Assisted Diagnosis Of Childhood Pneumonia Based On Deep Learning

The early diagnosis of childhood pneumonia presents certain difficulties,which are more pronounced in areas with inadequate medical resources.Therefore,the fight against pneumonia requires more accurate,rapid,and low-cost diagnostic methods.Our research team previously developed a fiber optic sensing respiratory signal acquisition system and used supervised machine learning for childhood pneumonia diagnosis.However,there are still issues of long processing time and high cost of labeled samples.In this paper,we further improve the computational efficiency by using deep learning.Additionally,we combine the semi-supervised algorithm with model calibration to reduce the labeling cost of samples and better utilize unlabeled samples.The main contents are as follows:1、Data processing of respiratory vibration signals.In this paper,we used the samples collected by the fiber-optic sensing respiratory signal acquisition system built by our research group in the preliminary stage,with a sample size of 320 and data dimension of 40000×1.Due to the excessively long length of the sample sequences,we transformed the one-dimensional long sequence data into two-dimensional feature images,in order to obtain a more suitable data format for model training.2、Constructing a child pneumonia diagnostic model called Kid Pneumonia Detection Model based on deep learning.In this paper,we introduce the Convolutional Block Attention Module into the network and propose using Group Normalization as a batch processing method and improving the activation function to address the problem of small sample size datasets.The proposed algorithm achieved an accuracy of 96.0%on the test set with a processing time of 280 milliseconds per sample.3、Developing a child pneumonia diagnostic model called Semi-supervised Kid Pneumonia Detection Model based on semi-supervised learning.To address the issues of small data volume and high cost of acquiring labeled samples,this paper employs the generative model Denoising Diffusion Probabilistic Models for unsupervised data augmentation.Moreover,for model calibration,the paper introduces for the first time the differentiable expected calibration error loss based on the semi-supervised training algorithm Fix Match.The final model achieves a classification accuracy of 98.3%.This paper utilizes fiber optic sensing technology to collect children’s respiratory signals.To address the high-dimensional small-sample problem,the one-dimensional sequence data is transformed into two-dimensional feature maps.By constructing a deep learning-based model for pediatric pneumonia diagnosis,faster diagnosis of pediatric pneumonia is achieved.Additionally,this paper also establishes a semisupervised learning-based model for pediatric pneumonia diagnosis.In situations where labeled samples are limited,the model can better utilize unlabeled samples,reducing annotation costs while improving the accuracy and robustness of the model.