Multiple Instance Captioning: Learning Representations From Histopathology Textbooks and Articles

Figure 1: Four samples from ARCH, a multiple instance captioning computational pathology dataset. Samples on the left and right each consist of four image instances with a single caption; top-middle shows an image-caption pair while bottom-middle contains two image instances with a single caption. Labeled in color are examples of common tasks within computational pathology: diagnostic (orange); detection & classification (cyan); descriptive (yellow); special cell detection (red).