Research On Deep Learning-Based Lightweight And Label Fusion For Pulmonary Disease Recognition

In the present world,the incidence and mortality rates of lung diseases such as lung cancer are increasing rapidly,making it one of the most significant threats to public health.Early screening for lung diseases can greatly improve the chances of survival for patients.With the advancements in deep learning and computer-aided diagnosis technology,the application of deep learning-based computer-aided diagnosis techniques,combining knowledge from radiology,medical image processing,and pattern recognition,is gaining increasing popularity in clinical practice.Applying these techniques to medical image research has become a hot topic of study.However,existing deep learning models are often designed for large-scale datasets,leading to a large number of parameters and computational requirements.Balancing detection accuracy with improved speed has become a challenge.Furthermore,in the medical field,diseases exhibit complex dependencies on factors such as age,gender,and other existing conditions,which are often not adequately considered in traditional recognition methods.To address these issues,this research explores lung diseases based on deep learning principles and incorporates technologies such as Transformer and Graph Attention Networks(GAT)for in-depth analysis.The study focuses on the following aspects:(1)To achieve a lightweight deep learning network model that maintains detection accuracy while improving detection speed,a novel lightweight Transformer-based algorithm called LWTFNet is proposed for the identification of COVID-19 from CT scans.The algorithm employs an enhanced version of Mobile Net V2 for feature extraction,followed by multiple hierarchical Transformer layers to reduce parameter count.The introduction of neighborhood attention mechanism replaces the conventional self-attention mechanism,further reducing parameter quantity while extracting both global and local information to enhance intra-group and inter-group feature interaction.Finally,the integrate COVID-19 CT dataset is used for testing.Experimental results demonstrate that the proposed model significantly reduces parameter count compared to other models and exhibits excellent performance across multiple evaluation metrics,holding positive implications for computer-aided diagnosis of COVID-19.(2)To leverage the information among labels and enhance model performance,this paper presents a multi-label classification algorithm,named ICLGNet,which integrates contrastive learning and Graph Attention Networks.The algorithm employs Efficient Net V2 as the backbone network for extracting network features and feeds disease labels as well as auxiliary labels such as age and gender into the Graph Attention Network.The introduction of contrastive learning loss addresses the oversmoothing issue within the Graph Attention Network and strengthens the discriminative power of label embeddings.Subsequently,the label-dependent features extracted by the Graph Attention Network and the image features extracted by Efficient Net are fed into the Transformer Decoder layers of the two-stage decoding algorithm designed in this study.Initially,features such as age,gender,and the presence of disease are identified,followed by the identification of specific diseases based on these features.Finally,the proposed method is tested on the publicly available Chest Xray14 dataset.The comprehensive experimental results demonstrate that the proposed approach significantly improves the accuracy of multi-disease identification in the lungs,achieving an accuracy of 0.95 and an AUC of 0.8872,thereby greatly aiding in the diagnosis of lung diseases.