Mass detection in digital breast tomosynthesis: Deep convolutional neural network with transfer learning from mammography

• DOI: 10.1118/1.4967345
• Publication Date: 2016-11-29T18:01:47Z
• Authors: Ravi K. Samala; Heang‐Ping Chan; Lubomir Hadjiiski; Mark A. Helvie; Jun Wei; Kenny Cha
• Keywords: Data set; Overfitting; Digital Mammography; Transfer of learning; Breast imaging; Normalization

Develop a computer-aided detection (CAD) system for masses in digital breast tomosynthesis (DBT) volume using deep convolutional neural network (DCNN) with transfer learning from mammograms.A data set containing 2282 digitized film and mammograms 324 DBT volumes were collected IRB approval. The mass of interest on the images was marked by an experienced radiologist as reference standard. partitioned into training (2282 2461 230 views 228 masses) independent test (94 89 masses). For DCNN training, region (ROI) (true positive) extracted each image. False positive (FP) ROIs identified at prescreening their previously developed CAD systems. After augmentation, total 45 072 mammographic 37 450 obtained. Data normalization reduction non-uniformity across heterogeneous achieved background correction method applied to ROI. A four layers three fully connected (FC) first trained mammography data. Jittering dropout techniques used reduce overfitting. ROIs, all weights frozen, only last convolution layer FC randomly initialized again ROIs. authors compared performances two systems DBT: one DCNN-based approach other feature-based FP reduction. stage identical both systems, passing same candidates stage. system, 3D clustering active contour segmentation; morphological, gray level, texture features merged linear discriminant classifier score detected masses. five consecutive slices centered candidate passed through likelihood generated. evaluated free-response ROC curves performance difference analyzed non-parametric method.Before learning, AUC 0.99 classified 0.81 set. DBT, improved 0.90. breast-based set, sensitivity 83% 91%, respectively, 1 FP/DBT volume. between statistically significant (p-value < 0.05).The image patterns learned transferred DCNN. This study demonstrated that large sets are useful developing new alleviating problem effort collecting entirely modality.