Domain- and task-specific transfer learning for medical segmentation tasks.

Area of Science:

• Medical imaging analysis
• Machine learning in healthcare
• Deep learning for segmentation

Background:

• Transfer learning is crucial for training convolutional neural networks (CNNs) in medical image segmentation with limited datasets.
• Both source task and domain significantly impact transfer learning performance in medical image segmentation.

Purpose of the Study:

• To evaluate the performance of transfer learning for medical image segmentation across various source task and domain combinations.
• To identify optimal pre-training strategies for CNN-based medical image segmentation.

Main Methods:

• Pre-trained CNNs on classification, segmentation, and self-supervised tasks using natural images and T1 brain MRI domains.
• Fine-tuned pre-trained CNNs on three target T1 brain MRI segmentation tasks: stroke lesions, MS lesions, and brain anatomy.
• Evaluated segmentation accuracy using mIOU or Dice coefficients and detection accuracy for lesion tasks.

Main Results:

• Pre-training on a segmentation task within the same domain as the target task yielded superior or comparable segmentation accuracy.
• Pre-training on ImageNet showed similar lesion detection rates compared to same-domain pre-training, despite larger dataset size.

Conclusions:

• Optimal transfer learning for medical image segmentation requires pre-training on similar tasks and domains.
• CNNs can be effectively pre-trained on smaller datasets by aligning source and target domains and tasks.