Skip to content

[Showcase/Alternative] We achieved 25x GPU-accelerated parity using native PyTorch #947

Description

@helloerikaaa

Hi PyRadiomics team!

First, thank you for building the gold standard in radiomics. We've relied heavily on the mathematical rigor and IBSI compliance of PyRadiomics for our research.

We noticed that a lot of users (including ourselves) run into computational bottlenecks when scaling to massive clinical datasets on CPU. To help solve this, we recently built an open-source library called FastRadiomics (fastrad). It maps the exact mathematical definitions of PyRadiomics into native dense PyTorch tensors (torch.float64) for CUDA/GPU hardware acceleration.

Through rigorous testing, we've successfully achieved 100% compliance with the IBSI digital phantom and PyRadiomics outputs (within a 1e-4 tolerance) while getting up to a 25x speedup on NVIDIA GPUs without needing custom C++ kernels.

We wanted to share this with the maintainers in case you ever investigate a PyTorch/GPU backend, or if you'd be open to linking it in the README as a high-speed alternative for users hitting CPU limits! We'd love your feedback.

Preprint in SSRN

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions