Fig. 1: Overview of AdaSlide.
From: Adaptive compression framework for giga-pixel whole slide images
A Patch distributions of the PanCancer dataset used to train AdaSlide. Using 930 WSIs from 31 TCGA projects, we extracted 40× and 20× magnification patch images (each magnification has equal proportions). B AdaSlide pipeline. A WSI is tessellated into multiple patch images, and the compression ratios are decided using the CDA. Patch images selected for compression undergo encoding and decoding steps using the FIE. Finally, the patch images excluded from compression, along with the enhanced images, were collected and reconstructed for further analysis. C Overall performance of AdaSlide on 13 datasets. The baseline indicates the conventional supervised learning using original images. The degraded performance implies information loss during compression and enhancement. AdaSlide showed comparable performance to the baseline, whereas the other compression models (ESRGAN, VQVAE, SwinIR, LDM) did not. WSI whole-slide image, TCGA The Cancer Genome Atlas, CDA Compression Decision Agent, FIE Foundational Image Enhancer.
