MCP-MedImg: Reproducible Experimental Data for PACS-AI Integration via Model Context Protocol

This dataset contains the complete reproducible experimental data for the paper "MCP-MedImg: Standardizing PACS-AI Integration through the Model Context Protocol." Contents 1. Trained Model Weights- nnU-Net v2 5-fold cross-validation checkpoints for MSD Task02 Heart (left atrium segmentation)- 5 files, ~235.3 MB each, SHA256 verified- Dice score: 0.9324 ± 0.0062 (5-fold mean) 2. Training Logs- Complete logs from 5-fold serial training (~2.5 days on NVIDIA RTX 4090D)- 8 log files including per-fold training, preprocessing, and environment setup 3. Benchmark Results- LIDC-IDRI Inference Latency (n=200): Real GPU inference latency using SimpleUNet3D (22.6M parameters, randomly initialized). End-to-end mean: 5.40s. Latency distribution is bimodal: small volumes (single-patch, mean 178.6ms) vs large volumes (sliding-window, mean 4410.4ms).- MSD Heart Inference Latency (n=10): End-to-end latency using nnU-Net 5-fold ensemble. Mean: 18.20s (single-fold baseline also included).- MCP Protocol Overhead (n=98): Probe-based measurement across 4 architectural layers. Protocol overhead mean: 567.47ms, dominated by L1 transport (566.35ms, 99.8%). Validation mean error: 3.41% (3.19% after outlier removal). 4. Configuration & Scripts- nnU-Net auto-configured network plans and 5-fold split definitions- Reproduction scripts for all benchmarks (GPU, inference, protocol overhead)- P0 root-cause analysis scripts and summary data 5. Analysis Figures- LIDC inference latency variance analysis (bimodal distribution, quartile breakdown)- PROTO-004 validation error analysis (robustness, outlier identification) Hardware & Environment- GPU: NVIDIA GeForce RTX 4090D (24GB VRAM)- CPU: AMD EPYC 7K62 64-Core- CUDA: 12.1, PyTorch: 2.3.0, nnU-Net: v2- OS: Ubuntu 22.04 LTS (AutoDL Cloud, China West B) Notes- LIDC model weights are NOT included. The SimpleUNet3D architecture is fully defined in scripts/real_inference.py; the model was randomly initialized as the experiment focuses on inference latency benchmarking of the protocol stack, not segmentation accuracy.- Datasets (MSD Task02 Heart, LIDC-IDRI) are NOT included due to license restrictions. Download links and preprocessing instructions are provided in README.md. LicenseCC-BY 4.0 for academic and non-commercial use.