MLX-Radiomics: accelerating standardized regional and voxel-wise radiomics for CT, PET and MRI

Background: Differences among radiomics implementations complicate cross-study comparisons, while voxel-wise feature extraction remains computationally demanding. We developed MLX-Radiomics to accelerate regional feature extraction and voxel-wise mapping while preserving agreement with standardized reference measurements. Methods: MLX-Radiomics implements 172 feature definitions from the Image Biomarker Standardization Initiative (IBSI), 16 extensions and the IBSI 2 image filters. We assessed agreement with IBSI reference data and evaluated numerical accuracy against independent double-precision and analytic calculations. In matched comparisons on Apple M1 Max and M5 Max systems, we used the IBSI reference CT to compare feature values, extraction time and memory use with PyRadiomics, MIRP and Flash-Radiomics, and feature values with FastRad. Separately, we compared the time required to generate voxel-wise feature maps for regions of up to 10.1 million voxels with published GPU results obtained on different hardware and images. We evaluated agreement with MIRP under published settings in a 51-patient cohort with CT, FDG PET and T1-weighted MRI. Results: MLX-Radiomics agreed with all 2,160 IBSI reference measurements within the specified tolerances. With matching outputs, MLX-Radiomics was 2.66 to 7.27 times faster than PyRadiomics and MIRP for regional feature extraction and 198 to 218 times faster than PyRadiomics for voxel-wise mapping on a 24 x 24 x 24-voxel crop. It also outperformed the fastest Flash-Radiomics CPU configurations tested. On the M5 Max, generating 93 maps over 10.1 million voxels took 23.8 s. In the patient cohort, MLX-Radiomics and MIRP agreed on 98.0% of requested values. For tumors at the image edge, matching the libraries' resampling rules substantially improved agreement. Conclusions: MLX-Radiomics provides standardized regional measurements and voxel-wise maps for CT, PET and MRI, with evidence of faster extraction limited to CT. It can support quantitative imaging research on relatively inexpensive hardware without requiring a separate GPU workstation.