| Over the past decade, a machine learning technique called deep-learning has gained prominence in computer vision because of its ability to extract semantics from natural images. However, in contrast to the natural images, deep learning methods have been less effective for analyzing medical histology images. Analyzing histology images involves the classification of tissue according to cell types and states, where the differences in texture and structure are often subtle between states. These qualitative differences between histology and natural images make transfer learning difficult and limit the use of deep learning methods for histology image analysis.;This dissertation introduces two novel deep learning architectures, that address these limitations. Both provide intermediate hints to aid deep learning models. The first deep learning architecture is constructed based on stacked autoencoders with an additional layer, called a hyperlayer. The hyperlayer is an intermediate hint that captures image features at different scales. The second architecture is a two-tiered Convolutional Neural Networks (CNN), with an intermediate representation, called a pixel/region labeling. The pixel/region labels provide a normalized semantic description that can be used as an input to a subsequent image classifier.;The experiments show that by adding the hyperlayer, the architecture substantially outperforms fine-tuned CNN models trained without an intermediate target. In addition, the experiments suggest that the advantages of the labeling classifier are threefold. First, it generalizes to other related vision tasks. Second, image classification does not require extremely accurate pixel labeling. The architecture is robust and not susceptible to the noise. Lastly, labeling model captures low-level texture information and converts them to valuable hints. |
