Multi-level progressive transfer learning for cervical cancer dose prediction

Recently, deep learning has accomplished the automation of radiation therapy planning, enhancing its quality and efficiency. However, such progress comes at the cost of a large amount of clinical data. For some low-incidence cancers, i.e., cervical cancer, with limited available data, current data-hungry deep models fail to achieve satisfactory performance. To address this, in this paper, considering that cervical cancer and rectum cancer share the same scanning area and organs at risk (OARs), we resort to trans-fer learning to transfer the knowledge acquired from rectum cancer (source domain) to cervical cancer (target domain) to perform dose map prediction task. To overcome the possible negative transferring problem, we design a two-phase paradigm to progressively transfer knowledge. In the first phase, we ag-gregate the data of the two domains by linear interpolation and pre-train an aggregated network with the aggregated data to perceive the target dose distribution beforehand. In the second phase, we elaborately design two modules, i.e., a Feature-level Transfer (FT) Module, and an Image-level Transfer (IT) Module, to selectively transfer knowledge in multi-level. Specifically, the FT module aims to preserve those fil-ters that are more helpful while the IT module tries to highlight those samples with more target-specific knowledge. Extensive experiments proclaim the exemplary performance of our proposed method com-pared with other state-of-the-art methods.(c) 2023 Elsevier Ltd. All rights reserved.