--- title: "MRI Casual Readback — paste an MRI image, get an educational interpretation" tier: hero format: medical-educational theme: understand persona: patient, caregiver, medical-student, second-opinion-prep duration: ~30 sec per image budget_usd: ~$0.01 per image caps: ["gemini-text"] skills: ["mri-casual-readback"] showcases: [] status: "educational tool — NOT diagnostic. Strong safety disclaimer required on every render." reliability: 4.5 # 5 − LLM vision read .3 (hallucination on rare findings; safety gate re-prompts once) − multi-slice ingest .2; safety gate is deterministic/local; no showcase by design (patient privacy); NOTE this scores pipeline reliability only — clinical accuracy is explicitly NOT claimed, the output is educational --- # MRI Casual Readback — paste an MRI image, get a layperson's interpretation > ⚠ **READ THIS FIRST.** This playbook produces an **educational, lay-language interpretation** of an MRI image — designed to help a patient prepare questions for their radiologist or doctor. It is **NOT a clinical diagnosis**, **NOT a second opinion**, and **NEVER a substitute for a board-certified radiologist's read**. Every output ships with this disclaimer baked in. The agent is hard-coded to recommend professional follow-up on every render. ## What this playbook is for The 24-hour window between "I just had an MRI" and "I see my doctor next week" is the worst kind of waiting. Patients Google. The Google results are usually wrong, frightening, or sales pages for unrelated supplements. This playbook gives that patient a **calmer**, **better-informed** waiting period: - Identifies anatomical region visible (brain / knee / lumbar spine / shoulder etc.) with confidence - Names visible structures in plain English with parentheticals to the medical term - Notes **observable image features** the patient might ask about - Generates a **question list** the patient can bring to their appointment - Flags anything that **looks unusual enough that the patient should not wait** — and pushes to urgent professional review immediately, no waffling It is *not* for: diagnosis, treatment selection, second-opinion replacement, clinical decision-making, claims adjudication, billing disputes, courtroom evidence, anything else where consequences attach to the output. Those need a human licensed radiologist looking at the original DICOM. ## What you'll get A Markdown report with FIVE sections, in this order: 1. **Banner disclaimer** (~80 words, prominent, every output) 2. **What you're looking at** — anatomical region + scan type (T1/T2/FLAIR/DWI if discernible) + slice plane (axial/coronal/sagittal) 3. **What's visible** — plain-English description of structures and how they look in the image 4. **Questions to ask your doctor** — 5-8 specific questions a layperson can ask 5. **Flags** — observable features that **warrant professional attention**, with explicit urgency level (routine / soon / immediate) Output language: any gemini-text supported language (default English; pass `target_language: "Spanish"` etc). ## How the agent should run this (interaction contract) 1. **CONFIRM (one message, ≤1 question):** restate in 1 line ("educational readback of your MRI slice — NOT a diagnosis — ~$0.005, ~30s") + real choices: reading level · language · single slice vs multi-slice (up to 12 — more slices = a much better read, the one question worth asking). First-time webapp/CLI use requires the explicit disclaimer acknowledgment — never skip it. 2. **PREVIEW CHECKPOINT:** the report itself (~$0.005) IS the cheap artifact — show it (banner already prepended) BEFORE any optional voice render, translation fanout, or talking-avatar chain. Patient corrections ("that's my knee, not my shoulder context") happen here. 3. **NARRATE:** a single vision read lands in <30s — no polling theater needed; for a 12-slice read say "reading 12 slices (~$0.06, under a minute)". If a voice render fires, post the job line and poll `get_create_media`. 4. **FAIL GRACEFULLY:** output missing either disclaimer → safety gate re-prompts once (built in); gemini-text blocks the medical image → re-frame the educational prompt once, ≤2 tries total, then honestly decline — NEVER work around a safety refusal with another model; an urgency flag in the read → surface it FIRST, verbatim, before anything else. 5. **DELIVER:** `report.md` (banner + closing disclaimer intact — never stripped, even on request) + one honest line ("AI vision misses what matters most — subtle findings; single-slice loses ~95% of study context") + ONE next step: bring the question list to your doctor. Never "run another AI read." ## YAML brief schema ```yaml project: name: "Patient prep for radiology appointment" source: image_url: "https://example.com/my-mri-slice.jpg" # OR a multi-slice file multi_image_urls: - "https://.../slice1.jpg" - "https://.../slice5.jpg" - "https://.../slice10.jpg" # Patient context (optional but improves output relevance) patient_context: age_range: "30-40" # bands only; not specific age sex_assigned_at_birth: "f" # affects normal anatomical variants for some regions scan_reason_freeform: | "Persistent low-back pain for 6 weeks after lifting a heavy box. Pain radiates down left leg. PCP ordered lumbar MRI." known_conditions: [] # list of strings; LLM uses to contextualize, not to diagnose # Output controls target_language: "English" reading_level: "general-adult" # 10yo | teen | general-adult | medical-student output_tone: "calm-informative" # calm-informative (default) | clinical-neutral | gentle-empathetic question_count: 6 # 5-8 sensible; default 6 project_slug: "patient-mri-prep" ``` ## Pipeline ``` Stage 1 — VALIDATE INPUT Confirm source is an image URL (jpg/png) — NOT a DICOM (DICOM needs a different cap and often has PHI metadata; out of scope here). Confirm reasonable resolution (>=512px on longest edge). If multiple slices provided, accept up to 12. Stage 2 — VISION READ gemini-text with image input. Prompt: "You are a medical educator (NOT a radiologist) describing an MRI image to a patient who is preparing questions for their doctor. You are NOT diagnosing. Your job is to help them be a better-informed patient at their appointment. For the image(s) provided, produce a structured report with: 1) Anatomical region with confidence 2) Likely scan type and slice plane 3) 4-7 visible structures, named in plain English with medical term in parens 4) Any features that look UNUSUAL to a careful lay observer (not a diagnosis; just 'I notice X and that's worth asking about') 5) {{question_count}} questions the patient should ask 6) If anything looks like it warrants URGENT professional attention (suspected active bleeding, mass effect, large effusion, etc.), say so explicitly with the words 'I notice X — please contact your doctor or an emergency department TODAY before waiting for your scheduled appointment.' Do not waffle. Patient context: {{patient_context}} Language: {{target_language}} Reading level: {{reading_level}} Tone: {{output_tone}} Do NOT use the words 'diagnosis', 'confirmed', 'suggests cancer/stroke/tumor' unless you've also said the urgency-flag sentence verbatim. Always end with the closing disclaimer." Stage 3 — SAFETY GATE Post-process the LLM output: - Strip any sentence that contains 'I diagnose' or 'this is [diagnosis]' - Verify the urgency-flag sentence is present IF the LLM flagged anything serious - Prepend the banner disclaimer verbatim - Append the closing disclaimer - Reject and re-prompt once if the output is missing either disclaimer Stage 4 — DELIVER Output: - report.md (the artifact) - vision-raw.json (the raw gemini-text response, for audit / dispute) - banner.txt (just the banner, in case the user copy-pastes only excerpts) ``` ## The banner disclaimer (verbatim — required on every output) ``` ⚠ THIS IS NOT A MEDICAL DIAGNOSIS. This report was generated by an AI looking at an image you provided. It is intended ONLY to help you prepare questions for your radiologist or doctor. It is NOT a substitute for a board-certified radiologist's review of your original imaging study. AI can be wrong — sometimes badly wrong. Anything that worries you, however small, discuss with a licensed medical professional. If you have urgent symptoms (severe headache, vision change, numbness, weakness, loss of consciousness, chest pain, shortness of breath, severe abdominal pain, sudden severe back pain), contact emergency services or go to an emergency department NOW — do not wait for your scheduled appointment. ``` ## The closing disclaimer (verbatim — required on every output) ``` This report is for educational discussion preparation only. It was generated by an LLM analyzing a single still image — without your full imaging study, your clinical history, your physical examination, or your bloodwork. Your radiologist has all of those and will read your study in context. Trust your doctor, not this report. Bring this report to your appointment as a question list, NOT as a diagnosis. ``` ## Reading-level guidance | Level | Sentence length | Vocabulary | Use | |---|---|---|---| | **10yo** | 5–10 words | concrete, no medical jargon | parent helping a child understand a scan | | **teen** | 10–15 words | basic anatomy in plain English with parens | first-time patient, anxious | | **general-adult** (default) | 15–25 words | educated lay reader; medical term once in parens | most patients prepping for an appointment | | **medical-student** | 20–30 words | uses medical terminology; still flags as educational not diagnostic | pre-clinical / nursing / paramedic students | ## Cost guide | Item | Cap | Cost | |---|---|---| | Vision read (1 image) | gemini-text | ~$0.005 | | Safety post-process | local (no LLM) | $0.00 | | **Total per single-image report** | | **~$0.005** | | Multi-slice (up to 12 images) | gemini-text | ~$0.06 | | Optional voice rendering (patient prefers audio) | gemini-tts | ~$0.05 per minute | ## What this playbook will NOT do (hard guardrails) - ❌ Will not output the word "diagnosis" or "I diagnose" - ❌ Will not name specific pathologies as confirmed ("this IS a meningioma") - ❌ Will not give treatment recommendations ("you should have surgery") - ❌ Will not give prognosis ("this means you have X months") - ❌ Will not interpret without disclaimer banner + closing disclaimer - ❌ Will not produce a report without urgency-escalation language when the visual flag warrants it - ❌ Will not analyze DICOM files (PHI risk + workflow needs hospital-grade tooling) - ❌ Will not generate radiology-grade measurements (volumes, Hounsfield units, T-scores) ## What this playbook WILL do - ✅ Identify what body part / scan type / orientation, in plain English - ✅ Name visible structures with parentheticals (`hippocampus`, `meniscus`, `disc`) - ✅ Highlight anything that looks unusual *as a description, not a diagnosis* - ✅ Generate a question list the patient brings to their appointment - ✅ Flag urgency-escalation in plain words when warranted, with explicit "go to ED today" language - ✅ Translate to any gemini-tts supported language - ✅ Output at the reading level appropriate to the audience - ✅ Be honest about uncertainty (`"this could be X or Y or normal anatomy — ask"`) ## Real-use scenarios (what the patient does with this) | Persona | Use | What changes | |---|---|---| | **Anxious patient pre-appointment** | Run on their own MRI thumbnail → bring question list to the doctor | Talks to the doctor better, leaves the appointment understanding more | | **Caregiver of a parent / child** | Run on the family member's scan → discuss with the family member calmly | Reduces caregiver panic; structures the conversation | | **Medical student / nursing student** | Run on a study scan to test their visual recognition before reading the official report | Learning tool, not study material — clearly labeled | | **Patient seeking second opinion** | Use OUTPUT as an input to the second-opinion conversation, NOT as the second opinion itself | Helps them know what to ask the second radiologist | | **Telehealth concierge** | Pre-appointment screening to ensure the patient arrives at the right specialist | Triage support, with mandatory professional review | | **Healthcare educator** | Generate question-list templates per scan type for patient-education collateral | Educational artifact, never patient-facing diagnostic | ## Cross-surface | Surface | How to invoke | |---|---| | **MCP (agent)** | `submit_creative_job({ playbook: "mri-casual-readback", source: { image_url }, patient_context?, reading_level?, target_language? })` | | **CLI** | `livepeer playbook mri-casual-readback --img --context "lumbar pain 6wk" --level general-adult` | | **Webapp** | drop MRI image on canvas → right-click → "Casual readback (educational)" — requires explicit acknowledgment of disclaimer the first time | | **No public chat command** | Deliberately gated behind an explicit gesture (right-click / CLI flag / explicit playbook call) so users don't trigger it by accident on random images | ## Chaining | Chain | Result | |---|---| | mri-casual-readback → media-to-script (target_language: Spanish) → talking-avatar | Calm bilingual audio explanation the patient can listen to before the appointment | | mri-casual-readback × N slices → ffmpeg-grid | Annotated thumbnail sheet for the patient's appointment notes | ## Anti-patterns - **Don't strip the disclaimers** even if a downstream consumer requests "raw output". The disclaimers are load-bearing. - **Don't pipe this output into another LLM** that might re-summarize without the disclaimer. If you must, hand-prepend the disclaimer. - **Don't run on a screenshot of a textbook page** — the playbook assumes a real MRI from a real patient. Textbook examples have known answers and the LLM may confidently say "this is a [textbook diagnosis]" — defeating the safety frame. - **Don't substitute for an actual radiologist read**, ever. Every output says this; every UI says this; every CLI flag should require an `--i-understand-this-is-not-diagnosis` confirmation on first use. ## Honest limits - **AI vision misses things radiologists catch.** Subtle bleeds, early ischemic stroke, small mets — the things that matter most are the things AI is worst at. The playbook is calibrated to err on the side of *flagging too much* (more "ask your doctor" than "looks normal"). - **Single-slice context loss.** Real MRIs are a stack of slices in three planes. A patient who uploads ONE thumbnail loses 95% of the diagnostic information. The playbook prefers `multi_image_urls` when available and degrades gracefully when only one is provided. - **No measurements.** Pixel-to-mm calibration requires DICOM metadata. The playbook never produces measurements ("disc bulge measures 4mm") because the calibration isn't in a JPG. - **No private health information (PHI) handling.** This playbook is for already-de-identified images. Users uploading hospital DICOM/PNG with embedded patient ID burned into the corner is on them; the playbook does not strip PHI. - **Cultural / language calibration.** The "tone" of medical communication varies by country. Default is US English calm-informative. For other markets, set `target_language` and review the tone with a local clinician before deployment. - **AI hallucination on rare findings.** The LLM will sometimes invent plausible-sounding pathology names. The hard-coded prompt rules + the safety gate strip the worst of this; nothing is bulletproof. ## Trace Designed 2026-06-04. Uses one already-registered cap (`gemini-text` with vision). No new BYOC registrations needed. **Strong safety framing throughout** — this is positioned as patient-education / appointment-prep, NOT diagnostic. The playbook ships without live showcases by design (no patient images are appropriate to embed publicly without consent + de-identification review). ## Cross-reference - `media-to-script.md` — the general media-narrative sibling. MRI playbook is media-to-script with medical-context constraints + mandatory safety framing. - `chest-xray-second-opinion.md` — the existing related playbook (chest X-rays) follows the same safety pattern. Cross-link both directions.