Keywords: Transfer learning, histopathology, segmentation
TL;DR: Pre-training on histopathology datasets may significantly improve segmentation performance relative to ImageNet pre-trained weights.
Abstract: In limited data settings, transfer learning has proven useful in initializing model parameters. In this work, we compare random initialization, pre-training on ImageNet, and pre-training on histopathology datasets for 2 model architectures across 4 segmentation histopathology datasets. We show that pre-training on histopathology datasets does not always significantly improve performance relative to ImageNet pre-trained weights for both model architectures. We conclude that unless larger labeled datasets or semi-supervised techniques are leveraged, ImageNet pre-trained weights should be used in initializing segmentation models for histopathology.