TPT Spectra

Open-source, GPU-accelerated DICOM reconstruction engine, written in Rust.

TPT Spectra turns raw medical sensor data (CT, MRI, Ultrasound, PET) into 3D diagnostic voxel volumes in seconds, then runs third-party AI diagnostic models directly on the reconstructed data.

It is part of the TPT ecosystem: tpt-healthcare-nz stores the FHIR patient record, while TPT Spectra manages the imaging machines and pushes reconstructed images and AI confidence scores back into the patient's record.

License

Dual-licensed under either of:

• MIT License (LICENSE-MIT)
• Apache License, Version 2.0 (LICENSE-APACHE)

at your option. Copyright (c) TPT Solutions.

Architecture

+-------------------------------------------------------+
|                      spectra-cli                       |
+-------------------+-----------------------------------+
|   dicom-parser    |  Zero-alloc DICOM file parsing     |
|   recon-core      |  Shared volume model + Modality    |
|   gpu-backend     |  wgpu compute (FBP, iterative)     |
|   ai-bridge       |  ONNX Runtime model registry       |
+-------------------------------------------------------+

Crate	Responsibility
dicom-parser	DICOM tags, datasets, file meta, transfer syntaxes, pixel data
recon-core	Shared volume data structures and the Modality trait
gpu-backend	wgpu device/pipeline init and compute-shader reconstruction
ai-bridge	ONNX Runtime bindings, voxel→tensor adapter, model registry
cli	End-to-end parse → reconstruct → AI command-line tool

Current Status

Active development. Phases 0–5 are implemented and the workspace builds and passes cargo test, cargo fmt --check, and cargo clippy -- -D warnings. See todo.md for the phased roadmap.

• Cargo workspace + crate layout
• Dual-license + README + governance docs + CI
• DICOM parser: data model, file meta, transfer syntaxes, pixel data, writer
• GPU reconstruction backend (wgpu device/pipeline init + WGSL kernels)
• Modality implementations: CT FBP (working), MRI iterative (working), Ultrasound / PET (stubs)
• AI diagnostic bridge (ONNX Runtime, voxel adapter, model registry)
• CLI end-to-end pipeline (parse / reconstruct / diagnose / run)
• Verification: phantom tests, DCMTK parity harness (gated), benchmark, FDA 510(k) groundwork docs

Roadmap gaps

• GPU compute path is scaffolded (device init + WGSL) but the FBP compute dispatch is not yet wired to a full GPU pipeline run.
• Ultrasound and PET reconstruction are stubbed (interface defined, algorithms pending).
• Real ONNX model files (models/*.onnx) and full DCMTK pixel-parity are not bundled; the parity harness is gated on SPECTRA_DCMTK_BIN.

Building

Requirements:

• Rust 1.74+ (install via rustup)
• A GPU with Vulkan / Metal / DX12 support (for gpu-backend)

cargo build --workspace
cargo test  --workspace

Running the CLI

cargo run -p spectra-cli -- parse path/to/study.dcm

Verification & Compliance

Long-term goals (Phase 6) include phantom-image test suites, pixel-perfect parity against DCMTK, and FDA 510(k) validation groundwork. This software is not a certified medical device; do not use it for clinical diagnosis without appropriate regulatory validation.