1. Introduction

• 적은 annotation의 이미지로 diagnostic accuracy를 개선할 수 있을까?
• Semi-supervised Learning (SSL)
  • an auxiliary loss term defined on unlabeled data (consistency regularization term)
  • linear interpolations of both labeled and unlabeled data (MixUp augmentation)
• FocalMix
  • a generic generalization of the focal loss that allows the usage of soft-target training labeles with skewed distributions
  • a target-prediction strategy that leverages anchor-level ensembles of augmented image patches by rotation and flipping
  • MixUp augmentation
  • MixMatch 사용

2. Background and Preliminaries

2.1. Object Detection in 3D Medical Images

2.1.1. Anchor boxes

• target object를 맞추기 위해 tiling된 이미지의 predefined bounding boxes

2.1.2. Focal Loss

• foreground-background imbalance를 완화하기 위해

  $$ FL(p_t)=-\alpha_t(1-p_t)^\gamma\log(p_t) \tag{1} $$

  $$ \begin{cases} p &\text{if }y=1 \\ 1-p &\text{otherwise.} \end{cases} \tag{2} $$

2.2. Semi-supervised Learning

• 이 연구는 MixMatch를 기반으로 진행

  • entropy minization, consistency regularization, MixUp augmentation 등 좋은 것들은 다 넣음

  • prediction의 average ensemble 사용

    $$ \bar{y}=\frac{1}{K}\sum^K_{k=1}\text{p}_\text{Model}(\hat{u}_k;\theta) \tag{3} $$

    • $\bar{y}$ : guessed labeles
    • $\theta$ : parameters of the current model
    • $\hat{u}_k$ : unlabeled training sample $u$에 대해 K번의 augmentation

  • guessed label을 sharpening operator로 한 번 더 변환

    $$ \text{Sharpen}(\bar{y},T)_i=\bar{y}^{\frac{1}{T}}i/\sum^L{j=1}\bar{y}^{\frac{1}{T}}_j \tag{4} $$

    • $T$ : temperature which controls the smoothness of the output distribution

  • MixUp augmentation으로 또 다른 training example $(x',y')$을 생성

    $$ \begin{aligned} &\lambda\sim\text{Beta}(\eta,\eta) &&&&&&&&&&&&&&&&(5)\\ &\tilde{\lambda}=\text{max}(\lambda,1-\lambda) &&&&&&&&&&&&&&&&(6)\\ &\hat{x}=\tilde{\lambda}x+(1-\tilde{\lambda})x' &&&&&&&&&&&&&&&&(7)\\ &\hat{y}=\tilde{\lambda}y+(1-\tilde{\lambda})y' &&&&&&&&&&&&&&&&(8) \end{aligned} $$

3. Methodology

• target prediction + MixUp augmentation

3.1. Soft-target Focal Loss

3.2. Anchor-level Target Prediction

3.3. MixUp Augmentation for Detection

4. Experiments