Research On Few-shot Medical Image Segmentation Based On Meta-Transfer Learning

With the development of deep learning,semantic segmentation has gradually replaced manual segmentation and traditional segmen-tation methods and become an important step in medical diagnosis and image analysis.Segmentation annotation of medical image has huge time and labor costs.The lack of high-quality training data often becomes an important factor restricting the performance of deep learning semantic segmentation model.Therefore,the research of medical image segmentation algorithm and the research of few-shot algorithm are two effective ways to solve the problem,and they are also the current hot research direction.After the investigation and research of medical image segmenta-tion algorithms at home and abroad,this paper proposes a multi-scale segmentation algorithm based on high-density dense connection and a few-shot medical image segmentation algorithm based on meta-transfer learning.The former combines dense connection,multi-scale information collection and medical image characteristics,proposes a high-density dense connection module,optimizes multi-scale fusion module and introduces attention mechanism,and achieves Dice coef-ficient scores of 0.948 and 0.945 in training set and test set.The latter improves the existing model independent meta-learning method,and introduces the transfer learning method to maintain the model param-eters from forgetting damage during domain migration,which greatly improves the model performance in the case of few samples.In the case of sufficient training data,the algorithm can also improve the accuracy.In this paper,we train and test the algorithm based on cardiac MRI data provided by the Department of radiology of Peking Union Medical College Hospital and Medical Segmentation Decathlon com-petition data.Compared with other algorithms,experiments and re-sults analysis verify the effectiveness of the proposed algorithm.