Clinical AI Agents, MCP Tooling, and Vector Search with dicomtk.ai

Overview

dicomtk.ai extends dicomtk with ellmer-based clinical imaging agents, a Model Context Protocol (MCP) server via mcptools, pluggable AI segmentation backends with uncertainty maps, clinical report NLP, RadLex / LOINC-RSNA protocol harmonisation, and DuckDB vector similarity search.

1. Guardrailed 5-Stage Agent Workflow

Every de-identification workflow enforces the PS3.15 monotonic join-semilattice (K < C < U < D < Z < X), preventing any LLM or caller from weakening the baseline profile:

library(dicomtk)
library(dicomtk.ai)

synth <- dicom_synthesize(modality = "CT", dims = c(16L, 16L, 4L), phi_profile = "basic")
out_dir <- tempfile("ai_vignette_out_")

plan <- ai_agent_plan(
  task = "deid_and_qc",
  path = synth[["out_dir"]],
  baseline_profile = dicom_deid_profile("ps3.15"),
  chat = ai_chat("offline")
)
dry <- ai_agent_dry_run(plan)
approved <- ai_agent_approve(dry, auto_approve = TRUE, approver = "vignette")
exec_res <- ai_agent_execute(approved, out_dir = out_dir)
ver <- ai_agent_verify(exec_res)
ver$passed
#> [1] TRUE

2. Protocol Harmonisation & Weak-Label Extraction

hdrs <- tibble::tibble(
  SeriesDescription = c("CHEST WO CON 3.0mm", "BRAIN_T1_POST_GAD"),
  Modality = c("CT", "MR"),
  BodyPartExamined = c("CHEST", "HEAD")
)
ai_harmonize_protocol(hdrs)
#> # A tibble: 2 × 9
#>   SeriesDescription  Modality rpid    long_common_name    short_name body_region
#>   <chr>              <chr>    <chr>   <chr>               <chr>      <chr>      
#> 1 CHEST WO CON 3.0mm CT       RPID29  CT Chest WO contra… CT Chest … CHEST      
#> 2 BRAIN_T1_POST_GAD  MR       RPID102 MR Brain WO and W … MR Brain … HEAD       
#> # ℹ 3 more variables: radlex_id <chr>, contrast <lgl>, confidence <dbl>

3. Cleanup