Medical images are valuable for clinical diagnosis and decision making. Image modality is an important primary step, as it is capable of aiding clinicians to access required medical image in retrieval systems. Traditional methods of modality classification are dependent on the choice of hand-crafted features and demand a clear awareness of prior domain knowledge. The feature learning approach may detect efficiently visual characteristics of different modalities, but it is limited to the number of training datasets. To overcome the absence of labeled data, on the one hand, we take deep convolutional neural networks (VGGNet, ResNet) with different depths pre-trained on ImageNet, fix most of the earlier layers to reserve generic features of natural images, and only train their higher-level portion on ImageCLEF to learn domain-specific features of medical figures. Then, we train from scratch deep CNNs with only six weight layers to capture more domain-specific features. On the other hand, we employ two data augmentation methods to help CNNs to give the full scope to their potential characterizing image modality features. The final prediction is given by our voting system based on the outputs of three CNNs. After evaluating our proposed model on the subfigure classification task in ImageCLEF2015 and ImageCLEF2016, we obtain new, state-of-the-art results—76.87% in ImageCLEF2015 and 87.37% in ImageCLEF2016—which imply that CNNs, based on our proposed transfer learning methods and data augmentation skills, can identify more efficiently modalities of medical images.
展开▼
