Study On Ambiguity Assistant Diagnosis Of Lumbar Disc Degeneration

Ambiguity often exists in diagnosis of lumbar disc degeneration（LDD）based on the Pfirrmann grading system,in which,the imaging physicians might make the disagree or uncertain judgment on the adjacent degenerative grades.At present,few studies have paid attentions to the ambiguous diagnosis of LDD.In this paper,based on the discal metabolomics detected by the magnetic resonance image（MRI）,an ambiguity-aware classifier of LDD has been proposed by means of the subjective probability quantitation and label distribution learning.Such classifier can provide assistance for the clinical diagnosis of doctors.First of all,the related theories are systematically introduced,including the structure and functions of LDD,the basic theory of label distribution learning,the subjective probability quantification method,decomposition of label distribution data sets,and data normalization formulas.The modeling ideas and optimization of the label distribution classifiers,including PT-SVM,PT-Bayes,AA-KNN,AA-BP,and SA-IIS,emphatically described.Secondly,390 samples are clinically collected,where the instance of each sample consists of three discal biochemical metabolic indexes including T2* value,and upper and lower vertebral fat fraction（FF）which are detected by MRI,the label of each sample is represented as a probability distribution generated by labelers with the help of the interval bisection method.Therefore,such metabolomic sample is described by three metabolic indexes and a probability distribution label.To preprocess the sample set,a K-fold cross-validation method for the label distribution dataset is proposed to divide the sample set into the training set and test set,which are then normalized,respectively.According to the modeling ideas and optimization method of PT-SVM,PT-Bayes,AA-KNN,and AA-BP algorithms,the parameters of the four classifiers are optimized by using the training set.The performances of the four classifiers are then analyzed and compared by using the test set.A novel metric,Consensus of Ordered Categories（COC）,is proposed to evaluate the accuracy of the label distribution learning algorithm.Thirdly,to further study the diagnostic ability of the ambiguity-aware classifiers to early LDD,three ambiguity-aware classifiers,PT-Bayes,AA-BP,and SA-IIS,are modeled by using early degenerative samples such as those of grades I and II.The mean-similarity-based splitting sample,a partitioning method available to the small sets of label distribution samples,is used to decompose the sample set into the training set and the test set.Moreover,the parameters of the classifiers are optimized by using the training set,and the performances of the classifiers are analyzed and compared in the test set using some selected metrics.Final part of this paper discusses how the proposed ambiguity-aware classifiers assist doctors to eliminate ambiguity and uncertainty of LDD diagnosis.Some suggestions for the interpretability of the ambiguity-aware classifiers are presented.