Figure 2.

Overview of transfer learning framework in our paper. Top row: the CNN-A is pretrained on the ImageNet database for classification, which consists of many convolutional blocks and fully connected layers. Bottom row: after modifying the structure of fully connected layers, the CNN-B model (except fully connected layers) is initialized with the previous trained weights from CNN-A, the first n convolutional blocks of which are locked, while the left are unlocked. Then the entire network is trained on breast ultrasound images to fine-tune the remaining unlocked layers.