Few-Shot Tumor Detection via Feature Reweighting and Knowledge Transferring

Tumor detection is a fundamental and difficult task for computer aided diagnose system. Due to the morphology of tumors varies greatly, it is difficult to train a general tumor detector for all kinds of tumors. Existing tumor detection method always train a special tumor detector in a supervision way. However, training a supervised tumor detector always needs many annotated samples. It is difficult and costly to obtain annotated samples in medical image domain, especially for rare tumor. To overcome these problems, we address the few-shot tumor detection with limited annotated samples. In this paper, we present a few-shot tumor detection method, which extracts a latent task representation from few supervisions, and optimize the architecture of the detector end-to-end for efficient and accurate few-shot object detection. Our method can transfer to new type of tumor without further optimization and quickly update when given few more samples. We reported the results for three type of tumors detection from one annotation per class and show motivation to two other kinds of tumor detection. Our approach reweights features across space for new class tumor. The result show that our method has an acceptable accuracy and transportability.