Differentiable Zooming for Multiple Instance Learning on Whole-Slide Images