Evaluation Of Axillary Lymph Node Metastasis Of Breast Cancer Based On Dual-modal Ultrasound Image And Deep Learning

The incidence of breast cancer ranks first among female cancers,accounting for 30% of all newly diagnosed cancers.It is the most menacing for women’s health.The status of lymph node(LN)metastasis is one of the most important prognostic factors for breast cancer.Timely and accurate detection of axillary lymph node(ALN)metastasis is very important to guide the clinical treatment of breast cancer.ALN dissection is a gold standard for clinical judgment of LN metastasis.However,ALN dissection is an invasive surgical method,which can lead to many postoperative complications,such as lymphedema,seroma,and infectious neuropathy,and will result in over-treatment for the patients with early breast cancer.Axillary ultrasound is a method for assessing ALNs in patients with breast disease,which can be divided into two categories.One is B-mode ultrasound,which mainly detects malignant LNs based on morphological features;the other is shear wave Elastography(SWE)ultrasound,which uses SWE to measure the stiffness of LN tissues to assess the LN metastasis.However,the interpretation of ultrasound images relies on the subjective assessment of radiologists,which leads to differences between different observers and affects the accuracy of diagnosis.Thus,the diagnostic efficiency of noninvasive detection of ALN metastasis is limited and it is necessary to study an objective,accurate and reliable method to assist doctors in improving the diagnosis of ALN metastasis.The thesis mainly describes the following contents.1.With the cooperation of the Department of Ultrasound,Sun Yat-sen University Affiliated Tumor Hospital,214 breast cancer patients admitted for treatment from 2018 to2020 have been collected among these patients,440 B-mode ultrasound and SWE ultrasound images of the patients have been acquired as the experimental dataset in our study.The LN regions of interests are delineated from the images by means of an annotation software.2.464 various image features are extracted from ALN B-mode ultrasound images by feature engineering,including first-order statistical features,morphological features,grayscale texture features,and wavelet features.Then,three feature selection methods are applied to select various feature subsets,which are univariate statistics,model-based selection and recursively elimination.Next,three classifiers,including Logistic regression,support vector machine,and multi-layer perceptron,are employed to build a classification model,which can explore different combinations of feature subsets for the diagnostic performance of ALN metastasis.Also,different classifiers are evaluated.3.A deep learning based heterogeneous model(DLHM)is proposed in this thesis.It employs convolution neural network to automatically mine intrinsic medical hints contained in ultrasound images,which can deal with the problem that the key medical information existing between the LN and surrounding tissues may be ignored when the outline of the LN is manually delineated.The feature-level fusion can organically combine radiomics features with deep image features,which effectively improves the diagnostic efficiency of the model.Experimental results show that the DLHM can achieve the AUC,accuracy,sensitivity and specificity of 0.911,86.2%,82.8%,and 90.2%,respectively,for on an independent test cohort,which demonstrates good diagnostic evaluation of LN metastasis.Also,it indicates that multi-source heterogeneous data can effectively improve the generalization performance of deep learning on small-scale data sets.