Rethinking ImageNet Pre-training for Computational Histopathology.

Abstract

Transfer learning from ImageNet pretrained weights is widely used when training Deep Learning models on a Histopathology dataset. However, the visual features of the two domains are different. Rather than ImageNet pretrained weights, pre-training on a Histopathology dataset may provide better initialization. To prove this hypothesis, we train two commonly used Deep Learning model architectures - ResNet and DenseNet on a complex Histopathology classification dataset, and compare transfer learning performance with ImageNet pretrained weights. Based on the fine-tuning on three histopathology datasets including two different stains (H&E and IHC), we show that the domain specific pretrained weights are better suited for transfer learning. This is reflected by higher performance, lower training time as well as better feature reuse. Clinical Relevance - The paper establishes merit of using Histopathology domain specific pretrained weights rather than ImageNet pretrained weights.