International Journal of Imaging Systems and Technology最新文献

Diabetic retinopathy (DR) poses a serious threat to vision, emphasising the need for early detection. Manual analysis of fundus images, though common, is error-prone and time-intensive. Existing automated diagnostic methods lack precision, particularly in the early stages of DR. This paper introduces the Soft Convolutional Block Attention Module-based Network (Soft-CBAMNet), a deep learning network designed for severity detection, which features Soft-CBAM attention to capture complex features from fundus images. The proposed network integrates both the convolutional block attention module (CBAM) and the soft-attention components, ensuring simultaneous processing of input features. Following this, attention maps undergo a max-pooling operation, and refined features are concatenated before passing through a dropout layer with a dropout rate of 50%. Experimental results on the APTOS dataset demonstrate the superior performance of Soft-CBAMNet, achieving an accuracy of 85.4% in multiclass DR grading. The proposed architecture has shown strong robustness and general feature learning capability, achieving a mean AUC of 0.81 on the IDRID dataset. Soft-CBAMNet's dynamic feature extraction capability across all classes is further justified by the inspection of intermediate feature maps. The model excels in identifying all stages of DR with increased precision, surpassing contemporary approaches. Soft-CBAMNet presents a significant advancement in DR diagnosis, offering improved accuracy and efficiency for timely intervention.

Medical image labeling requires specialized knowledge; hence, the solution to the challenge of medical image classification lies in efficiently utilizing the few labeled samples to create a high-performance model. Building a high-performance model requires a complicated convolutional neural network (CNN) model with numerous parameters to be trained which makes the test quite expensive. In this paper, we propose optimizing a lightweight deep learning model with only five convolutional layers using the particle swarm optimization (PSO) algorithm to find the best number of kernel filters for each convolutional layer. For colored red, green, and blue (RGB) images acquired from different data sources, we suggest using stain separation using color deconvolution and horizontal and vertical flipping to produce new versions that can concentrate the representation of the images on structures and patterns. To mitigate the effect of training with incorrectly or uncertainly labeled images, grades of disease could have small variances, we apply a second-pass training excluding uncertain data. With a small number of parameters and higher accuracy, the proposed lightweight deep learning model optimization (LDLMO) algorithm shows strong resilience and generalization ability compared with most recent research on four MedMNIST datasets (RetinaMNIST, BreastMNIST, DermMNIST, and OCTMNIST), Medical-MNIST, and brain tumor MRI datasets.

Skin cancer is a pressing global health issue, with high incidence and mortality rates. Convolutional neural network (CNN) models have been proven to be effective in improving performance in skin lesion image classification and reducing the medical burden. However, the inherent class imbalance in training data caused by the difficulty of collecting dermoscopy images leads to categorical overfitting, which still limits the effectiveness of these data-driven models in recognizing few-shot categories. To address this challenge, we propose a dynamic multi-output convolutional neural network (DMO-CNN) model that incorporates exit nodes into the standard CNN structure and includes feature refinement layers (FRLs) and an adaptive output scheduling (AOS) module. This model improves feature representation ability through multi-scale sub-feature maps and reduces the inter-layer dependencies during backpropagation. The FRLs ensure efficient and low-loss down-sampling, while the AOS module utilizes a trainable layer selection mechanism to refocus the model's attention on few-shot lesion categories. Additionally, a novel correction factor loss is introduced to supervise and promote AOS convergence. Our evaluation of the DMO-CNN model on the HAM10000 dataset demonstrates its effectiveness in multi-class skin lesion classification and its superior performance in recognizing few-shot categories. Despite utilizing a very simple VGG structure as the sole backbone structure, DMO-CNN achieved impressive performance of 0.885 in BACC and 0.983 in weighted AUC. These results are comparable to those of the ensemble model that won the ISIC 2018 challenge, highlighting the strong potential of DMO-CNN in dealing with few-shot skin lesion data.

In this paper, a reduced accelerated adaptive fast iterative shrinkage threshold algorithm based on Smooth-Lasso regularization (SL-RAFISTA-BB) is proposed for fluorescence molecular tomography (FMT) 3D reconstruction. This method uses the Smooth-Lasso regularization to fuse the group sparse prior information which can balance the relationship between the sparsity and smoothness of the solution, simplifying the process of calculation. In particular, the convergence speed of the FISTA is improved by introducing a reduction strategy and Barzilai-Borwein variable step size factor, and constructing a continuation strategy to reduce computing costs and the number of iterations. The experimental results show that the proposed algorithm not only accelerates the convergence speed of the iterative algorithm, but also improves the positioning accuracy of the tumor target, alleviates the over-sparse or over-smooth phenomenon of the reconstructed target, and clearly outlines the boundary information of the tumor target. We hope that this method can promote the development of optical molecular tomography.

Diabetic retinopathy (DR) is an important cause of blindness. If not diagnosed and treated in a timely manner, it can lead to irreversible vision loss. The diagnosis of DR relies heavily on specialized ophthalmologists. In recent years, with the development of artificial intelligence a number of diagnostics using this technique have begun to appear. One method for diagnosing diseases in this field is to segment four common kinds of lesions from color fundus images, including: exudates (EX), soft exudates (SE), hemorrhages (HE), and microaneurysms (MA). In this paper, we propose a segmentation model for DR based on deep learning. The main part of the model consists of two layers of improved U-Net network based on transformer, corresponding to the two stages of coarse segmentation and fine segmentation, respectively. The model can segment four common kinds of lesions from the input color fundus image at the same time. To validate the performance of our proposed model, we test our model on three public datasets: IDRiD, DDR, and DIARETDB1. The test results show that our proposed model achieves competitive results compared with the existing methods in terms of PR-AUC, ROC-AUC, Dice, and IoU, especially for lesions segmentation of SE and MA.

The skin, a crucial organ, plays a protective role in the human body, emphasizing the significance of early detection of skin diseases to prevent potential progression to skin cancer. The challenge lies in diagnosing these diseases at their early stages, where visual resemblance complicates differentiation, highlighting the need for an innovative automated method for precisely identifying skin lesions in biomedical images. This paper introduces a holistic methodology that combines DenseNet, multi-scale feature boundary module (MFBM), and feature fusion and decoding engine (FFDE) to tackle challenges in existing deep-learning image segmentation methods. Furthermore, a convolutional neural network model is designed for the classification of segmented images. The DenseNet encoder efficiently extracts features at four resolution levels, leveraging dense connectivity to capture intricate hierarchical features. The proposed MFBM plays a crucial role in extracting boundary information, employing parallel dilated convolutions with various dilation rates for effective multi-scale information capture. To overcome potential disadvantages related to the conversion of features during segmentation, our approach ensures the preservation of context features. The proposed FFDE method adaptively fuses features from different levels, restoring skin lesion location information while preserving local details. The evaluation of the model is conducted on the HAM10000 dataset, which consists of 10 015 dermoscopy images, yielding promising results.

Automatic segmentation of the fundus retinal vessels and accurate classification of the arterial and venous vessels play an important role in clinical diagnosis. This article proposes a fundus retinal vascular segmentation and arteriovenous classification network that combines the adversarial training and attention mechanism to address the issues of fundus retinal arteriovenous classification error and ambiguous segmentation of fine blood vessels. It consists of three core components: discriminator, generator, and segmenter. In order to address the domain shift issue, U-Net is employed as a discriminator, and data samples for arterial and venous vessels are generated with a generator using an unsupervised domain adaption (UDA) approach. The classification of retinal arterial and venous vessels (A/V) as well as the segmentation of fine vessels is improved by adding a self-attention mechanism to improve attention to vessel edge features and the terminal fine vessels. Non-strided convolution and non-pooled downsampling methods are also used to avoid losing fine-grained information and learning less effective feature representations. The performance of multi-class blood vessel segmentation is as follows, per test results on the DRIVE dataset: F1-score (F1) has a value of 0.7496 and an accuracy of 0.9820. The accuracy of A/V categorization has increased by 1.35% when compared to AU-Net. The outcomes demonstrate that by enhancing the baseline U-Net, the strategy we suggested enhances the automated classification and segmentation of blood vessels.

Chest x-ray (CXR) is an effective method for diagnosing heart failure, and identifying important features such as cardiomegaly, effusion, and edema on patient chest x-rays is significant for aiding the treatment of heart failure. However, manually identifying a vast amount of CXR data places a huge burden on physicians. Deep learning's progression has led to the utilization of this technology in numerous research aimed at tackling these particular challenges. However, many of these studies utilize global learning methods, where the contribution of each pixel to the classification is considered equal, or they overly focus on small areas of the lesion while neglecting the global context. In response to these issues, we propose the Global Local Attention Network (GLAN), which incorporates an improved attention module on a branched structure. This enables the network to capture small lesion areas while also considering both local and global features. We evaluated the effectiveness of the proposed model by testing it on multiple public datasets and real-world datasets. Compared to the state-of-the-art methods, our network structure demonstrated greater accuracy and effectiveness in the identification of three key features: cardiomegaly, effusion, and edema. This provides more targeted support for diagnosing and treating heart failure.

Thyroid nodules are a common endocrine system disorder for which accurate ultrasound image segmentation is important for evaluation and diagnosis, as well as a critical step in computer-aided diagnostic systems. However, the accuracy and consistency of segmentation remains a challenging task due to the presence of scattering noise, low contrast and resolution in ultrasound images. Therefore, we propose a deep learning-based CAD (computer-aided diagnosis) method, STU³Net in this paper, aiming at automatic segmentation of thyroid nodules. The method employs a modified Swin Transformer combined with a CNN encoder, which is capable of extracting morphological features and edge details of thyroid nodules in ultrasound images. In decoding through the features for image reconstruction, we introduce a modified three-layer U-Net network with cross-layer connectivity to further enhance image reduction. This cross-layer connectivity enhances the network's capture and representation of the contained image feature information by creating skip connections between different layers and merging the detailed information of the shallow network with the abstract information of the deeper network. Through comparison experiments with current mainstream deep learning methods on the TN3K and BUSI datasets, we validate the superiority of the STU³Net method in thyroid nodule segmentation performance. The experimental results show that STU³Net outperforms most of the mainstream models on the TN3K dataset, with Dice and IoU reaching 0.8368 and 0.7416, respectively, which are significantly better than other methods. The method demonstrates excellent performance on these datasets and provides radiologists with an effective auxiliary tool to accurately detect thyroid nodules in ultrasound images.