Unsupervised Pre-training Across Image Domains Improves Lung Tissue Classification

The detection and classification of anomalies relevant for disease diagnosis or treatment monitoring is important during computational medical image analysis. Often, obtaining sufficient annotated training data to represent natural variability well is unfeasible. At the same time, data is frequently collected across multiple sites with heterogeneous medical imaging equipment. In this paper we propose and evaluate a semi-supervised learning approach that uses data from multiple sites (domains). Only for one small site annotations are available. We use convolutional neural networks to capture spatial appearance patterns and classify lung tissue in high-resolution computed tomography data. We perform domain adaptation via unsupervised pre-training of convolutional neural networks to inject information from sites or image classes for which no annotations are available. Results show that across site pre-training as well as pre-training on different image classes improves classification accuracy compared to random initialisation of the model parameters.