About

Medical Image Classification : scarce source data to train the model → Use Transfer Learning

Transfer Learning?

부족한 데이터로 하는 머신러닝! '전이 학습'

1. First training on a larger, source dataset(natural or medical images)
2. And then further fine-tuning on a smaller, target dataset(medical images)

⇒ Can reuse features learned by source data, smaller amount of target data needed

Related Work

• ImageNet is a popular source dataset, but medical과 property 차이가 너무 크다
• 몇 연구는 source와 target 간 차이가 클수록 feature reuse의 효과성이 감소된다고 한다
• 일반적으로 large enough and similar enough source data, but exact definition is not mentioned
• All of them use different dataset. Hard to compare

Datasets

Experiment