Research On Medical Image Feature Fusion Algorithm Based On Deep Learning

The feature fusion of medical images involves extracting features from medical images and utilizing the complementarity between features to construct comprehensive feature attribute descriptions from multiple perspectives.Deep learning has achieved great success in image-based disease diagnosis.However,in order to achieve the best results,deep learning methods require the use of a large number of images and require medical experts to accurately annotate the images,which consumes a lot of labor costs.On the contrary,in clinical practice,doctors do not need such a large number of samples to quickly improve their diagnostic level,thanks to their ability to identify and compare anatomical structures in images.This article focuses on the diagnosis of glaucoma,drawing on the diagnostic thinking of doctors,and constructing an intelligent diagnosis system that integrates expert knowledge and small samples under the framework of integrated deep learning.In this system,expert knowledge is used to obtain multi view information of medical images,including the defect features of optic disc,optic cup and optic nerve fiber layer in fundus photography,and is combined through different global and local angles;Small samples are used to construct specific classifiers at each angle.Then,ensemble learning is used to fuse the predictions of all classifiers to obtain the final diagnostic result.In clinical data from Beijing Tongren Hospital,hospitals in Tibet and Ningxia Autonomous Region of China,and the First Hospital of Harbin Medical University,the intelligent diagnostic system has achieved accurate prediction results.We obtained classified AUC values of0.998,0.998,and 0.984 on three batches of data from different hospitals.Compared to classic deep learning classification methods,the prediction accuracy has been significantly improved.In order to explore the universality of the intelligent diagnosis system,we further applied the scheme to the diagnosis of diabetes retinopathy,pathological myopia and age-related macular disease,especially in the case of small sample learning.The intelligent diagnostic system achieved classification AUC values of 0.916(basic network 0.898),0.986(basic network 0.970),and 0.953(basic network 0.720)in three disease classifications.This result indicates that the intelligent diagnostic system designed in this article can introduce multi angle information for classification,and has a certain improvement effect on classification accuracy performance,confirming the generalization performance of the intelligent diagnostic system.The expansion of the three diseases further confirms the importance of intelligent diagnostic systems for current medical image classification tasks.In order to further elaborate on the effectiveness principle of intelligent diagnostic systems,this article also proposes a theoretical framework for weighted ensemble learning convergence analysis,which analyzes the effectiveness of intelligent diagnostic systems from a mathematical perspective.