Few shot learning for brain tumor segmentation
Résumé
Automated brain tumor segmentation from magnetic resonance imaging (MRI) scans is of great significant for brain diagnosis. However, how to segment tumor regions accurately with very limited labeled MRI images is still an appealing challenge. To tackle this issue, the present paper proposes to use few-shot under meta-learning setup. The idea is to exploit knowledge derived from a handful annotated support images during episodic learning to guide the segmentation of query images. Specifically, for each episode, the encoder extract feature maps for the both support and query images. Then, a masked average pooling is performed with the support mask to get the guidance features by only considering the target categories belonging to the support image. We use convolution operation to construct the relationship between the guidance features and query feature maps. With an aim to achieve better generalization on few-shot semantic segmentation, decoder based U-Net architecture is used. The proposed method is applied on the benchmark BraTS2021 dataset for brain tumor segmentation. The experimental results in terms of quantitative and qualitative are satisfactory in generating brain tumor segments. The presented segmentation method might be useful to help doctors perform a subtle diagnosis so that the life expectancy of patients becomes longer.