Two-Phase Progressive Deep Transfer Learning for Cervical Cancer Dose Map Prediction

Recently, deep learning has enabled the automation of radiation therapy planning, improving its quality and efficiency. However, such progress comes at the cost of amounts of data. For some low incidence cancers, e.g., cervical cancer, the available data is limited, which could degrade the performance of conventional deep learning models. To alleviate this, in this paper, we resort to transfer learning to accomplish the task of dose prediction on a small amount of cervical cancer data. Considering the same scanning areas of the cervical cancer and the rectum cancer and their shared organs at risk, we are inspired to transfer the knowledge learned from rectum cancer (source domain) to cervical cancer (target domain). Specifically, to narrow the huge gap between the source domain and the target domain, we propose a two-phase transfer strategy. Firstly, we aggregate the data distributions of two domains by linear interpolation, and train an aggregated network to perceive the target domain in advance. Secondly, we transfer the knowledge from the well-trained aggregated network to the target network through an innovatively designed Weighted Feature Transfer Module (WFTM), thus ensuring that the target network can learn more valuable knowledge. Experimental results on 130 rectum cancer patients and 42 cervical cancer patients demonstrate the effectiveness of our method.