MAR ILS Dataset Generator & Evaluation Framework

Overview

This is the reference evaluation framework for conducting standardized Interlaboratory Studies (ILS) on CT Metal Artifact Reduction (MAR) algorithms. Following the task-based signal detection framework of Vaishnav et al. (Medical Physics, 2020), this toolset provides:

1. Standardized Dataset Generation: Physics-based fan-beam sinogram synthesis (v7.0.0).
2. Reference MAR Algorithms: Implementations of iMAR, MBIR, and Dual-Energy Spectral MAR.
3. Objective Statistical Scoring: Quantifying lesion detectability ($\Delta AUC$) using a Channelized Hotelling Observer (CHO) with normative internal noise ($\sigma=15$).
4. AI-Native Diagnostics: Integrated Model Context Protocol (MCP) servers for automated data auditing and visualization.

---

April 2026 Status: Metrology Baseline Locked

The framework has officially transitioned from parallel-beam research tiers to a Normative Fan-Beam Geometry (SID=570mm, SDD=1040mm).

ILS Reference Baseline ($N=40$)

The following values are established as the normative floor for ASTM WKXXXXX Rev 04:

Metric	Value	Status
Baseline AUC_noMAR	0.8294	LOCKED
95% Confidence Interval	[0.7612, 0.9025]	Verified
Lesion Contrast	12.0 HU	Normative
Integrity Check	SHA-256	Validated

---

Performance Breakthrough: Numba Acceleration

The v7.0.0 engine features a JIT-compiled Batch Backprojector optimized for Apple Silicon (M3) and modern x86_64 architectures, providing a 24x speedup over standard NumPy implementations.

• Throughput: ~61s per realization (256 slices).
• Efficiency: Full $N=40$ dataset generation in <30 minutes on standard workstations.

Quick Start

# Generate the canonical fan-beam dataset
python generator_v7_0_0.py --output-dir ./astm_reference_dataset

# Self-test (validates CHO pipeline — ΔAUC = 0 by definition)
python run_cho_analysis_v7_0.py \
    --dataset-dir ./astm_reference_dataset \
    --self-test --internal-noise-sigma 15

# ILS mode (score your MAR algorithm)
python run_cho_analysis_v7_0.py \
    --dataset-dir ./astm_reference_dataset \
    --mar-output-dir ./mar_recon \
    --internal-noise-sigma 15

---

AI-Integrated Laboratory (MCP)

This repository includes Model Context Protocol (MCP) servers, allowing AI assistants to act as "Laboratory Hands" on your local workstation.

• Data Inspector (mcp_data_inspector.py): Automatically parses CHO JSON results and summarizes statistical significance.
• Visualization (mcp_visualization.py): Renders HDF5 sinogram slices and ROI comparisons directly in the AI chat interface.

---

Validation & Audit Tools

To ensure inter-laboratory consistency, use the following in-memory physics auditors:

• plot_spectral_transparency.py: Visualizes the "Transparency Jump" between 60 keV and 140 keV.
• check_metal_overflow.py: Audits the normative 3000 HU metal-ROI hard-set requirement.

---

Deliverables & Repository Structure

Path	Description
generator_v7_0_0.py	Normative fan-beam dataset generator (Rev 04).
run_cho_analysis_v7_0.py	Normative 2D CHO scoring tool (Rev 04).
patch_2026b_metadata.py	One-time DICOM 2026b CP-2575 metadata patcher.
docs_and_references/ASTM_MAR_Standard.md	Draft standard text (Rev 04, machine-readable)
docs_and_references/IEC_203_6_7_101_compliance_statement_proposal.md	Proposed §203.6.7.101.1 compliance-statement amendment binding ASTM FXXXX as the TYPE TEST
docs_and_references/FDA_guidance_framework.md	Draft FDA guidance for acceptance criteria
/algorithms	Reference MAR implementations (iMAR, MBIR, Spectral).
/legacy	Archived v6.0.0 parallel-beam research framework.

---

Metadata Standard: DICOM 2026b Compliant

This is the first reference implementation of the DICOM CP-2575 Metal Artifact Reduction Macro (PS3.3 C.8.15.3.15), finalized in DICOM 2026b.

All DICOM files produced by the generator and the patch_2026b_metadata.py utility include:

Tag	Keyword	Value
(0018,9390)	Metal Artifact Reduction Sequence	Present (1 item)
(0018,9391)	Metal Artifact Reduction Applied	NO

The noMAR reconstructions are tagged NO by definition. Laboratories applying MAR algorithms must set (0018,9391) to YES and optionally populate (0018,9392) Metal Artifact Reduction Algorithm with a value from CID 10036 (e.g., MAR_IMAR, MAR_SPECTRAL).

Verification (hex-tag method)

Because pydicom's bundled data dictionary does not yet include the 2026b additions, auditing requires hex-tag access:

import pydicom
ds = pydicom.dcmread('astm_reference_dataset/noMAR_recon/LP/realization_001/slice_0129.dcm')

# Access the MAR Macro via hex tags
mar_seq = ds[0x00189390]                          # Metal Artifact Reduction Sequence
mar_applied = mar_seq.value[0][0x00189391].value   # → "NO"
print(f'(0018,9391) Metal Artifact Reduction Applied = {mar_applied}')

---

Regulatory Framework

Layer	Document	Role
1 — What to record	DICOM CP-2575 (2026b)	MAR metadata tags in DICOM
2 — Must have / describe / record	IEC 60601-2-44 Ed. 4 §203.6.7.101 (.1 method, .2 user info, .3 DICOM)	Regulatory mandate
3 — How to measure (bound by compliance statement)	ASTM FXXXX (formerly WKXXXXX Rev 04) — proposed compliance statement of §203.6.7.101.1	Quantitative ΔAUC TYPE TEST
4 — Acceptance	FDA guidance (proposed)	Non-degradation / superiority thresholds

---

Technical Contact

Christopher D. Cocchiaraley Consumer Member, ASTM International Committee F04
Executor of the Estate of Veronica M. Cocchiaraley