaging aids in diagnosis but is often limited by manual interpretation, which can be time–intensive and error–prone. Automated classification systems using convolutional neural networks (CNNs) have shown significant promise in enhancing accuracy and efficiency. To further improve classification performance, we propose a novel deep learning-based Multi–Scale Feature Map Fusion (MFMF) model that extracts and fuses features from multiple convolutional layers. The MFMF module effectively combines these multi–scale features, enabling robust feature capture even with limited datasets. Additionally, a hair removal algorithm is proposed to enhance prediction accuracy through improved image preprocessing. On the HAM10000 dataset, our proposed model achieves an overall accuracy of 96.12%, an AUC of 98.7%, and an F1 Score of 93.29%, along with notable improvements in precision, recall, and sensitivity.