[P] fastrad: GPU-native radiomics library — 25× faster than PyRadiomics, 100% IBSI-compliant, all 8 feature classes

Developer Erika has released fastrad, an open-source radiomics library that leverages PyTorch's GPU acceleration to process medical imaging data 25 times faster than the industry-standard PyRadiomics. While PyRadiomics has been the de facto tool for extracting quantitative features from CT and MRI scans, its CPU-only architecture creates bottlenecks at scale, taking approximately 3 seconds per scan. Fastrad reimplements all eight IBSI-compliant feature classes—including first-order statistics, GLCM, GLRLM, and shape descriptors—as native tensor operations that automatically route to available GPUs.

On an RTX 4070 Ti GPU, fastrad processes scans in just 0.116 seconds compared to PyRadiomics' 2.9 seconds, with per-class speedups ranging from 12.9x to 49.3x. Remarkably, even in single-threaded CPU mode, it outperforms PyRadiomics' 32-thread implementation by 2.63x on x86 and 3.56x on Apple Silicon. The library maintains exceptional accuracy, validating against the IBSI Phase 1 digital phantom with maximum deviation below 10⁻¹³% and matching PyRadiomics' outputs on real NSCLC CT scans within 10⁻¹¹ tolerance.

Available now on GitHub with a pre-print detailing the implementation, fastrad represents a significant leap for medical imaging research. By dramatically reducing computation time while maintaining compliance with international radiomics standards, it enables researchers to process larger datasets and iterate more quickly on cancer detection algorithms, potentially accelerating discoveries in personalized medicine and treatment response prediction.