OR-JRD: Observation Region-wise Just Recognition Distortion Guided Whole Slide Image Compression for Histopathology Classification

This is the authors' abstract. We don't add key points for this paper.

Whole Slide Images (WSIs), as the digital version of pathological tissue sections, are of high resolution with gigapixel size. The size of WSI brings significant computational burden not only on storage and transmission, but also on histopathology analysis, e.g., histopathology classification. Therefore, we propose an efficient WSI compression framework to minimize the bit rate while maintaining the classification performance of compressed histopathology images, so as to advance the digitalization of histopathology. Firstly, as pathologists tend to follow the “localization-then-diagnosis” paradigm in the diagnosis, we divide the pathology image into non-overlapping patches and introduce an Observation Region (OR) selection module based on Reinforcement Learning (RL) to identify the patches that are interesting to the pathology classification. Secondly, we explore the minimum distortion in the selected OR that affect the accuracy of pathology classification by predicting the OR-wise Just Recognizable Distortion (OR-JRD) threshold. We construct a OR-JRD dataset with multi-level compressed OR and corresponding OR-JRD thresholds, i.e., Quantization Parameter (QP) in the HEVC standard. An OR-JRD threshold predictor is trained from the constructed OR-JRD dataset. Finally, we apply the predicted OR-JRD on compressing ORs and explore the effect of the compression and transmission setting of patches besides ORs on the pathology classification performance. Experimental results show that the proposed framework can significantly reduce bitrate while maintaining or improving the classification performance compared to the HEVC test model version 18.0. This validate the effectiveness and practicality of the proposed compression framework for pathology classification.