Room: Exhibit Hall
Purpose: The recent trend in medical image understanding demands that more descriptive information be extracted from the image for use in training machine learning models to perform clinically relevant inference. The use of local features is becoming more important as researchers look inside tumors to understand their composition and gain more insight into the threats presented. There have been many tools released to date that automate this process for region-based features, but no efficient tools are available that perform this processing on a local scale.
Methods: A GPU accelerated, local (patch-based) feature calculation toolkit has been developed that parallelizes the task of local histogram construction and feature calculation for commonly used radiomics features. A python library was implemented with a CUDA backend, supporting flexible calculation of first order statistical features (mean, variance, etc.) and many 2D histogram features in the GLCM, GLRLM, and GLSZM families. Support for 2D local support and 3D local support in histogram construction is included offering slice-by-slice or volumetric feature calculation. Multi-GPU support is built-in, greatly increasing the efficiency of large-batch feature calculation. Timing comparisons for a reference CPU and the proposed GPU implementation have been performed on natural and medical images.
Results: Calculation of GLCM entropy for a 512x512 natural image with a patch size of 11x11 and a stride of 1 requires 75 seconds for the CPU implementation and only 1.2 seconds for the proposed GPU implementation. For GLCM entropy on a 512x512x120 chest CT image with a patch size of 11x11 and stride of 1, the CPU method requires 1.32 hours while our multi-GPU method only requires 56 seconds.
Conclusion: We have developed an extremely efficient toolkit to meet the needs of local radiomics feature extraction using a multi-GPU approach that greatly outperforms a standard CPU implementation for 2D and 3D images.
Funding Support, Disclosures, and Conflict of Interest: This research is supported by the following grants: NIH U19AI067769 DE-SC0017687 NIH R21CA228160 DE-SC0017057 NIH R44CA183390 NIH R43CA183390 NIH R01CA188300