Sleep apnea (SA) detection method based on traditional machine learning needs a lot of efforts in feature engineering and classifier design. We constructed a one-dimensional convolutional neural network (CNN) model, which consists in four convolution layers, four pooling layers, two full connection layers and one classification layer. The automatic feature extraction and classification were realized by the structure of the proposed CNN model. The model was verified by the whole night single-channel sleep electrocardiogram (ECG) signals of 70 subjects from the Apnea-ECG dataset. Our results showed that the accuracy of per-segment SA detection was ranged from 80.1% to 88.0%, using the input signals of single-channel ECG signal, RR interval (RRI) sequence, R peak sequence and RRI sequence + R peak sequence respectively. These results indicated that the proposed CNN model was effective and can automatically extract and classify features from the original single-channel ECG signal or its derived signal RRI and R peak sequence. When the input signals were RRI sequence + R peak sequence, the CNN model achieved the best performance. The accuracy, sensitivity and specificity of per-segment SA detection were 88.0%, 85.1% and 89.9%, respectively. And the accuracy of per-recording SA diagnosis was 100%. These findings indicated that the proposed method can effectively improve the accuracy and robustness of SA detection and outperform the methods reported in recent years. The proposed CNN model can be applied to portable screening diagnosis equipment for SA with remote server.
Coronavirus disease 2019 (COVID-19) has spread rapidly around the world. In order to diagnose COVID-19 more quickly, in this paper, a depthwise separable DenseNet was proposed. The paper constructed a deep learning model with 2 905 chest X-ray images as experimental dataset. In order to enhance the contrast, the contrast limited adaptive histogram equalization (CLAHE) algorithm was used to preprocess the X-ray image before network training, then the images were put into the training network and the parameters of the network were adjusted to the optimal. Meanwhile, Leaky ReLU was selected as the activation function. VGG16, ResNet18, ResNet34, DenseNet121 and SDenseNet models were used to compare with the model proposed in this paper. Compared with ResNet34, the proposed classification model of pneumonia had improved 2.0%, 2.3% and 1.5% in accuracy, sensitivity and specificity respectively. Compared with the SDenseNet network without depthwise separable convolution, number of parameters of the proposed model was reduced by 43.9%, but the classification effect did not decrease. It can be found that the proposed DWSDenseNet has a good classification effect on the COVID-19 chest X-ray images dataset. Under the condition of ensuring the accuracy as much as possible, the depthwise separable convolution can effectively reduce number of parameters of the model.
Attention deficit/hyperactivity disorder (ADHD) is a behavioral disorder syndrome found mainly in school-age population. At present, the diagnosis of ADHD mainly depends on the subjective methods, leading to the high rate of misdiagnosis and missed-diagnosis. To solve these problems, we proposed an algorithm for classifying ADHD objectively based on convolutional neural network. At first, preprocessing steps, including skull stripping, Gaussian kernel smoothing, et al., were applied to brain magnetic resonance imaging (MRI). Then, coarse segmentation was used for selecting the right caudate nucleus, left precuneus, and left superior frontal gyrus region. Finally, a 3 level convolutional neural network was used for classification. Experimental results showed that the proposed algorithm was capable of classifying ADHD and normal groups effectively, the classification accuracies obtained by the right caudate nucleus and the left precuneus brain regions were greater than the highest classification accuracy (62.52%) in the ADHD-200 competition, and among 3 brain regions in ADHD and the normal groups, the classification accuracy from the right caudate nucleus was the highest. It is well concluded that the method for classification of ADHD and normal groups proposed in this paper utilizing the coarse segmentation and deep learning is a useful method for the purpose. The classification accuracy of the proposed method is high, and the calculation is simple. And the method is able to extract the unobvious image features better, and can overcome the shortcomings of traditional methods of MRI brain area segmentation, which are time-consuming and highly complicate. The method provides an objective diagnosis approach for ADHD.
When applying deep learning to the automatic segmentation of organs at risk in medical images, we combine two network models of Dense Net and V-Net to develop a Dense V-network for automatic segmentation of three-dimensional computed tomography (CT) images, in order to solve the problems of degradation and gradient disappearance of three-dimensional convolutional neural networks optimization as training samples are insufficient. This algorithm is applied to the delineation of pelvic endangered organs and we take three representative evaluation parameters to quantitatively evaluate the segmentation effect. The clinical result showed that the Dice similarity coefficient values of the bladder, small intestine, rectum, femoral head and spinal cord were all above 0.87 (average was 0.9); Jaccard distance of these were within 2.3 (average was 0.18). Except for the small intestine, the Hausdorff distance of other organs were less than 0.9 cm (average was 0.62 cm). The Dense V-Network has been proven to achieve the accurate segmentation of pelvic endangered organs.
With the rapid development of network structure, convolutional neural networks (CNN) consolidated its position as a leading machine learning tool in the field of image analysis. Therefore, semantic segmentation based on CNN has also become a key high-level task in medical image understanding. This paper reviews the research progress on CNN-based semantic segmentation in the field of medical image. A variety of classical semantic segmentation methods are reviewed, whose contributions and significance are highlighted. On this basis, their applications in the segmentation of some major physiological and pathological anatomical structures are further summarized and discussed. Finally, the open challenges and potential development direction of semantic segmentation based on CNN in the area of medical image are discussed.
The application of minimally invasive surgical tool detection and tracking technology based on deep learning in minimally invasive surgery is currently a research hotspot. This paper firstly expounds the relevant technical content of the minimally invasive surgery tool detection and tracking, which mainly introduces the advantages based on deep learning algorithm. Then, this paper summarizes the algorithm for detection and tracking surgical tools based on fully supervised deep neural network and the emerging algorithm for detection and tracking surgical tools based on weakly supervised deep neural network. Several typical algorithm frameworks and their flow charts based on deep convolutional and recurrent neural networks are summarized emphatically, so as to enable researchers in relevant fields to understand the current research progress more systematically and provide reference for minimally invasive surgeons to select navigation technology. In the end, this paper provides a general direction for the further research of minimally invasive surgical tool detection and tracking technology based on deep learning.
Alzheimer’s Disease (AD) is a progressive neurodegenerative disorder. Due to the subtlety of symptoms in the early stages of AD, rapid and accurate clinical diagnosis is challenging, leading to a high rate of misdiagnosis. Current research on early diagnosis of AD has not sufficiently focused on tracking the progression of the disease over an extended period in subjects. To address this issue, this paper proposes an ensemble model for assisting early diagnosis of AD that combines structural magnetic resonance imaging (sMRI) data from two time points with clinical information. The model employs a three-dimensional convolutional neural network (3DCNN) and twin neural network modules to extract features from the sMRI data of subjects at two time points, while a multi-layer perceptron (MLP) is used to model the clinical information of the subjects. The objective is to extract AD-related features from the multi-modal data of the subjects as much as possible, thereby enhancing the diagnostic performance of the ensemble model. Experimental results show that based on this model, the classification accuracy rate is 89% for differentiating AD patients from normal controls (NC), 88% for differentiating mild cognitive impairment converting to AD (MCIc) from NC, and 69% for distinguishing non-converting mild cognitive impairment (MCInc) from MCIc, confirming the effectiveness and efficiency of the proposed method for early diagnosis of AD, as well as its potential to play a supportive role in the clinical diagnosis of early Alzheimer's disease.
Glaucoma is the leading cause of irreversible blindness, but its early symptoms are not obvious and are easily overlooked, so early screening for glaucoma is particularly important. The cup to disc ratio is an important indicator for clinical glaucoma screening, and accurate segmentation of the optic cup and disc is the key to calculating the cup to disc ratio. In this paper, a full convolutional neural network with residual multi-scale convolution module was proposed for the optic cup and disc segmentation. First, the fundus image was contrast enhanced and polar transformation was introduced. Subsequently, W-Net was used as the backbone network, which replaced the standard convolution unit with the residual multi-scale full convolution module, the input port was added to the image pyramid to construct the multi-scale input, and the side output layer was used as the early classifier to generate the local prediction output. Finally, a new multi-tag loss function was proposed to guide network segmentation. The mean intersection over union of the optic cup and disc segmentation in the REFUGE dataset was 0.904 0 and 0.955 3 respectively, and the overlapping error was 0.178 0 and 0.066 5 respectively. The results show that this method not only realizes the joint segmentation of cup and disc, but also improves the segmentation accuracy effectively, which could be helpful for the promotion of large-scale early glaucoma screening.
Cardiac auscultation is the basic way for primary diagnosis and screening of congenital heart disease(CHD). A new classification algorithm of CHD based on convolution neural network was proposed for analysis and classification of CHD heart sounds in this work. The algorithm was based on the clinically collected diagnosed CHD heart sound signal. Firstly the heart sound signal preprocessing algorithm was used to extract and organize the Mel Cepstral Coefficient (MFSC) of the heart sound signal in the one-dimensional time domain and turn it into a two-dimensional feature sample. Secondly, 1 000 feature samples were used to train and optimize the convolutional neural network, and the training results with the accuracy of 0.896 and the loss value of 0.25 were obtained by using the Adam optimizer. Finally, 200 samples were tested with convolution neural network, and the results showed that the accuracy was up to 0.895, the sensitivity was 0.910, and the specificity was 0.880. Compared with other algorithms, the proposed algorithm has improved accuracy and specificity. It proves that the proposed method effectively improves the robustness and accuracy of heart sound classification and is expected to be applied to machine-assisted auscultation.
With the rapid improvement of the perception and computing capacity of mobile devices such as smart phones, human activity recognition using mobile devices as the carrier has been a new research hot-spot. The inertial information collected by the acceleration sensor in the smart mobile device is used for human activity recognition. Compared with the common computer vision recognition, it has the following advantages: convenience, low cost, and better reflection of the essence of human motion. Based on the WISDM data set collected by smart phones, the inertial navigation information and the deep learning algorithm-convolutional neural network (CNN) were adopted to build a human activity recognition model in this paper. The K nearest neighbor algorithm (KNN) and the random forest algorithm were compared with the CNN network in the recognition accuracy to evaluate the performance of the CNN network. The classification accuracy of CNN model reached 92.73%, which was much higher than KNN and random forest. Experimental results show that the CNN algorithm model can achieve more accurate human activity recognition and has broad application prospects in predicting and promoting human health.