Knowledge transfer between brain lesion segmentation tasks with increased model capacity.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Convolutional neural networks (CNNs) are effective for brain lesion segmentation but require extensive annotated data.
• Limited annotated data is a significant challenge in medical imaging for training CNNs.
• Knowledge transfer via pretraining and fine-tuning is a common strategy to address data scarcity.

Purpose of the Study:

• To improve brain lesion segmentation accuracy when training data is limited.
• To adapt the strategy of fine-tuning with increased model capacity for segmentation tasks.
• To develop a novel spatially adaptive augmentation strategy for fine-tuning CNNs.

Main Methods:

• Extended fine-tuning with increased model capacity (width augmentation) for brain lesion segmentation.
• Developed a spatially adaptive augmentation module to enhance fine-tuning effectiveness.
• Applied the proposed method to ischemic stroke lesion segmentation using a model pretrained on brain tumor segmentation.

Main Results:

• The proposed method, incorporating width augmentation and spatial adaptivity, demonstrated benefits in brain lesion segmentation.
• Fine-tuning with increased model capacity improved adaptation to the target segmentation task.
• Spatially adaptive augmentation further enhanced the performance over vanilla width augmentation.

Conclusions:

• Fine-tuning CNNs with increased model capacity is a viable strategy for brain lesion segmentation with limited data.
• The developed spatially adaptive augmentation method offers improved performance for segmentation tasks.
• The approach shows promise for applications like ischemic stroke lesion segmentation.