Iris Digital Pathology Deterministic Network Image Compression

TECHNOLOGY NUMBER: 2025-482
Technology No. 2025-482

OVERVIEW

Advanced generative neural network compression pipeline for whole slide digital pathology imaging.
  • Reduces computational loads, storage costs, and network streaming overhead significantly over standard codecs. 
  • Efficiently manages whole slide image archiving and real-time viewing in digital pathology workflows.


BACKGROUND

Digital whole slide imaging generates files often exceeding 1.5 GB per slide, with thousands of pixels in each dimension. Tertiary medical centers accessioning hundreds of thousands of annual surgical cases face immense operational challenges. Storing these massive files requires petabytes of new storage capacity each year, creating severe computational, storage, and financial burdens. Modern conventional file structures rely on standard compression techniques like JPEG, J2K, WebP, and HEIC. These methods utilize frequency decomposition or regional feature encoding while remaining mostly agnostic of global redundant visual information. Conventional codecs fail to exploit the unique properties of digital pathology images, such as their limited dichromatic vocabulary and repeating tissue textures. Furthermore, non-linear and non-sequential slide evaluation by pathologists causes significant GPU, CPU, and network resource lags. Consequently, an improved compression methodology is needed to reduce data overhead without sacrificing critical diagnostic visual detail.


INNOVATION

The proposed technology is a deterministic generative neural network compression pipeline that utilizes super-resolution convolutional neural networks (SRCNN) and adjacent feature networks (AFCNN). It processes whole slide images organized in resolution pyramids and fixed-dimension tiles. The system takes a low-resolution image input, predicts higher-resolution subregions, and incorporates stochastic residuals to correct errors, enabling both minimally lossy and fully lossless decompression modes. By leveraging the repetitive histological textures and dichromatic characteristics of specific tissue types, the predictive networks achieve exceptionally high compression rates. Residual corrections are iteratively applied across subregions and zoom layers to refine consecutive predictions. This framework streams compressed byte streams directly to the convolutional neural network for on-the-fly rendering. Real-world applications include high-speed slide rendering, reduced streaming latency, lower hardware processing requirements, and substantial enterprise-level digital pathology storage cost savings.


ADDITIONAL INFORMATION

INTELLECTUAL PROPERTY

Patent Pending


RELATED TECHNOLOGY


  • expand_more mode_edit Inventor (2)
    Ryan Landvater
    Ulysses Balis
  • expand_more cloud_download Supporting documents (1)
    Product brochure
    Iris Digital Pathology Deterministic Network Image Compression.pdf
Questions about this technology?