Research On Intelligent Recognition Method Of Pneumothorax Based On Attention Mechanism

Deep learning techniques have been developed in recent years and are widely used in various fields,and their role in the medical field is becoming more and more prominent.Pneumothorax is a common lung disease with rapid onset and high risk,and its diagnosis often relies on manual identification by X-ray films,which has certain limitations.As the application of artificial intelligence in medical image recognition deepens,more and more diseases can be diagnosed with the aid of computers,which can not only reduce the work pressure of medical staffs,but also effectively reduce the probability of misdiagnosis.This paper proposes two intelligent recognition models for pneumothorax based on attention mechanism,which are based on Hard Attention and Soft Attention respectively.The former uses a lung segmentation sub-model to extract lung regions from X-ray chest films for pneumothorax recognition.The latter combines Swin-Transformer with convolutional neural networks to propose a Hybrid EfficientNet Swin-Transformer pneumothorax intelligent recognition model with gradient accumulation technique for training and reduced memory usage limitation.This paper uses pneumothorax X-ray chest films from the Chest X-rays dataset and provided by Nanyang Central Hospital in Henan Province.Traditional convolutional neural networks have the disadvantages of limited ability to extract features,easy loss of image edge information and lack of globalization.Among the two models proposed in this paper,the hard attention-based pneumothorax recognition model uses an improved U-Net network for effective extraction of lung regions from X-ray chest films with an Io U coefficient of 0.968;and the Dense Net169 model is selected as a classifier for pneumothorax recognition of the segmented lung region images.The results show that,compared to the single Dense Net classification network,the intelligent pneumothorax recognition model combined with the lung segmentation algorithm has a 0.17 improvement in accuracy,with an accuracy of 0.97.In addition,the proposed Hybrid EfficientNet Swin-Transformer model based on soft attention has a pneumothorax recognition accuracy of 0.96.Compared to the single Swin-Transfomer model,the hybrid model combined with EffcientNet has an 0.08 improvement in recognition accuracy.