Multi-Lesion Recognition In Endoscopic Images Based On Improved ADMM Algorithm

Capsule endoscopy(CE),as a non-invasive and painless auxiliary diagnostic equipment,has been widely used in the diagnosis of human gastrointestinal tract.A large number of image samples will be generated in CE diagnosis.Accurate and efficient use of these image information to achieve lesion diagnosis has become a hotspot in recent years.Machine Learning(ML)and Deep Learning(DL)are the main methods to achieve CE image lesion recognition.The former uses candidate features,while the latter often uses Convolutional Neural Network(CNN)to extract depth.feature.However,candidate features are limited by existing knowledge,and deep features are highly dependent on the completeness of the sample set,both of which are ineffective in CE image multi-lesion recognition.Considering the coupling effect between the knowledge corresponding to the candidate features and the semantic features contained in the deep features,how to extract and select appropriate features for fusion,and then form an integrated classifier for multi-label classification is a feasible way to solve the multi-lesion recognition in CE images.Three CNN models are selected as feature extractors for deep features,and these features are combined with three traditional features of color,texture,and shape to form a feature library.The Alternating Direction Method of Multipliers(ADMM)assigns weights to different feature components,and achieves feature selection by eliminating redundant feature components with too low weights.On this basis,the feature combination is adjusted to form a multi-label classifier with high generalization ability and less computational redundancy.In this process,aiming at the problem of unbalanced and incomplete lesion samples,data enhancement is used to balance,and for the special requirements of medical images for misclassification,a penalty matrix Δ is designed to reduce the missed detection rate of lesions.Since different lesions have far different feature expressions,it is difficult to distinguish different lesions with the same feature vector,and multi-classifier ensemble can effectively improve the classification effect.Therefore,a classifier ensemble method based on Bayesian theory is proposed.For four common lesions such as ulcers and bleeding,the above ADMM-based feature selection method was used to form four three-label classifiers.The classification effects of these classifiers in the independent test set were used as prior knowledge,and the results of each classifier were analyzed Fusion to get an ensemble classifier.The experimental results verify the complementarity between deep features and candidate features.The features obtained through ADMM feature selection have better representation ability and can meet the requirements of multi-type lesion identification.Based on the fusion of traditional features and deep features,ADMM algorithm is used for feature selection to remove redundant features,which can achieve better recognition results.At the same time,the method of ensemble fusion of classifier results based on Bayes’ theorem further improves our identification accuracy.These experimental results show the advanced nature of the research methods in this paper and have certain application prospects.