Squeeze-and-Excitation Block Research Articles

Deep Learning-Based Ovarian Cancer Subtype Classification Using VGG16 and MobileNetV2 with Squeeze-and-Excitation Blocks

Background: Ovarian carcinoma remains one of the deadliest gynecological malignancies due to its heterogeneity and late-stage detection. Accurate classification of its subtypes—High-Grade Serous Carcinoma (HGSC), Clear-Cell Carcinoma (CC), Endometrioid Carcinoma (EC), Low-Grade Serous Carcinoma (LGSC), and Mucinous Carcinoma (MC)—is essential for tailored treatments, yet traditional histopathological methods often lack precision. This study aimed to develop a deep learning (DL) model to enhance ovarian cancer subtype classification using histopathological images. Methods: Histopathological images were collected from medical repositories, and data augmentation techniques were applied to increase dataset diversity. Two convolutional neural network (CNN) architectures, VGG16 and MobileNetV2, were fine-tuned using transfer learning to classify the subtypes. The models were pre-trained on the ImageNet dataset and evaluated through accuracy, precision, recall, F1-score, and ROC-AUC, with K-fold cross-validation ensuring robustness. Results: Results indicated that VGG16 improved classification over baseline CNN models, while MobileNetV2, with Squeeze-and-Excitation (SE) blocks, achieved the highest performance, offering greater accuracy and computational efficiency. MobileNetV2’s lightweight architecture captured intricate tissue patterns more effectively, making it the superior model. Conclusion: This study highlights the potential of advanced DL models, particularly MobileNetV2 with SE block attention, for improving ovarian cancer subtype classification. These findings offer promising implications for clinical practice and personalized treatment approaches. Future research should focus on larger datasets and integrating multimodal data for further advancements.

Detection of Pneumonia Using A Hybrid Approach Consisting of MobileNetV2 and Squeeze-and-Excitation Network

Pneumonia is a global health concern, responsible for a significant number of deaths. Its diagnostic challenge arises from visual similarities it shares with various respiratory diseases, such as tuberculosis, complicating accurate identification. Furthermore, the variability in acquiring and processing chest X-ray (CXR) images can impact image quality, posing a hurdle for dependable algorithm development. To address this, resilient data-centric algorithms, trained on comprehensive datasets and validated through diverse imaging methods and radiology expertise, are imperative. This study presents a deep learning approach designed to distinguish between normal and pneumonia cases. The model, a hybrid of MobileNetV2 and the Squeeze-and-Excitation (SE) block, aims to reduce learnable parameters while enhancing feature extraction and classification. Integration of the SE block enhances classification performance, despite a slight parameter increase. The model was trained and tested on a dataset of 5856 CXR images from Kaggle's medical imaging challenge. Results demonstrated the model's exceptional performance, achieving an accuracy of 98.81%, precision of 98.79%, recall rate of 98.24%, and F1-score of 98.51%. Comparative analysis with various Convolutional neural network-based pre-trained models and recent literature studies confirmed its superiority, solidifying its potential as a robust tool for pneumonia detection, thus addressing a critical healthcare need.

MetaSwin: a unified meta vision transformer model for medical image segmentation.

Transformers have demonstrated significant promise for computer vision tasks. Particularly noteworthy is SwinUNETR, a model that employs vision transformers, which has made remarkable advancements in improving the process of segmenting medical images. Nevertheless, the efficacy of training process of SwinUNETR has been constrained by an extended training duration, a limitation primarily attributable to the integration of the attention mechanism within the architecture. In this article, to address this limitation, we introduce a novel framework, called the MetaSwin model. Drawing inspiration from the MetaFormer concept that uses other token mix operations, we propose a transformative modification by substituting attention-based components within SwinUNETR with a straightforward yet impactful spatial pooling operation. Additionally, we incorporate of Squeeze-and-Excitation (SE) blocks after each MetaSwin block of the encoder and into the decoder, which aims at segmentation performance. We evaluate our proposed MetaSwin model on two distinct medical datasets, namely BraTS 2023 and MICCAI 2015 BTCV, and conduct a comprehensive comparison with the two baselines, i.e., SwinUNETR and SwinUNETR+SE models. Our results emphasize the effectiveness of MetaSwin, showcasing its competitive edge against the baselines, utilizing a simple pooling operation and efficient SE blocks. MetaSwin's consistent and superior performance on the BTCV dataset, in comparison to SwinUNETR, is particularly significant. For instance, with a model size of 24, MetaSwin outperforms SwinUNETR's 76.58% Dice score using fewer parameters (15,407,384 vs 15,703,304) and a substantially reduced training time (300 vs 467 mins), achieving an improved Dice score of 79.12%. This research highlights the essential contribution of a simplified transformer framework, incorporating basic elements such as pooling and SE blocks, thus emphasizing their potential to guide the progression of medical segmentation models, without relying on complex attention-based mechanisms.

PGDS-YOLOv8s: An Improved YOLOv8s Model for Object Detection in Fisheye Images

Recently, object detection has become a research hotspot in computer vision, which often detects regular images with small viewing angles. In order to obtain a field of view without blind spots, fisheye cameras, which have distortions and discontinuities, have come into use. The fisheye camera, which has a wide viewing angle, and an unmanned aerial vehicle equipped with a fisheye camera are used to obtain a field of view without blind spots. However, distorted and discontinuous objects appear in the captured fisheye images due to the unique viewing angle of fisheye cameras. It poses a significant challenge to some existing object detectors. To solve this problem, this paper proposes a PGDS-YOLOv8s model to solve the issue of detecting distorted and discontinuous objects in fisheye images. First, two novel downsampling modules are proposed. Among them, the Max Pooling and Ghost’s Downsampling (MPGD) module effectively extracts the essential feature information of distorted and discontinuous objects. The Average Pooling and Ghost’s Downsampling (APGD) module acquires rich global features and reduces the feature loss of distorted and discontinuous objects. In addition, the proposed C2fs module uses Squeeze-and-Excitation (SE) blocks to model the interdependence of the channels to acquire richer gradient flow information about the features. The C2fs module provides a better understanding of the contextual information in fisheye images. Subsequently, an SE block is added after the Spatial Pyramid Pooling Fast (SPPF), thus improving the model’s ability to capture features of distorted, discontinuous objects. Moreover, the UAV-360 dataset is created for object detection in fisheye images. Finally, experiments show that the proposed PGDS-YOLOv8s model on the VOC-360 dataset improves mAP@0.5 by 19.8% and mAP@0.5:0.95 by 27.5% compared to the original YOLOv8s model. In addition, the improved model on the UAV-360 dataset achieves 89.0% for mAP@0.5 and 60.5% for mAP@0.5:0.95. Furthermore, on the MS-COCO 2017 dataset, the PGDS-YOLOv8s model improved AP by 1.4%, AP50 by 1.7%, and AP75 by 1.2% compared with the original YOLOv8s model.

Comparison and analysis of deep neural networks in facial expression recognition

Recognizing facial expressions automatically is of importance to interactive computer systems since facial expression is an efficient means of conveying emotions. Over the past few years, many researchers have attempted to use deep learning for expression recognition. The advantage of deep learning lies in its ability to learn features from datasets automatically, without relying on hand-crafted features. The paper analyzes what mechanism is useful for expression recognition in deep learning by comparing the performance of different popular models and algorithms from recent research. The paper trains the models on the Facial Expression Recognition Dataset (FER-2013), which is a relatively small and imbalanced dataset. After that, model performance is assessed on the private test dataset of FER-2013. Specifically, Residual Network (ResNet), Visual Geometry Group Network (VGGNet), and MobileNet are evaluated in the experiment. The evaluation is based on running time, the number of parameters, and the accuracy. Squeeze and Excitation (SE) block is utilized in the ResNet, which enables the models to learn useful features from global information. In the paper, ResNet34 inserted with SE block (SE-ResNet34) get the highest private test accuracy of 70.80%. Experimental results show that the residual learning enables models to go deeper without degradation and the SE block is beneficial for the model to learn global information.

A NEW INTELLIGENT MODEL BASED ON IMPROVED INCEPTION-V3 FOR ORAL CANCER AND CYST CLASSIFICATION

Oral cancer, which is also called mouth cancer, is cancer of the lining of the mouth, lips, or upper throat that has appeared in more than 355,000 people worldwide and caused more than 177,000 deaths, so it is essential to diagnose it as early as possible. Computed tomography (CT) scan is conducive to oral cancer diagnosis, but classifying oral CT images to cancer and cyst manually is difficult and time-consuming. A novel intelligent model based on improved Inception-v3 for classifying oral cancer and cyst CT images automatically is proposed in this paper. We replace the conventional convolution block in Inception-v3 with the Inverted Bottleneck Block and introduce Squeeze-and-Excitation Block (SEB) and Convolutional Block Attention Block (CBAB). The proposed model in this paper is trained on a dataset consisting of CT images of two classes (oral cancer and cyst), and the proposed model achieves 84.053% accuracy, 82.364% sensitivity, 84.508% specificity for oral cancer classification and outperforms other common models in classifying oral CT images.

Decoding motor imagery based on dipole feature imaging and a hybrid CNN with embedded squeeze-and-excitation block

Motor imagery (MI) decoding is the core of an intelligent rehabilitation system in brain computer interface, and it has a potential advantage by using source signals, which have higher spatial resolution and the same time resolution compared to scalp electroencephalography (EEG). However, how to delve and utilize the personalized frequency characteristic of dipoles for improving decoding performance has not been paid sufficient attention. In this paper, a novel dipole feature imaging (DFI) and a hybrid convolutional neural network (HCNN) with an embedded squeeze-and-excitation block (SEB), denoted as DFI-HCNN, are proposed for decoding MI tasks. EEG source imaging technique is used for brain source estimation, and each sub-band spectrum powers of all dipoles are calculated through frequency analysis and band division. Then, the 3D space information of dipoles is retrieved, and by using azimuthal equidistant projection algorithm it is transformed to a 2D plane, which is combined with nearest neighbor interpolation to generate multi sub-band dipole feature images. Furthermore, a HCNN is designed and applied to the ensemble of sub-band dipole feature images, from which the importance of sub-bands is acquired to adjust the corresponding attentions adaptively by SEB. Ten-fold cross-validation experiments on two public datasets achieve the comparatively higher decoding accuracies of 84.23% and 92.62%, respectively. The experiment results show that DFI is an effective feature representation, and HCNN with an embedded SEB can enhance the useful frequency information of dipoles for improving MI decoding.

SE_SPnet: Rice leaf disease prediction using stacked parallel convolutional neural network with squeeze‐and‐excitation

AbstractRice is one of the significant crops, and the early identification and prevention of its diseases are essential to ensure adequate and healthy availability to the world's growing population. The use of image processing is an encouraging method for automatic rice leaf disease identification and detection. In particular, the recent advancements indicate the effectiveness of convolutional neural network (CNN) based deep learning approaches. In this direction, the present work proposes a novel stacked parallel convolution layers‐based network (SPnet) with the squeeze‐and‐excitation (SE) architecture, named (SE_SPnet), for classifying diseased rice leaf images. The stacked parallel network block comprises four parallel convolution layers with different kernel sizes for abstractions of the global and local features. The SE block extracts feature information automatically while removing invalid ones. We compare the SE_SPnet model with state‐of‐the‐art CNN models such as VGG16, DenseNet121, and InceptionV3 based on computational effort, accuracy, sensitivity, specificity, precision, recall, and F1‐score. The experimental results show that the SE_SPnet outperforms standard CNN models for the considered rice leaf disease image datasets. In particular, the SE_SPnet achieves the highest accuracy (99.2%), sensitivity (98.2%), specificity (98.5%), precision (98.4%), recall (98.2%), and F1‐score (98.5%) while using stochastic gradient descent (with momentum) optimizer with a 0.01 learning rate. Furthermore, the SE_SPnet also exhibits to outperform when compared with some of the most recent and relevant existing works.

Permeability prediction of considering organic matter distribution based on deep learning

At present, researchers predict permeability through core experiments that require specific experimental conditions and methods, which are difficult and time-consuming. Conventional simulation methods for predicting permeability require considerable computational resources. Therefore, deep learning can be used as a pore-scale simulation prediction method. In this study, we established a workflow for directly predicting permeability from images. Considering that the mineral properties of the nanopore wall of shale oil have a large influence on the flow, a core dataset with organic distribution was constructed with random circles, and the slip influence of organic pores was considered. From our dataset, we found that the average permeability with organic distribution was 32.3% higher than that without organic distribution. Therefore, to simulate the microscopic flow and predict the permeability of shale oil, considering the differences in the pore flow mechanisms of different minerals is necessary. We designed a convolutional network for the dataset, adopted the structure of SE-ResNet, added the squeeze-and-excitation (SE) module to the double-layer residual module of ResNet18, and combined the characteristics of the SE block with the attention mechanism and ResNet to effectively obtain the information between channels and avoid the problem of gradient disappearance or explosion. Using SE-ResNet for directly predicting the apparent permeability from images, the accuracy of the test set reached 88.5%. The model had strong generalization ability, and the SE-ResNet could map the image of the core to the apparent permeability, which was approximately 100 times faster than the direct flow simulation.

A Mental Workload Classification Method Based on GCN Modified by Squeeze-and-Excitation Residual

In some complex labor production and human–machine interactions, such as subway driving, to ensure both the efficient and rapid completion of work and the personal safety of staff and the integrity of operating equipment, the level of mental workload (MW) of operators is monitored at all times. In existing machine learning-based MW classification methods, the association information between neurons in different regions is almost not considered. To solve the above problem, a graph convolution network based on the squeeze-and-excitation (SE) block is proposed. For a raw electroencephalogram (EEG) signal, the principal component analysis (PCA) dimensionality reduction operation is carried out. After that, combined with the spatial distribution between brain electrodes, the dimensionality reduction data can be converted to graph structure data, carrying association information between neurons in different regions. In addition, we use graph convolution neural network (GCN) modified by SE residual to obtain final classification results. Here, to adaptively recalibrate channel-wise feature responses by explicitly modelling interdependencies between channels, the SE block is introduced. The residual connection can ease the training of networks. To discuss the performance of the proposed method, we carry out some experiments using the raw EEG signals of 10 healthy subjects, which are collected using the MATB-II platform based on multi-task aerial context manipulation. From the experiment results, the structural reasonableness and the performance superiority of the proposed method are verified. In short, the proposed GCN modified by the SE residual method is a workable plan of mental workload classification.

An n-Sigmoid Activation Function to Improve the Squeeze-and-Excitation for 2D and 3D Deep Networks

The Squeeze-and-Excitation (SE) structure has been designed to enhance the neural network performance by allowing it to execute positive channel-wise feature recalibration and suppress less useful features. SE structures are generally adopted in a plethora of tasks directly in existing models and have shown actual performance enhancements. However, the various sigmoid functions used in artificial neural networks are intrinsically restricted by vanishing gradients. The purpose of this paper is to further improve the network by introducing a new SE block with a custom activation function resulting from the integration of a piecewise shifted sigmoid function. The proposed activation function aims to improve the learning and generalization capacity of 2D and 3D neural networks for classification and segmentation, by reducing the vanishing gradient problem. Comparisons were made between the networks with the original design, the addition of the SE block, and the proposed n-sigmoid SE block. To evaluate the performance of this new method, commonly used datasets, CIFAR-10 and Carvana for 2D data and Sandstone Dataset for 3D data, were considered. Experiments conducted using SE showed that the new n-sigmoid function results in performance improvements in the training accuracy score for UNet (up 0.25% to 99.67%), ResNet (up 0.9% to 95.1%), and DenseNet (up 1.1% to 98.87%) for the 2D cases, and the 3D UNet (up 0.2% to 99.67%) for the 3D cases. The n-sigmoid SE block not only reduces the vanishing gradient problem but also develops valuable features by combining channel-wise and spatial information.

Discriminative kernel convolution network for multi-label ophthalmic disease detection on imbalanced fundus image dataset

It is feasible to recognize the presence and seriousness of eye disease by investigating the progressions in retinal biological structures. Fundus examination is a diagnostic procedure to examine the biological structure and anomalies present in the eye. Ophthalmic diseases like glaucoma, diabetic retinopathy, and cataracts are the main cause of visual impairment worldwide. Ocular Disease Intelligent Recognition (ODIR-5K) is a benchmark structured fundus image dataset utilized by researchers for multi-label multi-disease classification of fundus images. This work presents a Discriminative Kernel Convolution Network (DKCNet), which explores discriminative region-wise features without adding extra computational cost. DKCNet is composed of an attention block followed by a Squeeze-and-Excitation (SE) block. The attention block takes features from the backbone network and generates discriminative feature attention maps. The SE block takes the discriminative feature maps and improves channel interdependencies. Better performance of DKCNet is observed with InceptionResnet backbone network for multi-label classification of ODIR-5K fundus images with 96.08 AUC, 94.28 F1-score, and 0.81 kappa score. The proposed method splits the common target label for an eye pair based on the diagnostic keyword. Based on these labels, over-sampling and/or under-sampling are done to resolve the class imbalance. To check the bias of the proposed model towards training data, the model trained on the ODIR dataset is tested on three publicly available benchmark datasets. It is observed that the proposed DKCNet gives good performance on completely unseen fundus images also.

MS-HNN: Multi-Scale Hierarchical Neural Network With Squeeze and Excitation Block for Neonatal Sleep Staging Using a Single-Channel EEG.

Most existing neonatal sleep staging appro- aches applied multiple EEG channels to obtain good performance. However, it potentially increased the computational complexity and led to an increased risk of skin disruption to neonates during data acquisition. In this paper, a multi-scale hierarchical neural network (MS-HNN) with a squeeze and excitation (SE) block for neonatal sleep staging is presented in this study on the basis of a single EEG channel. MS-HNN composes of multi-scale convolutional neural network (MSCNN), temporal information learning (TIL) module, and squeeze and excitation (SE) block. MSCNN can extract features from different scales and frequencies, and TIL module can acquire the transition information among adjacent stages. In addition, for these extracted features, SE block can selectively concentrate on informative features and weaken redundant features for achieving better performance. The proposed approach was validated on a clinical dataset involving 64 neonates from the Children's Hospital of Fudan University (CHFU). The proposed network achieves an accuracy of 75.4% and 76.5% for three-class automatic neonatal sleep staging with the single-EEG channel and the eight-EEG channels, respectively. The experimental results show that the proposed method can maintain good performance by making full use of the information in the single channel while reducing the channels to control the computational overhead.

MfeeU-Net: A multi-scale feature extraction and enhancement U-Net for automatic liver segmentation from CT Images.

Medical image segmentation of the liver is an important prerequisite for clinical diagnosis and evaluation of liver cancer. For automatic liver segmentation from Computed Tomography (CT) images, we proposed a Multi-scale Feature Extraction and Enhancement U-Net (mfeeU-Net), incorporating Res2Net blocks, Squeeze-and-Excitation (SE) blocks, and Edge Attention (EA) blocks. The Res2Net blocks which are conducive to extracting multi-scale features of the liver were used as the backbone of the encoder, while the SE blocks were also added to the encoder to enhance channel information. The EA blocks were introduced to skip connections between the encoder and the decoder, to facilitate the detection of blurred liver edges where the intensities of nearby organs are close to the liver. The proposed mfeeU-Net was trained and evaluated using a publicly available CT dataset of LiTS2017. The average dice similarity coefficient, intersection-over-union ratio, and sensitivity of the mfeeU-Net for liver segmentation were 95.32%, 91.67%, and 95.53%, respectively, and all these metrics were better than those of U-Net, Res-U-Net, and Attention U-Net. The experimental results demonstrate that the mfeeU-Net can compete with and even outperform recently proposed convolutional neural networks and effectively overcome challenges, such as discontinuous liver regions and fuzzy liver boundaries.

Citrus Tree Crown Segmentation of Orchard Spraying Robot Based on RGB-D Image and Improved Mask R-CNN

Orchard spraying robots must visually obtain citrus tree crown growth information to meet the variable growth-stage-based spraying requirements. However, the complex environments and growth characteristics of fruit trees affect the accuracy of crown segmentation. Therefore, we propose a feature-map-based squeeze-and-excitation UNet++ (MSEU) region-based convolutional neural network (R-CNN) citrus tree crown segmentation method that intakes red–green–blue-depth (RGB-D) images that are pixel aligned and visual distance-adjusted to eliminate noise. Our MSEU R-CNN achieves accurate crown segmentation using squeeze-and-excitation (SE) and UNet++. To fully fuse the feature map information, the SE block correlates image features and recalibrates their channel weights, and the UNet++ semantic segmentation branch replaces the original mask structure to maximize the interconnectivity between feature layers, achieving a near-real time detection speed of 5 fps. Its bounding box (bbox) and segmentation (seg) AP50 scores are 96.6 and 96.2%, respectively, and the bbox average recall and F1-score are 73.0 and 69.4%, which are 3.4, 2.4, 4.9, and 3.5% higher than the original model, respectively. Compared with bbox instant segmentation (BoxInst) and conditional convolutional frameworks (CondInst), the MSEU R-CNN provides better seg accuracy and speed than the previous-best Mask R-CNN. These results provide the means to accurately employ autonomous spraying robots.

DMSENet: Deep multi-modal squeeze and excitation network for the diagnosis of Alzheimer's disease

ABSTRACT Alzheimer's Disease (AD) is an incurable neurodegenerative disorder that affects millions of older people worldwide. Compared to conventional methods, neuroimaging modalities along with the machine learning techniques detect the onset of AD more successfully. It has been established that multimodal classification provides better accuracy than single modal classification. Exploring the synergy between several multimodal neuroimages remains challenging due to the lack of available fusion techniques. The proposed Deep Multimodal Squeeze and Excitation network (DMSENet) uses the ResNet Squeeze and Excitation (SE) block to extract relevant features from MRI and PET images. Using the hierarchical fusion method, the extracted features are fused; subsequently, using the fused Feature Map (FM), the ResNet SE block retrieves additional higher-level and lower-level features. Hierarchical fusion methodology ensures the efficiency of Multimodal Fusion (MMF); the Attention Model(AM) then assigns the fusion ratio automatically by prioritizing the multimodal data. Moreover, the depth and efficiency of the attention network are ensured by the combination of identity mapping and Residual Block. In the DMSENet, both higher-level and lower-level features could be utilized simultaneously by employing both early and late fusion methodologies. The suggested framework is investigated using the ADNI dataset, providing greater precision than state-of-the-art methods.

Identifying rumen protozoa in microscopic images of ruminant with improved YOLACT instance segmentation

Identification of rumen protozoa in ruminants is significant for species identification, ecological population structure survey, and protozoa behaviour analysis. At present, rumen protozoa identification still needs to be done manually, which is time-consuming and inefficient. To address this issue, this paper proposes a deep learning method for accurate segmentation and recognition of protozoa instances in rumen microscopic images for the first time. Firstly, a microscopic image dataset of protozoa was constructed, which consists of 2671 images of 17 species, 11 genera, and 2 orders, and the images were annotated in detail. Secondly, by comparing two instance segmentation models, an improved YOLACT was presented for efficient protozoan image segmentation. In the proposed method, the SE (Squeeze-and-Excitation) block is integrated into the shallowest layer of the backbone network to enhance the expression of features, and the ReLU activation function of all layers is replaced by FReLU activation function to eliminate the spatial insensitivity of ReLU activation function. The experimental results show that the segmentation accuracy of the improved YOLACT method is 89.45%, which is 0.55% higher than that of the original YOLACT and 2.63% higher than that of the Mask R-CNN method. Meanwhile, the detection speed of YOLACT is 1.7 times that of Mask R-CNN. This work reduces the labour cost of rumen protozoa identification and provides a new method for accurate identification of the microscopic images of protozoa. Besides, the microscopic image dataset of protozoa and the prediction models constructed in this paper will be publicly available. • We use deep neural network to recognise rumen protozoa in microscopic image. • YOLACT outperforms Mask R-CNN in the accuracy of protozoa segmentation. • Improved YOLACT among all methods achieves the best performance. • This work provides a technical reference to the related protozoa research.

Land-Sea Target Detection and Recognition in SAR Image Based on Non-Local Channel Attention Network

Synthetic aperture radar (SAR) target recognition is essential for SAR image interpretation. It has been widely used in national defense and national economy. At present, the SAR image detection and recognition methods based on convolutional neural network (CNN) have problems such as insufficient extraction of the feature information of SAR image targets, false targets caused by the interference of complex backgrounds, and low detection performance. The main reason is that the feature extraction of CNN is a local operation in space and time, which ignores the correlation between pixels and regions and the dependencies between channels in SAR images. In this paper, a non-local channel attention network (NLCANet) SAR image target recognition method is proposed based on the GoogLeNet structure combined with asymmetric pyramid non-local block (APNB) and squeeze-and-excitation block (SEB). APNB is added to the GoogLeNet framework to capture more context information and enhance the correlation between pixels and regions. SEB is added to the Inception structure to become Inception-SEB (ISEB), through which channel dependencies based on the fusion of different scale features can be obtained. The experimental results based on the moving and stationary target acquisition and recognition (MSTAR) dataset and the SAR ship detection dataset (SSDD) show that the proposed method improves the detection ability of targets in complex backgrounds and achieves better land-sea target recognition performance.

RAFF-Net: An improved tongue segmentation algorithm based on residual attention network and multiscale feature fusion.

Due to the complexity of face images, tongue segmentation is susceptible to interference from uneven tongue texture, lips and face, resulting in traditional methods failing to segment the tongue accurately. To address this problem, RAFF-Net, an automatic tongue region segmentation network based on residual attention network and multiscale feature fusion, was proposed. It aims to improve tongue segmentation accuracy and achieve end-to-end automated segmentation. Based on the UNet backbone network, different numbers of ResBlocks combined with the Squeeze-and-Excitation (SE) block was used as an encoder to extract image layered features. The decoder structure of UNet was simplified and the number of parameters of the network model was reduced. Meanwhile, the multiscale feature fusion module was designed to optimize the network parameters by combining a custom loss function instead of the common cross-entropy loss function to further improve the detection accuracy. The RAFF-Net network structure achieved Mean Intersection over Union (MIoU) and F1-score of 97.85% and 97.73%, respectively, which improved 0.56% and 0.46%, respectively, compared with the original UNet; ablation experiments demonstrated that the improved algorithm could contribute to the enhancement of tongue segmentation effect. This study combined the residual attention network with multiscale feature fusion to effectively improve the segmentation accuracy of the tongue region, and optimized the input and output of the UNet network using different numbers of ResBlocks, SE block, multiscale feature fusion and weighted loss function, increased the stability of the network and improved the overall effect of the network.

Detecting Small Anatomical Structures in 3D Knee MRI Segmentation by Fully Convolutional Networks

Accurately identifying the pixels of small organs or lesions from magnetic resonance imaging (MRI) has a critical impact on clinical diagnosis. U-net is the most well-known and commonly used neural network for image segmentation. However, the small anatomical structures in medical images cannot be well recognised by U-net. This paper explores the performance of the U-net architectures in knee MRI segmentation to find a relative structure that can obtain high accuracies for both small and large anatomical structures. To maximise the utilities of U-net architecture, we apply three types of components, residual blocks, squeeze-and-excitation (SE) blocks, and dense blocks, to construct four variants of U-net, namely U-net variants. Among these variants, our experiments show that SE blocks can improve the segmentation accuracies of small labels. We adopt DeepLabv3plus architecture for 3D medical image segmentation by equipping SE blocks based on this discovery. The experimental results show that U-net with SE block achieves higher accuracy in parts of small anatomical structures. In contrast, DeepLabv3plus with SE block performs better on the average dice coefficient of small and large labels.