RRT MIL

Feature Re-Embedding: Towards Foundation Model-Level Performance in Computational Pathology

data:

CAMELYON-16 [1] (C16)、TCGA-BRCA和TCGA-NSCLC

TCGA-LUAD、TCGA-LUSC 和 TCGA-BLCA

code: Github

方法

\(\begin{equation*} \hat{Y}\succ \mathcal{M}(X):=\mathcal{C}(\mathcal{A}(\mathcal{R}(\mathcal{F}(X)))), \tag{2} \end{equation*}\)

创新点就是R()重新嵌入的映射函数

\(\begin{align*} & \boldsymbol{ Step 1: } H \in \mathbb{R}^{I \times D} \xrightarrow{\text{Squaring }} H \in \mathbb{R}^{L^2 \times M^2 \times D}, \\ & \boldsymbol{ Step 2: } H \xrightarrow{\text{Partition }}\left\{H^l\right\}_{l=1}^{L^2}, H^l \in \mathbb{R}^{M \times M \times D},\\ & \boldsymbol{ Step 3: } \hat{Z}:=\left\{\hat{Z}^l\right\}_{l=1}^{L^2}, \hat{Z}^l=\mathcal{S}\left(H^l\right) \in \mathbb{R}^{M \times M \times D}, \tag{5} \end{align*}\)

s()是EPEG

结果

参考文献

Tang, Wenhao et al. “Feature Re-Embedding: Towards Foundation Model-Level Performance in Computational Pathology.” 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (2024): 11343-11352.