A Study Of Light-weight And Multi-scale Methods In Biomedical Image Processing

As an interdisciplinary subject of computer and biomedicine,biomedical engineering has been using computer technology to assist clinical medicine very early.Among them,the processing of biomedical images is particularly important.As the most used auxiliary diagnosis and treatment method in modern medicine,fast and effective analysis and processing of biomedical images has become a current research hotspot.More and more researchers have begun to deepen the research of biomedical image processing.This paper mainly conducts in-depth research on lightweight multi-scale methods in the field of biomedical image segmentation.The main research contents are as follows:(1)In the study of lightweight and multi-scale biomedical image segmentation methods,a new model named PyConvU-Net was proposed for biomedical images in view of the huge volume of deep learning models and the problem of relying on more computing resources.Segmentation,the model uses pyramid convolution to replace traditional convolution,which can reduce the size of the model and reduce the computational complexity while processing multi-scale features.The model has MIoU and Dice values of 0.963 and 0.9339 on the lung segmentation dataset,0.7050 and 0.8277 on the kidney segmentation dataset,and 0.8385 on the cell segmentation dataset.and 0.9177,these values are basically in the first or second position,so the model performs well in accuracy.The values in the number of parameters and computational complexity are 3.7MB and 10.65 GMac,which are much higher than the second place.The experimental results show that PyConvU-Net can ensure that the parameters and computational complexity of the model can be fully reduced on the premise of ensuring accuracy in the segmentation of biomedical images.(2)In the research of biomedical image segmentation methods based on selfsupervision,a new model named SimTrans is proposed for the small number of samples in biomedical image datasets and the high cost of labeling biomedical image data.Obtain a good performing biomedical image segmentation model in fewer cases.The model uses a self-supervised learning strategy to construct pseudo-labels of data through proxy tasks,and then uses pseudo-labels to train the model,so that the model learns the distribution characteristics of data from the data space itself,and then applies the trained model to downstream tasks,such as classification,segmentation,etc.The MIoU and Dice values of this model on the stroke segmentation dataset are 0.6424 and 0.7585,respectively,higher than the second place.The model successfully breaks away from the dependence of the true labels of the data,learns data features through proxy tasks,and achieves good performance in downstream tasks.The PyConvU-Net and SimTrans methods proposed in this paper have achieved good results in biomedical image segmentation and provide some guidance for exploring biomedical image segmentation.