Objective To develop a deep learning system for CT images to assist in the diagnosis of thoracolumbar fractures and analyze the feasibility of its clinical application. Methods Collected from West China Hospital of Sichuan University from January 2019 to March 2020, a total of 1256 CT images of thoracolumbar fractures were annotated with a unified standard through the Imaging LabelImg system. All CT images were classified according to the AO Spine thoracolumbar spine injury classification. The deep learning system in diagnosing ABC fracture types was optimized using 1039 CT images for training and validation, of which 1004 were used as the training set and 35 as the validation set; the rest 217 CT images were used as the test set to compare the deep learning system with the clinician’s diagnosis. The deep learning system in subtyping A was optimized using 581 CT images for training and validation, of which 556 were used as the training set and 25 as the validation set; the rest 104 CT images were used as the test set to compare the deep learning system with the clinician’s diagnosis. Results The accuracy and Kappa coefficient of the deep learning system in diagnosing ABC fracture types were 89.4% and 0.849 (P<0.001), respectively. The accuracy and Kappa coefficient of subtyping A were 87.5% and 0.817 (P<0.001), respectively. Conclusions The classification accuracy of the deep learning system for thoracolumbar fractures is high. This approach can be used to assist in the intelligent diagnosis of CT images of thoracolumbar fractures and improve the current manual and complex diagnostic process.
Magnetic resonance imaging(MRI) can obtain multi-modal images with different contrast, which provides rich information for clinical diagnosis. However, some contrast images are not scanned or the quality of the acquired images cannot meet the diagnostic requirements due to the difficulty of patient's cooperation or the limitation of scanning conditions. Image synthesis techniques have become a method to compensate for such image deficiencies. In recent years, deep learning has been widely used in the field of MRI synthesis. In this paper, a synthesis network based on multi-modal fusion is proposed, which firstly uses a feature encoder to encode the features of multiple unimodal images separately, and then fuses the features of different modal images through a feature fusion module, and finally generates the target modal image. The similarity measure between the target image and the predicted image in the network is improved by introducing a dynamic weighted combined loss function based on the spatial domain and K-space domain. After experimental validation and quantitative comparison, the multi-modal fusion deep learning network proposed in this paper can effectively synthesize high-quality MRI fluid-attenuated inversion recovery (FLAIR) images. In summary, the method proposed in this paper can reduce MRI scanning time of the patient, as well as solve the clinical problem of missing FLAIR images or image quality that is difficult to meet diagnostic requirements.
In the extraction of fetal electrocardiogram (ECG) signal, due to the unicity of the scale of the U-Net same-level convolution encoder, the size and shape difference of the ECG characteristic wave between mother and fetus are ignored, and the time information of ECG signals is not used in the threshold learning process of the encoder’s residual shrinkage module. In this paper, a method of extracting fetal ECG signal based on multi-scale residual shrinkage U-Net model is proposed. First, the Inception and time domain attention were introduced into the residual shrinkage module to enhance the multi-scale feature extraction ability of the same level convolution encoder and the utilization of the time domain information of fetal ECG signal. In order to maintain more local details of ECG waveform, the maximum pooling in U-Net was replaced by Softpool. Finally, the decoder composed of the residual module and up-sampling gradually generated fetal ECG signals. In this paper, clinical ECG signals were used for experiments. The final results showed that compared with other fetal ECG extraction algorithms, the method proposed in this paper could extract clearer fetal ECG signals. The sensitivity, positive predictive value, and F1 scores in the 2013 competition data set reached 93.33%, 99.36%, and 96.09%, respectively, indicating that this method can effectively extract fetal ECG signals and has certain application values for perinatal fetal health monitoring.
The diagnosis of hypertrophic cardiomyopathy (HCM) is of great significance for the early risk classification of sudden cardiac death and the screening of family genetic diseases. This research proposed a HCM automatic detection method based on convolution neural network (CNN) model, using single-lead electrocardiogram (ECG) signal as the research object. Firstly, the R-wave peak locations of single-lead ECG signal were determined, followed by the ECG signal segmentation and resample in units of heart beats, then a CNN model was built to automatically extract the deep features in the ECG signal and perform automatic classification and HCM detection. The experimental data is derived from 108 ECG records extracted from three public databases provided by PhysioNet, the database established in this research consists of 14,459 heartbeats, and each heartbeat contains 128 sampling points. The results revealed that the optimized CNN model could effectively detect HCM, the accuracy, sensitivity and specificity were 95.98%, 98.03% and 95.79% respectively. In this research, the deep learning method was introduced for the analysis of single-lead ECG of HCM patients, which could not only overcome the technical limitations of conventional detection methods based on multi-lead ECG, but also has important application value for assisting doctor in fast and convenient large-scale HCM preliminary screening.
In recent years, epileptic seizure detection based on electroencephalogram (EEG) has attracted the widespread attention of the academic. However, it is difficult to collect data from epileptic seizure, and it is easy to cause over fitting phenomenon under the condition of few training data. In order to solve this problem, this paper took the CHB-MIT epilepsy EEG dataset from Boston Children's Hospital as the research object, and applied wavelet transform for data augmentation by setting different wavelet transform scale factors. In addition, by combining deep learning, ensemble learning, transfer learning and other methods, an epilepsy detection method with high accuracy for specific epilepsy patients was proposed under the condition of insufficient learning samples. In test, the wavelet transform scale factors 2, 4 and 8 were set for experimental comparison and verification. When the wavelet scale factor was 8, the average accuracy, average sensitivity and average specificity was 95.47%, 93.89% and 96.48%, respectively. Through comparative experiments with recent relevant literatures, the advantages of the proposed method were verified. Our results might provide reference for the clinical application of epilepsy detection.
Magnetic resonance imaging (MRI) is an important medical imaging method, whose major limitation is its long scan time due to the imaging mechanism, increasing patients’ cost and waiting time for the examination. Currently, parallel imaging (PI) and compress sensing (CS) together with other reconstruction technologies have been proposed to accelerate image acquisition. However, the image quality of PI and CS depends on the image reconstruction algorithms, which is far from satisfying in respect to both the image quality and the reconstruction speed. In recent years, image reconstruction based on generative adversarial network (GAN) has become a research hotspot in the field of magnetic resonance imaging because of its excellent performance. In this review, we summarized the recent development of application of GAN in MRI reconstruction in both single- and multi-modality acceleration, hoping to provide a useful reference for interested researchers. In addition, we analyzed the characteristics and limitations of existing technologies and forecasted some development trends in this field.
ObjectiveTo study a deep learning-based dual-modality fundus camera which was used to study retinal blood oxygen saturation and vascular morphology changes in eyes with branch retinal vein occlusion (BRVO). MethodsA prospective study. From May to October 2020, 31 patients (31 eyes) of BRVO (BRVO group) and 20 healthy volunteers (20 eyes) with matched gender and age (control group) were included in the study. Among 31 patients (31 eyes) in BRVO group, 20 patients (20 eyes) received one intravitreal injection of anti-vascular endothelial growth factor drugs before, and 11 patients (11 eyes) did not receive any treatment. They were divided into treatment group and untreated group accordingly. Retinal images were collected with a dual-modality fundus camera; arterial and vein segments were segmented in the macular region of interest (MROI) using deep learning; the optical density ratio was used to calculate retinal blood oxygen saturation (SO2) on the affected and non-involved sides of the eyes in the control group and patients in the BRVO group, and calculated the diameter, curvature, fractal dimension and density of arteriovenous in MROI. Quantitative data were compared between groups using one-way analysis of variance. ResultsThere was a statistically significant difference in arterial SO2 (SO2-A) in the MROI between the affected eyes, the fellow eyes in the BRVO group and the control group (F=4.925, P<0.001), but there was no difference in the venous SO2 (SO2-V) (F=0.607, P=0.178). Compared with the control group, the SO2-A in the MROI of the affected side and the non-involved side of the untreated group was increased, and the difference was statistically significant (F=4.925, P=0.012); there was no significant difference in SO2-V (F=0.607, P=0.550). There was no significant difference in SO2-A and SO2-V in the MROI between the affected side, the non-involved side in the treatment group and the control group (F=0.159, 1.701; P=0.854, 0.197). There was no significant difference in SO2-A and SO2-V in MROI between the affected side of the treatment group, the untreated group and the control group (F=2.553, 0.265; P=0.088, 0.546). The ophthalmic artery diameter, arterial curvature, arterial fractal dimension, vein fractal dimension, arterial density, and vein density were compared in the untreated group, the treatment group, and the control group, and the differences were statistically significant (F=3.527, 3.322, 7.251, 26.128, 4.782, 5.612; P=0.047, 0.044, 0.002, <0.001, 0.013, 0.006); there was no significant difference in vein diameter and vein curvature (F=2.132, 1.199; P=0.143, 0.321). ConclusionArterial SO2 in BRVO patients is higher than that in healthy eyes, it decreases after anti-anti-vascular endothelial growth factor drugs treatment, SO2-V is unchanged.
[Abstract]Automatic and accurate segmentation of lung parenchyma is essential for assisted diagnosis of lung cancer. In recent years, researchers in the field of deep learning have proposed a number of improved lung parenchyma segmentation methods based on U-Net. However, the existing segmentation methods ignore the complementary fusion of semantic information in the feature map between different layers and fail to distinguish the importance of different spaces and channels in the feature map. To solve this problem, this paper proposes the double scale parallel attention (DSPA) network (DSPA-Net) architecture, and introduces the DSPA module and the atrous spatial pyramid pooling (ASPP) module in the “encoder-decoder” structure. Among them, the DSPA module aggregates the semantic information of feature maps of different levels while obtaining accurate space and channel information of feature map with the help of cooperative attention (CA). The ASPP module uses multiple parallel convolution kernels with different void rates to obtain feature maps containing multi-scale information under different receptive fields. The two modules address multi-scale information processing in feature maps of different levels and in feature maps of the same level, respectively. We conducted experimental verification on the Kaggle competition dataset. The experimental results prove that the network architecture has obvious advantages compared with the current mainstream segmentation network. The values of dice similarity coefficient (DSC) and intersection on union (IoU) reached 0.972 ± 0.002 and 0.945 ± 0.004, respectively. This paper achieves automatic and accurate segmentation of lung parenchyma and provides a reference for the application of attentional mechanisms and multi-scale information in the field of lung parenchyma segmentation.
This article aims to combine deep learning with image analysis technology and propose an effective classification method for distal radius fracture types. Firstly, an extended U-Net three-layer cascaded segmentation network was used to accurately segment the most important joint surface and non joint surface areas for identifying fractures. Then, the images of the joint surface area and non joint surface area separately were classified and trained to distinguish fractures. Finally, based on the classification results of the two images, the normal or ABC fracture classification results could be comprehensively determined. The accuracy rates of normal, A-type, B-type, and C-type fracture on the test set were 0.99, 0.92, 0.91, and 0.82, respectively. For orthopedic medical experts, the average recognition accuracy rates were 0.98, 0.90, 0.87, and 0.81, respectively. The proposed automatic recognition method is generally better than experts, and can be used for preliminary auxiliary diagnosis of distal radius fractures in scenarios without expert participation.
Motor imagery electroencephalogram (EEG) signals are non-stationary time series with a low signal-to-noise ratio. Therefore, the single-channel EEG analysis method is difficult to effectively describe the interaction characteristics between multi-channel signals. This paper proposed a deep learning network model based on the multi-channel attention mechanism. First, we performed time-frequency sparse decomposition on the pre-processed data, which enhanced the difference of time-frequency characteristics of EEG signals. Then we used the attention module to map the data in time and space so that the model could make full use of the data characteristics of different channels of EEG signals. Finally, the improved time-convolution network (TCN) was used for feature fusion and classification. The BCI competition IV-2a data set was used to verify the proposed algorithm. The experimental results showed that the proposed algorithm could effectively improve the classification accuracy of motor imagination EEG signals, which achieved an average accuracy of 83.03% for 9 subjects. Compared with the existing methods, the classification accuracy of EEG signals was improved. With the enhanced difference features between different motor imagery EEG data, the proposed method is important for the study of improving classifier performance.