Your browser doesn't support javascript.
loading
A feasibility study of building up deep learning classification model based on breast digital breast tomosynthesis image texture feature extraction of the simple mass lesions / 中华放射学杂志
Chinese Journal of Radiology ; (12): 668-672, 2018.
Article en Zh | WPRIM | ID: wpr-707977
Biblioteca responsable: WPRO
ABSTRACT
Objective To evaluate the diagnostic performance of digital breast tomosynthesis (DBT) breast X-ray photography image texture characteristics based deep learning classification model on differentiating malignant masses. Methods Retrospectively collected 132 cases with simplex breast lesions (89 benign lesions and 43 malignant lesions) which were confirmed by pathology and DBT during January 2016 to December 2016 in Nanfang Hospital. DBT was performed before biopsy and surgery. Image of cranio-caudal view (CC) and medio-lateral oblique (MLO) were captured. The lesion area was segmented to acquire ROI by ITK-SNAP software. Then the processed images were input into MATLAB R2015b to establish a feature model for extracting texture features. The characteristics with high correlation was analyzed from Fisher score and one sample t test. We built up support vector machine (SVM) classification model based on extracted texture and added neural network model (CNN) for deep learning classification model. We randomly assigned collected cases into training group and validation group. The diagnosis of benign and malignant lesions were served as the reference. The efficiency was evaluated by ROC classification model. Result We extracted 82 texture characteristics from 132 images of leisure (132 images of CC and 132 images of MLO) by establishing deep learning classification model of breast lesions. We randomly chose and combined characteristics from 15 texture characteristics with statistical significance, then differentiated benign and malignant by SVM classification model. After 50 iterations on each combination of characteristics, the average diagnostic efficacy was compared to obtained the one with higher efficacy. Nine of CC and 8 of MLO was selected. The result showed that the sensitivity, specificity, accuracy and area under curve (AUC) of the model to differentiate simplex breast lesions for CC were 0.68, 0.77, 0.74 and 0.74, for MLO were 0.71, 0.71, 0.71 and 0.76. Conclusions MLO has better diagnostic performance for the diagnosis than CC. The deep learning classification model on breast lesions which was built upon DBT image texture characteristics on MLO could differentiate malignant masses effectively.
Palabras clave
Texto completo: 1 Índice: WPRIM Tipo de estudio: Prognostic_studies Idioma: Zh Revista: Chinese Journal of Radiology Año: 2018 Tipo del documento: Article
Texto completo: 1 Índice: WPRIM Tipo de estudio: Prognostic_studies Idioma: Zh Revista: Chinese Journal of Radiology Año: 2018 Tipo del documento: Article