Multi-Task Pre-Training of Deep Neural Networks for Digital Pathology

Area of Science:

• Digital pathology
• Machine learning
• Computer vision

Background:

• Lack of large-scale datasets in digital pathology hinders model pre-training.
• Existing small and medium-sized datasets are fragmented.
• ImageNet, a large-scale dataset, is commonly used for pre-training but may lack domain specificity.

Purpose of the Study:

• Investigate multi-task learning (MTL) for pre-training digital pathology classification models.
• Develop a transferable model using a unified dataset.
• Evaluate the performance of MTL pre-trained models against ImageNet pre-trained models.

Main Methods:

• Assembled and transformed 22 digital pathology datasets into a unified pool of nearly 900,000 images.
• Proposed a simple architecture and training scheme for MTL.
• Implemented a robust evaluation and selection protocol.
• Compared performance using models as feature extractors and through fine-tuning.

Main Results:

• MTL pre-trained models significantly improved performance over ImageNet pre-trained models as feature extractors on specific target tasks.
• Comparable performance was achieved between MTL and ImageNet pre-trained models.
• Fine-tuning further enhanced performance, mitigating the domain specificity limitations of ImageNet features.
• Both pre-training approaches yielded comparable performance after fine-tuning.

Conclusions:

• Multi-task learning offers a viable strategy for pre-training models in digital pathology.
• MTL effectively leverages diverse, smaller datasets to create transferable representations.
• Fine-tuning is crucial for optimizing performance and adapting pre-trained features to specific downstream tasks.