| Computer-aided diagnosis (CAD) is a major research subject in medical imaging. Despite its proven effectiveness, the degradation of its performance in real clinical workflows prevents its implementation in practice and maintains its effects on the research field. A lot of effort has been put into designing more accurate machine learning models. However, the development of an optimum training strategy is often neglected by CAD designers. The aim of this proposal is to improve CAD system by designing an efficient training strategy using active learning techniques. Active learning has been successfully applied to other real-world domains but it is an unexplored field in medical image processing. With the proposed active learning approach, I aim to (1) efficiently select representative samples from large unlabeled medical image databases, combining exploitative and explorative query sampling strategies; (2) efficiently use the medical expert's time and effort, by means of a batch-mode querying approach and accurate stopping criteria; and (3) efficiently use the available unlabeled and labeled data in medical environments, by combining techniques for active and transfer learning. An evaluation of the proposed methodology will be carried out by assessing the improvement in performance of two different CAD applications: detection of breast cancer in mammograms and detection of lung cancer in chest CT scans. Databases of more than 350,000 mammograms and 40,000 CT scans are available for a large evaluation study of the proposal. The proposal involves a close collaboration with medical experts, providing the ideal framework to investigate the transfer of knowledge to potential users. |
