AdaSlide: adaptive compression framework for digital pathology slides

초록

Background & objectives: Digital pathology images require large amounts of storage space, naturally leading to cost issues. Here, we propose a compressing modality for whole slide images that adaptively employs the compression ratio based on information across regions of slides. Methods: AdaSlide, an Adaptive compression framework for digital pathology Slides, consists of a compression decision agent (CDA) and a foundational image enhancer (FIE). The CDA utilizes reinforcement S380 Virchows Archiv (2024) 485 (Suppl 1):S1–S546 1 3 learning to evaluate each patch’s necessity and degree of compression, ensuring minimal information loss and maintaining diagnostic integrity. The FIE, trained on diverse cancer types and magnifcations, ensures high-quality image restoration post-compression. Results: We assembled the PanCancer dataset from 30 TCGA projects, comprising 1.8 million patches extracted from 930 WSIs. Using this dataset, we trained two key modules: the FID and the CDA. The FID enhanced compressed images through encoding and decoding steps, while the CDA autonomously determined the compression level based on the image’s information content. We evaluated AdaSlide’s performance across various downstream tasks, including patch-level classifcation, segmentation, and slide-level classifcation. The results indicated minimal degradation in prediction performance when comparing pre- and post-AdaSlide outputs, demonstrating the efectiveness of AdaSlide in maintaining prediction accuracy while reducing image size. Conclusion: This framework paves the way for more efcient storage and transmission of digital pathology data without compromising diagnostic utility.