The accurate segmentation of breast ultrasound images is an important precondition for the lesion determination. The existing segmentation approaches embrace massive parameters, sluggish inference speed, and huge memory consumption. To tackle this problem, we propose T2KD Attention U-Net (dual-Teacher Knowledge Distillation Attention U-Net), a lightweight semantic segmentation method combined double-path joint distillation in breast ultrasound images. Primarily, we designed two teacher models to learn the fine-grained features from each class of images according to different feature representation and semantic information of benign and malignant breast lesions. Then we leveraged the joint distillation to train a lightweight student model. Finally, we constructed a novel weight balance loss to focus on the semantic feature of small objection, solving the unbalance problem of tumor and background. Specifically, the extensive experiments conducted on Dataset BUSI and Dataset B demonstrated that the T2KD Attention U-Net outperformed various knowledge distillation counterparts. Concretely, the accuracy, recall, precision, Dice, and mIoU of proposed method were 95.26%, 86.23%, 85.09%, 83.59%and 77.78% on Dataset BUSI, respectively. And these performance indexes were 97.95%, 92.80%, 88.33%, 88.40% and 82.42% on Dataset B, respectively. Compared with other models, the performance of this model was significantly improved. Meanwhile, compared with the teacher model, the number, size, and complexity of student model were significantly reduced (2.2×106 vs. 106.1×106, 8.4 MB vs. 414 MB, 16.59 GFLOPs vs. 205.98 GFLOPs, respectively). Indeedy, the proposed model guarantees the performances while greatly decreasing the amount of computation, which provides a new method for the deployment of clinical medical scenarios.
Lung cancer is the most threatening tumor disease to human health. Early detection is crucial to improve the survival rate and recovery rate of lung cancer patients. Existing methods use the two-dimensional multi-view framework to learn lung nodules features and simply integrate multi-view features to achieve the classification of benign and malignant lung nodules. However, these methods suffer from the problems of not capturing the spatial features effectively and ignoring the variability of multi-views. Therefore, this paper proposes a three-dimensional (3D) multi-view convolutional neural network (MVCNN) framework. To further solve the problem of different views in the multi-view model, a 3D multi-view squeeze-and-excitation convolution neural network (MVSECNN) model is constructed by introducing the squeeze-and-excitation (SE) module in the feature fusion stage. Finally, statistical methods are used to analyze model predictions and doctor annotations. In the independent test set, the classification accuracy and sensitivity of the model were 96.04% and 98.59% respectively, which were higher than other state-of-the-art methods. The consistency score between the predictions of the model and the pathological diagnosis results was 0.948, which is significantly higher than that between the doctor annotations and the pathological diagnosis results. The methods presented in this paper can effectively learn the spatial heterogeneity of lung nodules and solve the problem of multi-view differences. At the same time, the classification of benign and malignant lung nodules can be achieved, which is of great significance for assisting doctors in clinical diagnosis.
Glioma is a primary brain tumor with high incidence rate. High-grade gliomas (HGG) are those with the highest degree of malignancy and the lowest degree of survival. Surgical resection and postoperative adjuvant chemoradiotherapy are often used in clinical treatment, so accurate segmentation of tumor-related areas is of great significance for the treatment of patients. In order to improve the segmentation accuracy of HGG, this paper proposes a multi-modal glioma semantic segmentation network with multi-scale feature extraction and multi-attention fusion mechanism. The main contributions are, (1) Multi-scale residual structures were used to extract features from multi-modal gliomas magnetic resonance imaging (MRI); (2) Two types of attention modules were used for features aggregating in channel and spatial; (3) In order to improve the segmentation performance of the whole network, the branch classifier was constructed using ensemble learning strategy to adjust and correct the classification results of the backbone classifier. The experimental results showed that the Dice coefficient values of the proposed segmentation method in this article were 0.909 7, 0.877 3 and 0.839 6 for whole tumor, tumor core and enhanced tumor respectively, and the segmentation results had good boundary continuity in the three-dimensional direction. Therefore, the proposed semantic segmentation network has good segmentation performance for high-grade gliomas lesions.
Speech imagery is an emerging brain-computer interface (BCI) paradigm with potential to provide effective communication for individuals with speech impairments. This study designed a Chinese speech imagery paradigm using three clinically relevant words—“Help me”, “Sit up” and “Turn over”—and collected electroencephalography (EEG) data from 15 healthy subjects. Based on the data, a Channel Attention Multi-Scale Convolutional Neural Network (CAM-Net) decoding algorithm was proposed, which combined multi-scale temporal convolutions with asymmetric spatial convolutions to extract multidimensional EEG features, and incorporated a channel attention mechanism along with a bidirectional long short-term memory network to perform channel weighting and capture temporal dependencies. Experimental results showed that CAM-Net achieved a classification accuracy of 48.54% in the three-class task, outperforming baseline models such as EEGNet and Deep ConvNet, and reached a highest accuracy of 64.17% in the binary classification between “Sit up” and “Turn over”. This work provides a promising approach for future Chinese speech imagery BCI research and applications.
In response to the issues of single-scale information loss and large model parameter size during the sampling process in U-Net and its variants for medical image segmentation, this paper proposes a multi-scale medical image segmentation method based on pixel encoding and spatial attention. Firstly, by redesigning the input strategy of the Transformer structure, a pixel encoding module is introduced to enable the model to extract global semantic information from multi-scale image features, obtaining richer feature information. Additionally, deformable convolutions are incorporated into the Transformer module to accelerate convergence speed and improve module performance. Secondly, a spatial attention module with residual connections is introduced to allow the model to focus on the foreground information of the fused feature maps. Finally, through ablation experiments, the network is lightweighted to enhance segmentation accuracy and accelerate model convergence. The proposed algorithm achieves satisfactory results on the Synapse dataset, an official public dataset for multi-organ segmentation provided by the International Conference on Medical Image Computing and Computer Assisted Intervention (MICCAI), with Dice similarity coefficient (DSC) and 95% Hausdorff distance (HD95) scores of 77.65 and 18.34, respectively. The experimental results demonstrate that the proposed algorithm can enhance multi-organ segmentation performance, potentially filling the gap in multi-scale medical image segmentation algorithms, and providing assistance for professional physicians in diagnosis.
Accurate segmentation of ground glass nodule (GGN) is important in clinical. But it is a tough work to segment the GGN, as the GGN in the computed tomography images show blur boundary, irregular shape, and uneven intensity. This paper aims to segment GGN by proposing a fully convolutional residual network, i.e., residual network based on atrous spatial pyramid pooling structure and attention mechanism (ResAANet). The network uses atrous spatial pyramid pooling (ASPP) structure to expand the feature map receptive field and extract more sufficient features, and utilizes attention mechanism, residual connection, long skip connection to fully retain sensitive features, which is extracted by the convolutional layer. First, we employ 565 GGN provided by Shanghai Chest Hospital to train and validate ResAANet, so as to obtain a stable model. Then, two groups of data selected from clinical examinations (84 GGN) and lung image database consortium (LIDC) dataset (145 GGN) were employed to validate and evaluate the performance of the proposed method. Finally, we apply the best threshold method to remove false positive regions and obtain optimized results. The average dice similarity coefficient (DSC) of the proposed algorithm on the clinical dataset and LIDC dataset reached 83.46%, 83.26% respectively, the average Jaccard index (IoU) reached 72.39%, 71.56% respectively, and the speed of segmentation reached 0.1 seconds per image. Comparing with other reported methods, our new method could segment GGN accurately, quickly and robustly. It could provide doctors with important information such as nodule size or density, which assist doctors in subsequent diagnosis and treatment.
Objective To explore the white matter microstructural abnormalities in patients with different subtypes of attention-deficit/hyperactivity disorder (ADHD) and establish a diagnostic classification model. Methods Patients with ADHD admitted to West China Hospital of Sichuan University between January 2019 and September 2021 and healthy controls recruited through advertisement were prospectively selected. All participants underwent diffusion tensor imaging scanning. The whole brain voxel-based analysis was used to compare the diffusion parameter maps of fractional anisotropy (FA) among patients with combined subtype of ADHD (ADHD-C), patients with inattentive subtype of ADHD (ADHD-I) and healthy controls. The support vector machine classifier and feature selection method were used to construct the individual ADHD diagnostic classification model and efficiency was evaluated between each two groups of the ADHD patients and healthy controls. Results A total of 26 ADHD-C patients, 24 ADHD-I patients and 26 healthy controls were included. The three groups showed significant differences in FA values in the bilateral sagittal stratum of temporal lobe (ADHD-C<ADHD-I<healthy controls) and the isthmus of corpus callosum (ADHD-C>ADHD-I>healthy controls) (P<0.005). The direct comparison between the two subtypes of ADHD showed that ADHD-C had higher FA than ADHD-I in the right middle frontal gyrus. The classification model differentiating ADHD-C and ADHD-I showed the highest efficiency, with a total accuracy of 76.0%, sensitivity of 88.5%, and specificity of 70.8%. Conclusions There is both commonality and heterogeneity in white matter microstructural alterations in the two subtypes of patients with ADHD. The white matter damage of the sagittal stratum of temporal lobe and the corpus callosum may be the intrinsic pathophysiological basis of ADHD, while the anomalies of frontal brain region may be the differential point between different subtypes of patients.
The processing mechanism of the human brain for speech information is a significant source of inspiration for the study of speech enhancement technology. Attention and lateral inhibition are key mechanisms in auditory information processing that can selectively enhance specific information. Building on this, the study introduces a dual-branch U-Net that integrates lateral inhibition and feedback-driven attention mechanisms. Noisy speech signals input into the first branch of the U-Net led to the selective feedback of time-frequency units with high confidence. The generated activation layer gradients, in conjunction with the lateral inhibition mechanism, were utilized to calculate attention maps. These maps were then concatenated to the second branch of the U-Net, directing the network’s focus and achieving selective enhancement of auditory speech signals. The evaluation of the speech enhancement effect was conducted by utilising five metrics, including perceptual evaluation of speech quality. This method was compared horizontally with five other methods: Wiener, SEGAN, PHASEN, Demucs and GRN. The experimental results demonstrated that the proposed method improved speech signal enhancement capabilities in various noise scenarios by 18% to 21% compared to the baseline network across multiple performance metrics. This improvement was particularly notable in low signal-to-noise ratio conditions, where the proposed method exhibited a significant performance advantage over other methods. The speech enhancement technique based on lateral inhibition and feedback-driven attention mechanisms holds significant potential in auditory speech enhancement, making it suitable for clinical practices related to artificial cochleae and hearing aids.
Manual segmentation of coronary arteries in computed tomography angiography (CTA) images is inefficient, and existing deep learning segmentation models often exhibit low accuracy on coronary artery images. Inspired by the Transformer architecture, this paper proposes a novel segmentation model, the double parallel encoder u-net with transformers (DUNETR). This network employed a dual-encoder design integrating Transformers and convolutional neural networks (CNNs). The Transformer encoder transformed three-dimensional (3D) coronary artery data into a one-dimensional (1D) sequential problem, effectively capturing global multi-scale feature information. Meanwhile, the CNN encoder extracted local features of the 3D coronary arteries. The complementary features extracted by the two encoders were fused through the noise reduction feature fusion (NRFF) module and passed to the decoder. Experimental results on a public dataset demonstrated that the proposed DUNETR model achieved a Dice similarity coefficient of 81.19% and a recall rate of 80.18%, representing improvements of 0.49% and 0.46%, respectively, over the next best model in comparative experiments. These results surpassed those of other conventional deep learning methods. The integration of Transformers and CNNs as dual encoders enables the extraction of rich feature information, significantly enhancing the effectiveness of 3D coronary artery segmentation. Additionally, this model provides a novel approach for segmenting other vascular structures.
The conventional fault diagnosis of patient monitors heavily relies on manual experience, resulting in low diagnostic efficiency and ineffective utilization of fault maintenance text data. To address these issues, this paper proposes an intelligent fault diagnosis method for patient monitors based on multi-feature text representation, improved bidirectional gate recurrent unit (BiGRU) and attention mechanism. Firstly, the fault text data was preprocessed, and the word vectors containing multiple linguistic features was generated by linguistically-motivated bidirectional encoder representation from Transformer. Then, the bidirectional fault features were extracted and weighted by the improved BiGRU and attention mechanism respectively. Finally, the weighted loss function is used to reduce the impact of class imbalance on the model. To validate the effectiveness of the proposed method, this paper uses the patient monitor fault dataset for verification, and the macro F1 value has achieved 91.11%. The results show that the model built in this study can realize the automatic classification of fault text, and may provide assistant decision support for the intelligent fault diagnosis of the patient monitor in the future.
