Swin Unet3D: a three-dimensional medical image segmentation network combining vision transformer and convolution.

BACKGROUND: Semantic segmentation of brain tumors plays a critical role in clinical treatment, especially for three-dimensional (3D) magnetic resonance imaging, which is often used in clinical practice. Automatic segmentation of the 3D structure of brain tumors can quickly help physicians understand the properties of tumors, such as the shape and size, thus improving the efficiency of preoperative planning and the odds of successful surgery. In past decades, 3D convolutional neural networks (CNNs) have dominated automatic segmentation methods for 3D medical images, and these network structures have achieved good results. However, to reduce the number of neural network parameters, practitioners ensure that the size of convolutional kernels in 3D convolutional operations generally does not exceed [Formula: see text], which also leads to CNNs showing limitations in learning long-distance dependent information. Vision Transformer (ViT) is very good at learning long-distance dependent information in images, but it suffers from the problems of many parameters. What's worse, the ViT cannot learn local dependency information in the previous layers under the condition of insufficient data. However, in the image segmentation task, being able to learn this local dependency information in the previous layers makes a big impact on the performance of the model. METHODS: This paper proposes the Swin Unet3D model, which represents voxel segmentation on medical images as a sequence-to-sequence prediction. The feature extraction sub-module in the model is designed as a parallel structure of Convolution and ViT so that all layers of the model are able to adequately learn both global and local dependency information in the image. RESULTS: On the validation dataset of Brats2021, our proposed model achieves dice coefficients of 0.840, 0.874, and 0.911 on the ET channel, TC channel, and WT channel, respectively. On the validation dataset of Brats2018, our model achieves dice coefficients of 0.716, 0.761, and 0.874 on the corresponding channels, respectively. CONCLUSION: We propose a new segmentation model that combines the advantages of Vision Transformer and Convolution and achieves a better balance between the number of model parameters and segmentation accuracy. The code can be found at https://github.com/1152545264/SwinUnet3D .

An imConvNet-based deep learning model for Chinese medical named entity recognition.

BACKGROUND: With the development of current medical technology, information management becomes perfect in the medical field. Medical big data analysis is based on a large amount of medical and health data stored in the electronic medical system, such as electronic medical records and medical reports. How to fully exploit the resources of information included in these medical data has always been the subject of research by many scholars. The basis for text mining is named entity recognition (NER), which has its particularities in the medical field, where issues such as inadequate text resources and a large number of professional domain terms continue to face significant challenges in medical NER. METHODS: We improved the convolutional neural network model (imConvNet) to obtain additional text features. Concurrently, we continue to use the classical Bert pre-training model and BiLSTM model for named entity recognition. We use imConvNet model to extract additional word vector features and improve named entity recognition accuracy. The proposed model, named BERT-imConvNet-BiLSTM-CRF, is composed of four layers: BERT embedding layer-getting word embedding vector; imConvNet layer-capturing the context feature of each character; BiLSTM (Bidirectional Long Short-Term Memory) layer-capturing the long-distance dependencies; CRF (Conditional Random Field) layer-labeling characters based on their features and transfer rules. RESULTS: The average F1 score on the public medical data set yidu-s4k reached 91.38% when combined with the classical model; when real electronic medical record text in impacted wisdom teeth is used as the experimental object, the model's F1 score is 93.89%. They all show better results than classical models. CONCLUSIONS: The suggested novel model (imConvNet) significantly improves the recognition accuracy of Chinese medical named entities and applies to various medical corpora.

One-stage and lightweight CNN detection approach with attention: Application to WBC detection of microscopic images.

White blood cell (WBC) detection in microscopic images is indispensable in medical diagnostics; however, this work, based on manual checking, is time-consuming, labor-intensive, and easily results in errors. Using object detectors for WBCs with deep convolutional neural networks can be regarded as a feasible solution. In this paper, to improve the examination precision and efficiency, a one-stage and lightweight CNN detector with an attention mechanism for detecting microscopic WBC images, and a white blood cell detection vision system are proposed. The method integrates different optimizing strategies to strengthen the feature extraction capability through the combination of an improved residual convolution module, hybrid spatial pyramid pooling module, improved coordinate attention mechanism, efficient intersection over union (EIOU) loss and Mish activation function. Extensive ablation and contrast experiments on the latest public Raabin-WBC dataset verify the effectiveness and robustness of the proposed detector for achieving a better overall detection performance. It is also more efficient than other existing studies for blood cell detection on two additional classic public BCCD and LISC datasets. The novel detection approach is significant and flexible for medical technicians to use for blood cell microscopic examination in clinical practice.

SMD-YOLO: An efficient and lightweight detection method for mask wearing status during the COVID-19 pandemic.

BACKGROUND AND OBJECTIVE: At present, the COVID-19 epidemic is still spreading worldwide and wearing a mask in public areas is an effective way to prevent the spread of the respiratory virus. Although there are many deep learning methods used for detecting the face masks, there are few lightweight detectors having a good effect on small or medium-size face masks detection in the complicated environments. METHODS: In this work we propose an efficient and lightweight detection method based on YOLOv4-tiny, and a face mask detection and monitoring system for mask wearing status. Two feasible improvement strategies, network structure optimization and K-means++ clustering algorithm, are utilized for improving the detection accuracy on the premise of ensuring the real-time face masks recognition. Particularly, the improved residual module and cross fusion module are designed to aim at extracting the features of small or medium-size targets effectively. Moreover, the enhanced dual attention mechanism and the improved spatial pyramid pooling module are employed for merging sufficiently the deep and shallow semantic information and expanding the receptive field. Afterwards, the detection accuracy is compensated through the combination of activation functions. Finally, the depthwise separable convolution module is used to reduce the quantity of parameters and improve the detection efficiency. Our proposed detector is evaluated on a public face mask dataset, and an ablation experiment is also provided to verify the effectiveness of our proposed model, which is compared with the state-of-the-art (SOTA) models as well. RESULTS: Our proposed detector increases the AP (average precision) values in each category of the public face mask dataset compared with the original YOLOv4-tiny. The mAP (mean average precision) is improved by 4.56% and the speed reaches 92.81 FPS. Meanwhile, the quantity of parameters and the FLOPs (floating-point operations) are reduced by 1/3, 16.48%, respectively. CONCLUSIONS: The proposed detector achieves better overall detection performance compared with other SOTA detectors for real-time mask detection, demonstrated the superiority with both theoretical value and practical significance. The developed system also brings greater flexibility to the application of face mask detection in hospitals, campuses, communities, etc.