An Empirical Study of Transfer Learning for Colorectal Polyps Image Segmentation

Abstract

Deep learning methods for medical image segmentation typically rely on pretrained models developed for natural images. The tremendous success of transfer learning raises the question: what makes a pretrained model good for medical image segmentation? In this paper, we explore properties of pretrained models on medical image segmentation. We compare the model performance on a polyp segmentation dataset and find that both the choice of network architecture and pretraining dataset are critical to the model’s transferability, while larger network does not always result in superior transfer learning performance.