Introduction

FastRadiomics (fastrad) is a modern, high-performance Python library designed for extracting radiomics features from medical images.

Built as a direct, highly optimized alternative to the widely used PyRadiomics library, fastrad achieves strict computational parity with PyRadiomics while enabling dramatically faster execution times. By implementing image feature extraction natively in PyTorch, fastrad bypasses the limitations of traditional CPU-bound scalar operations.

Why fastrad?

Traditional radiomics libraries, while scientifically robust and compliant with the Image Biomarker Standardisation Initiative (IBSI), often struggle to scale when processing large multi-modal datasets or high-resolution volumetric scans (like 512x512xN CT scans). This bottleneck can bring modern AI and machine learning pipelines to a halt.

fastrad solves this by treating radiomics feature extraction as a hardware-accelerated tensor problem:

  • Full Parity: Every single feature has been rigorously tested against PyRadiomics to ensure identical values extracted at high precision. You can swap fastrad into your existing scientific pipelines without degrading experimental validity.

  • Hardware Acceleration: Under the hood, fastrad leverages PyTorch block processing. This allows it to natively route extraction workloads across multi-core CPUs, CUDA-enabled NVIDIA GPUs, and even Apple Silicon (MPS).

  • Scalable Matrix Construction: Texture matrices (GLCM, GLRLM, GLSZM, GLDM, NGTDM) are notoriously slow to compute sequentially. fastrad heavily vectorizes these constructions.

Use Cases

  • High-Throughput Research: Compute features over cohorts of thousands of patients in a fraction of the time.

  • Real-Time Inference: Embed radiomics extractors seamlessly inside end-to-end PyTorch deep learning training loops or clinical deployment pipelines without facing severe CPU throttling.

  • Reproducible Science: Lean on our extensive unit-tests and stability analysis tools simulating adversarial 3D affine transformations to guarantee robust feature replication.

Note: Since statistical textural features demand extremely high 64-bit precision, CUDA or CPU are the primary recommended execution targets for perfect parity, as some PyTorch MPS (Apple Silicon) operations do not natively support FP64 operations yet.