Transfer learning in medical image segmentation: New insights from analysis of the dynamics of model parameters and learned representations

作者：

Highlights：

• We critically study and analyze transfer learning in medical image segmentation.

• We show that model weights change little from random initialization during training.

• We show viability of models with random encoders, challenging the established beliefs.

• We study evolution of learned representations, offering alternative training methods.

摘要

•We critically study and analyze transfer learning in medical image segmentation.•We show that model weights change little from random initialization during training.•We show viability of models with random encoders, challenging the established beliefs.•We study evolution of learned representations, offering alternative training methods.

论文关键词：Medical image segmentation,Fully convolutional neural networks,Deep learning,Transfer learning

论文评审过程：Received 16 February 2021, Revised 19 April 2021, Accepted 20 April 2021, Available online 23 April 2021, Version of Record 30 April 2021.

论文官网地址：https://doi.org/10.1016/j.artmed.2021.102078