HCTTI: High-Performance Heterogeneous Computing Toolkit for Tissue Image Stain Normalization

Whole slide imaging (WSI) has transformed diagnostic medicine, particularly in the field of cancer diagnosis and treatment. The use of deep learning algorithms for predicting WSIs has opened up new avenues for advanced medical diagnostics. Additionally, stain normalization can reduce the color and intensity variations present in WSI from different hospitals. As a result, deep learning classification accuracy improves. However, WSI reading and color normalization are still largely performed by using CPUs, leading to sub-optimal performance. We proposed a High-Performance Heterogeneous Computing Toolkit for Tissue Image (HCTTI) that integrates multiple computer system-level optimizations and encompasses WSI reading, tile normalization, and tile saving. We explored the potential advantages and limitations of different WSI readers and color normalization techniques in WSI analysis and the performance of different tile serialization formats for saving tiles. We found that HCTTI is 7 $\times$ faster than OpenSlide for reading WSIs, GPU implementation of the Macenko normalization algorithm is 9 $\times$ faster than TIAToolbox implementation, and HDF5 is faster than png and Zarr for storing normalized images in both writing (13 $\times$ acceleration compared to png) and reading (2 $\times$ acceleration compared to png), Specifically, HDF5 provides superior performance in handling large, complex datasets due to its efficient chunking and compression capabilities, as well as its broad support for hierarchical data management, making it very suitable for workloads like deep learning training I/O pattern that involves randomly reading large amount of small files in each training epoch. We also achieved linear acceleration in our multi-node distributed GPU implementation. To our knowledge, HCTTI is the first comprehensive toolkit that comprises distributed WSI reading, normalization, and serialization. It is 13 $\times$ speedup compared to TIAToolbox implementation for normalizing a single WSI. Our findings could help pave the way for more effective and efficient deep learning-based approaches to WSI analysis, with the potential to transform medical diagnosis and treatment for a wide range of conditions. The source code of HCTTI is available at https://github.com/wangbo00129/HCTTI .