Adaptive Transfer Learning To Enhance Domain Transfer In Brain Tumor Segmentation

Supervised deep learning has greatly catalyzed the development of medical image processing. However, reliable predictions require a large amount of labeled data, which is hard to attain due to the required expensive manual efforts. Transfer learning serves as a potential solution for mitigating the issue of data insufficiency. But up till now, most transfer learning strategies for medical image segmentation either fine-tune only the last few layers of a network or focus on the decoder or encoder parts as a whole. Thus, improving transfer learning strategies is of critical importance for developing supervised deep learning, further benefits medical image processing. In this work, we propose a novel strategy that adaptively fine-tunes the network based on policy value. Specifically, the encoder layers are fine-tuned to extract latent feature followed by a fully connected layer that generates policy value. The decoder is then adaptively fine-tuned according to these policy value. The proposed approach has been applied to segment human brain tumors in MRI. The evaluation has been performed using 769 volumes from public databases. Domain transfer from T2 to T1, T1ce, and Flair shows state-of-the-art segmentation accuracy.