Multi-layer Graph Sparse Non-negative Matrix Factorization And Its Application In Cervical Precancerous Lesion Recognition

Cervical cancer is one of the most common malignant tumors endangering women's health,and the incidence of cervical cancer in China ranks second in the world.At present,the diagnosis and screening methods of cervical cancer mainly include cervical biopsy scrape,surgical pathology,computed tomography(CT),magnetic resonance imaging(MRI)and ultrasound,etc.Compared with other methods,ultrasound is safe,non-invasive,fast,highly repeatable and economical.Cervical precancerous lesion screening is essential for the diagnosis and treatment of cervical cancer.However,most patients with cervical precancerous lesions have no obvious clinical symptoms on the ultrasound image data.Therefore,it is more difficult to timely and accurate diagnosis.In order to solve the above problems,the present research discusses the intelligent recognition of cervical precancerous lesion and and the analysis of their deep features based on the cervical ultrasound clinical data provided by Chongqing Maternal and Child Health Hospital,The main work is as follows:(1)In order to mine the characteristics of cervical ultrasound data,a multi-layer graph sparse nonnegative matrix factorization model is constructed,and its optimal solution and convergence are discussed.Firstly,the idea of manifold learning is introduced,the geometric structure information of actual data is considered,and the graph regularization term is added on the basis of the sparse non-negative matrix factorization model.In order to further alleviate the influence of random initial values and to dig deeper features,the deep representation and layer-by-layer pre-training strategies in deep learning are introduced,the multi-layer graph sparse non-negative matrix factorization model is constructed,and the optimal solution and convergence analysis of the model are carried out.(2)An intelligent recognition method for cervical precancerous lesions based on multi-layer graph sparse non-negative matrix factorization model is proposed.Combining cervical pathological data with cervical ultrasound data,the pathological diagnosis results are used as the category labels,which are the gold standard for clinical diagnosis.Based on the multi-layer graph sparse non-negative matrix factorization model,deep feature mining is carried out on ultrasonic data.In the case of insufficient labeled training samples of the same category,information is mined from the existing unlabeled data as much as possible,and inverse space sparse representation classification is applied to identify cervical precancerous lesions.(3)A method for trend analysis of cervical ultrasound features based on a multi-layer graph sparse non-negative matrix factorization model is proposed.In order to further explore the difference between normal and precancerous lesions and provide clinicians with a more intuitive display,Each ultrasound image is preprocessed,the areas of interest to the doctor are first divided into subsections,and arranged them in order from the outer mouth to the inner mouth.Then the characteristic trend of the sub-block ultrasonic data is analyzed by the multi-layer graph sparse non-negative matrix factorization model.