Transfer learning for brain lesion segmentation via data transfer

Automated brain lesion segmentation using deep learning (DL) is crucial for understanding neurological diseases and aiding in diagnosis. Efficiently training a DL segmentation model often involves transfer learning (TL), where knowledge from a source task helps in the target task. Traditionally, this involves model weight transfer, initializing the target model with weights from a pretrained source model and then fine-tuning it with target data. However, this approach limits direct interaction between source data and the target task. This work introduces a new TL paradigm for brain lesion segmentation called Brain Lesion Transfer (BLeT). Instead of transferring model weights, BLeT directly utilizes source training data by transferring useful information to the target task. For a given target brain lesion segmentation task, it is assumed that annotated data for a similar source task is available. BLeT transfers lesions from the source training data to the target annotated data, creating additional, diverse training images, thereby enhancing the training of the target segmentation model. To address the challenge of different appearances in images from different tasks, BLeT includes a lesion appearance transformation method. This method adjusts the lesions from the source task to be compatible with the target images. Experiments on public datasets demonstrate that BLeT outperforms conventional TL methods based on model weights for brain lesion segmentation.