Rethinking ImageNet Pre-training for Computational Histopathology

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Deep learning for medical imaging

Background:

• Transfer learning using ImageNet pretrained weights is common for Deep Learning in histopathology.
• Visual features differ significantly between natural images (ImageNet) and histopathology images.
• This domain gap may limit the effectiveness of ImageNet pretraining.

Purpose of the Study:

• To investigate if pretraining Deep Learning models on histopathology data improves transfer learning compared to ImageNet pretraining.
• To evaluate the impact of domain-specific pretraining on model performance, training efficiency, and feature reuse.

Main Methods:

• Trained ResNet and DenseNet architectures on a complex histopathology classification dataset.
• Compared transfer learning performance using ImageNet pretrained weights versus histopathology pretrained weights.
• Fine-tuned models on three distinct histopathology datasets with H&E and IHC stains.

Main Results:

• Histopathology domain-specific pretrained weights consistently outperformed ImageNet pretrained weights.
• Demonstrated higher classification performance with domain-specific weights.
• Observed reduced training times and improved feature reuse when using domain-specific pretraining.

Conclusions:

• Pretraining Deep Learning models on histopathology data provides superior initialization for transfer learning in this domain.
• Domain-specific pretraining is more effective than ImageNet pretraining for histopathology image analysis.
• Clinical relevance: Utilizing histopathology-specific pretrained weights is advantageous over ImageNet weights for improved model performance and efficiency.