RADIA: Radiology + AI A DICOM Viewer with AI on the Edge

Traditional medical imaging tools are heavy desktop software, expensive licenses and workflows that haven't changed in 20 years. A radiologist opens a PACS viewer, manually reviews slice by slice, dictates their report and sends it to the clinician. What if that entire workflow could live in a browser, with AI analyzing every slice in parallel, generating findings with coordinates and severity, and enabling real-time collaboration?

That's RADIA. Not a prototype or a demo. It's a production platform that processes real DICOM — from dental CBCTs with 300+ slices to mammograms, CT scans and clinical dermatology photography — with three AI models working in consensus.

The conventional radiological workflow has three fundamental problems that have remained unsolved for decades:

The most important decision was not to spin up servers. RADIA runs 100% on Cloudflare's edge: pages are static (Astro SSG), business logic runs on Workers, the database is D1 (distributed SQLite) and DICOM files live on R2 (S3-compatible). The result: 3-second deployments, global latency < 50ms, and zero infrastructure maintenance.

DICOM is the universal medical imaging standard, but it's a complex format: binary headers with nested tags, varied encodings (Little Endian, Big Endian, encapsulated) and JPEG 2000 compression that browsers don't natively understand.

Instead of parsing on the server (which would consume Worker resources), RADIA does all parsing on the client. The dicom-parser library extracts headers, and for JPEG 2000 pixels we use the OpenJPEG codec compiled to WebAssembly. The result: the Worker never touches pixel data, it only orchestrates.

ZIP (drag & drop)
  → JSZip extracts files
  → dicom-parser reads headers (patient, modality, geometry)
  → OpenJPEG WASM decodes JPEG 2000 → PNG thumbnails
  → POST /api/studies/upload (metadata)
  → PUT /api/dicom/{key} × N (binaries in parallel, 5 concurrent)
  → POST /api/studies/{id} action=finalize
  → Redirect → viewer with study ready

The trick is the dual storage layer: original DICOMs are stored in R2 for AI analysis, and in parallel we generate PNG thumbnails that are 10-50× smaller for fast navigation. The viewer preloads ±5 slices in each direction using the browser's Cache API, achieving an instant scroll experience even with 400-slice studies.

For 3D reconstruction of volumetric images (dental CBCT, CT), we wrote a ray marcher from scratch in GLSL. No Three.js (400KB+), no 3D dependencies. Just two shaders, a 3D texture and inline 4×4 matrix math.

The volume is cached in IndexedDB with a 7-day TTL, so reopening a study shows 3D instantly without rebuilding. On mobile, we reduce texture resolution and ray marching steps to maintain 30+ FPS with smooth touch rotation.

RADIA doesn't have a single "analyze" endpoint. It has five distinct analysis modes, each designed for a different moment in the doctor's workflow. Before running any analysis, a pre-analysis dialog lets the doctor choose whether to complement existing findings, critically review them, or wipe and re-analyze from scratch — with an Asteroid Mode toggle and a reason field to guide the AI's focus.

The pre-analysis dialog: the doctor chooses the mode, enables Asteroid, and provides clinical context before launching the AI.

AI-generated treatment suggestions: risk level, priority findings and staged plan with urgency and referrals.

The four models are chained in priority cascade:

Asteroid Mode is the heavy artillery. Instead of using a single model that falls back if it fails, it runs Gemini and MedGemma in parallel on each slice, cross-references findings and marks matches as multi-model consensus. After visual detection, DeepSeek Reasoner acts as a "Clinical Council": it receives all findings (text only, no images) and applies chain-of-thought reasoning to debate clinical consistency, detect likely false positives, and prioritize findings.

☄️ Asteroid Mode in action — 425-slice dental CT, multiplanar MPR with 3D volumetric, 113 findings detected with Gemini + MedGemma in parallel. Real-time progress bar with ETA.

☄️ Full MPR view — Four simultaneous panels: axial, 3D volumetric, coronal and sagittal. 17 suspicious and 24 indeterminate findings automatically classified. Findings panel with integrated search.

Slice N of study
  │
  ├─── Gemini 2.5 Flash ──────────┐
  │    "3 findings detected"       │
  │                                ├── Promise.allSettled()
  ├─── MedGemma 4B-IT ───────────┘
  │    "2 findings detected"
  │
  ▼
Cross-reference by (category + location + slice)
  │
  ├── Match → ☄️ Consensus
  │   • confidence × 1.15
  │   • ai_model = "gemini-2.5-flash+medgemma-4b-it"
  │
  └── Single model only → normal finding
      • confidence unchanged
      • ai_model = individual model

  ▼
DeepSeek Reasoner — Clinical Council (text-only)
  │
  Input: all findings JSON + clinical data
  │
  Output:
  ├── Expanded clinical differential with reasoning
  ├── Likely false positive detection
  ├── Cross-finding correlations
  └── Prioritization and recommendations

An important detail: AI consensus is never marked as doctor confirmation. The user_confirmed field always starts at 0 regardless of how many models agree. Human validation is a separate, intentional step — AI assists, it doesn't replace.

Each AI-generated finding is a rich structure with enough data for precise image localization, clinical categorization and validation workflow tracking. Findings on adjacent slices with the same category and location are automatically grouped (deduplicated), showing a slice range instead of repeated entries. The doctor can delete, hide from report, or navigate directly to each finding — individually or as a group.

Severity is classified into 5 levels with consistent color coding across the entire UI:

Findings with bounding boxes and slice numbers are projected as 3D markers onto the rendered volume. The bounding box (x, y) position maps to the volume coordinate system, slice_number determines Z depth, and the MVP (Model-View-Projection) matrix transforms them to screen coordinates every frame.

Tapping a marker shows a popup with finding information — title, severity, description, location — and three actions: go to findings list, jump to the 2D slice where the finding is, or open the MPR view centered on that position. The doctor navigates from 3D to 2D with a single touch. Additionally, Navigate Mode allows clicking anywhere on the 2D viewer to automatically jump to the nearest finding by Euclidean distance — ideal for exploring areas of interest without searching the list.

RADIA isn't just a dental viewer. It supports 16 study types organized in 6 medical sectors, each with specialized AI prompts and specific finding categories:

Detection is automatic: we analyze DICOM headers (Modality, SeriesDescription, StudyDescription) with a cascade of rules — first keywords in the description, then the DICOM modality code, then anatomy heuristics. The user can always correct manually, but in practice automatic detection is accurate 95%+ of the time.

The veterinary module isn't an afterthought bolted onto the medical engine — it's a dedicated branch with prompts tuned for animal anatomy, vet-specific clinical terminology, and a workflow built for 24/7 emergency cases. The AI distinguishes between laterolateral and ventrodorsal projections, recognizes the pathognomonic "double bubble" or "smurf hat" pattern, and prioritizes findings by life threat. In this real case, an abdominal radiograph of a Labrador Retriever triggered 19 automatic findings in under 30 seconds, with 6 classified as malignant.

Asteroid Mode activates the full ensemble of models in parallel and applies a second critical-review pass — designed for emergency cases where missing a finding is not an option. The image also shows the semantic boxes: each color encodes severity (red=malignant, orange=suspicious, yellow=undetermined, green=normal) and each box is linked to a navigable finding in the bottom panel.

In veterinary medicine the "patient" doesn't talk — and the owner is the one signing the clinical decision. That's why RADIA generates, alongside the technical report, a plain-language narrative aimed at the animal's owner. Same clinical content, two registers: the technical one for the medical record, and the owner-facing one explaining what's visible, why it matters and how urgent it is. All traceable: every owner explanation points back to the technical finding that backs it.

The PDF the veterinarian downloads isn't a screenshot — it's a structured report generated client-side with jsPDF, signed with the professional's data (multi-country license number: Spain, Mexico, Colombia, Argentina, Chile, etc.), optional scanned stamp and clinic tax ID. The final page is always the cartographic appendix: a panoramic view with all findings numbered over the original radiograph, mirroring the severity colors and linking each number back to the detailed finding from earlier pages.

That closes the loop: AI detects, the professional validates and edits, the system writes the owner-facing narrative, and the signed PDF goes to the medical record. In this GDV case, RADIA turned a suspicious radiograph into an actionable report in under two minutes — the kind of timing that matters when you're dealing with an emergency whose therapeutic window is measured in hours.

The chat isn't a generic chatbot you pass an image to. It's an assistant with full study context: DICOM metadata, all previous findings, conversation history, and the current viewer image (including the 3D view if active). If the doctor asks "what do you think about the periapical area of tooth 46?", the AI knows which slice is shown, what findings were already detected there, and can generate a contextualized analysis.

If the 3D viewer is active, the chat automatically captures the current view as PNG and includes it in the request. The doctor rotates the volume to the perspective they're interested in, types their question, and the AI sees exactly what the doctor sees. No manual screenshots.

The doctor generates a secure link with optional expiration. The link provides read-only access to the complete study — DICOM viewer, 3D, MPR, all findings — without the recipient needing an account or login. For doctor-to-doctor collaboration (where authentication is needed), there's a "collab" mode that requires Google login but doesn't need study ownership.

The DICOM proxy verifies the sharing token against the database on every request. There are no signed URLs that could leak — if you revoke the link, access is cut immediately.

Radiologists work with two or more monitors: one to review the image, another to consult the report, findings and history. Desktop PACS viewers support this natively, but web apps don't — opening two tabs of the same study gives two independent instances with no connection between them.

RADIA solves this with real-time synchronization between tabs. One click on "Multi-screen" opens a second window in viewer mode — optimized for the image, 3D and MPR — while the original becomes the editor with the findings list, AI chat and analysis tools. Everything the doctor does in one window is instantly reflected in the other.

Real multi-monitor: deep analysis with clinical differential in editor (left) and 3D volume with synced markers in viewer (right).

Synchronized 2D viewer: bounding boxes on the CBCT slice with labeled findings.

Findings list with severity filters and 3D with location markers.

Editor (monitor 1)                    Viewer (monitor 2)
  │                                       │
  │── NAVIGATE_SLICE { slice: 147 } ─────→│ Jumps to slice 147
  │                                       │
  │── SELECT_FINDING { id: "abc" } ──────→│ Shows bounding box
  │                                       │
  │── VIEW_MODE { mode: "3d" } ──────────→│ Activates 3D view
  │                                       │
  │── HIGHLIGHT_3D { id: "abc" } ────────→│ Marks on volume
  │                                       │
  │←── PING ──────────────────────────────│
  │── PONG ──────────────────────────────→│ Heartbeat every 3s
  │                                       │
  │── FINDINGS_CHANGED ──────────────────→│ Reloads findings

A PING/PONG heartbeat every 3 seconds detects if the other window is still active. If it doesn't respond within 6 seconds, an amber disconnection indicator appears. On reconnection, it turns green. The doctor always knows if their windows are synchronized.

The PDF report is generated entirely in the browser with jsPDF — patient data, selected findings with severity codes, overall impression, recommendations and doctor's notes. The Worker never generates PDFs; the client has all the necessary information. This avoids the complexity of rendering documents in a serverless environment with 128MB of memory.

RADIA proves that professional medical tools can live entirely on the edge, with no dedicated servers, no rented GPUs, no infrastructure to maintain. A doctor opens a browser, uploads their study, and three AI models analyze every slice while the 3D viewer rebuilds in WebGL2 at 30fps.

The full stack — Astro, React, Cloudflare Workers, D1, R2, Gemini 2.5 Flash, MedGemma, WebGL2 — deploys with npm run deploy in 15 seconds. No containers, no CI/CD pipelines, no clusters. The future of medical tooling is a static build with serverless edge functions, and it's already here.