ASTRO 2026 · BOSTON SEPT 27–29 · BOOTH 2814 Book a booth demo →
Features

What sets
Radial apart.

Radial turns a DICOM archive into a queryable dataset: structure names standardized, cases sorted into cohorts, plans scored against protocol limits, and metrics watched as new cases accrue.

The mapping loop

Structure mapping that learns your institution.

Every ROI gets a predicted TG-263 name with a confidence and the reason it matched, including geometry: a contour sitting a uniform 5 mm outside the cord reads as a PRV expansion, whatever its name. High confidence auto-confirms; the rest queue for one-keystroke review.

Every decision teaches Radial: confirm a variant once and Radial offers to backfill every existing structure that matches.

NAME + GEOMETRY EVIDENCE ONE-KEYSTROKE REVIEW DATABASE THAT LEARNS
CURATE · ROI MAP — UNMATCHED NAMES
REVIEW
cord_prv+5mm
14× occurrences · 3 mapped
Cord_PRV+5mmcord_PRV5+2
CONFIGURE
Map to existingSpinalCord_PRV05
SUGGESTEDSpinalCord_PRV05HIGH 0.89SpinalCord0.64
Uniform 5 mm expansion of SpinalCord · PRV-suffix token
APPLY
● BULK READYAdd variationAdd + map 11
Automatic cohort assignment

Every case sorted into
its cohort with evidence.

Radial groups incoming cases by treatment site, dose, and fractionation, then proposes an assignment per group, every case carrying a “why this assignment” drawer.

You still confirm and nothing files itself silently. But one decision covers a group, and mis-assigned legacy cases surface for moving.

GROUPED BY SITE × FX PER-CASE PROVENANCE YOU CONFIRM — IN BULK
COHORTS · ASSIGN
Oropharynx33 fx9Assign all → Choose cohort…
5 H&N OPC · 20243 Unassigned· 3 PENDING
ANON50121Plan v2H&N OPC · 2024×
ANON50127Plan v1 (boost)Assign cohort…
3 pending changesDiscardSave 3
COHORTS
LIVE COUNT AS YOU ASSIGN
H&N OPC · 202447+2
NPC · IMRT QA28
Unassigned12−2
Custom metrics

Your metrics, first-class.

Define any derived metric as free-form algebra over plan fields or dose endpoints bound to real structures. A live preview shows how many cases actually compute and how many come back null before you save; broken expressions are refused.

Saved metrics become columns, chart fields, and monitored signals, computed live for every cohort.

FREE-FORM ALGEBRA STRUCTURE-BOUND ENDPOINTS NULL-AWARE PREVIEW
CUSTOM METRICS · STUDIO
NAME
MU per prescription Gy
MODE
EquationFunction
FORMULA
total_MU / rx_dose
total_MUPlans · Total MUPER PLAN
rx_dosePlans · Rx dose [Gy]PER PLAN
aorta_D95Dose endpoint · D95%[Gy] @ A_AortaDATASET MAX
PREVIEW396 of 412 compute16 null — no MU recordedSave metric
Ambient monitoring

Every metric, watched every night.

Pin any metric and Radial charts it as cases accrue with a one-click toggle from raw to case-mix-adjusted, so a run of hard patients doesn’t read as a process change. Scope a chart to one machine, TPS or technique.

Alerts are controlled for false discovery, and every point drills down to its case.

RAW ⇄ CASE-MIX-ADJUSTED SCOPE TO ONE PROCESS POINT → CASE
MONITORING · Parotid_L Dmean — TRUEBEAM-2SIGNAL
RawAdjusted
Basis: EQD2 · α/β 3 — declared 2026-05-02Change α/β
UCL CL LCL
2 plans on 08-21 — open a case:ANON51204ANON51219
Case-mix modeladj R² 0.62n 218
CovariateCoeffpFitted range
ptv_volume0.0310.00341–612 cc
fractions−0.420.02115–35
The limit sheet

Every constraint on one axis.

Each constraint is rescaled as a fraction of its own limit, so every measure lands on one axis. Everything to the right is beyond the limit. The sheet names the widest offenders at the top.

MIXED UNITS, ONE SCALE BEYOND / WITHIN SPLIT AUTO-WRITTEN SUMMARY
COHORTS · LIMIT SHEET — PROSTATE 60/20
3 of 24 reporting measures sit beyond their reference limit — Rectum V60 by the widest margin.
Beyond the limit · 3 measures · median past the rule
Rectum V60+2.4%
FemoralHead_L Dmax+1.1 Gy
Within the limit · 21 measures
Bladder Dmean−4.1 Gy
SpinalCord D0.03cc−6.8 Gy
IQR WITHINIQR BEYONDREFERENCE0 = THE RULE'S EDGE
And more

Also worth knowing.

COHORT DISCOVERY

Queries that show their exclusions.

Every clause reports which cases it excludes, down to “this clause alone excludes 12 cases.” You can bypass one and re-count without touching the saved query.

CONTOUR EDITING

Fix a contour without leaving the cohort.

Draw, brush, livewire, nudge, and fill, plus slice interpolation, in the viewer. Each save writes a new version beside the original; the import is never overwritten.

PRIVACY MODE

Identifiers masked, at render time.

One switch swaps names, MRNs, and dates for synthetic data (or redacts them). Exports mask patient identifiers by default.

THE ASSISTANT · ALPHA

Ask in plain language.

Ask for a population DVH, a protocol evaluation, a benchmark. It runs the same tools as the rest of Radial and renders real charts and tables, not a text summary.

Assumed capabilities

And the fundamentals are all here.

The baseline capabilities you'd expect from any platform in this space, listed here rather than demonstrated, because they work exactly as you'd expect.

IMPORT
  • DICOM query/retrieve from PACS and TPS
  • Staged import queue, duplicate conflicts reconciled
  • Drag-and-drop upload
VIEWER & DOSE
  • Three-plane DICOM image viewer
  • Structure and isodose overlays
  • DVH engine based on dicompyler-core
  • Population DVH with percentile bands
PROTOCOLS & SCORING
  • Protocol libraries and in-house scorecards
  • Mayo-syntax clinical goals with cohort pass rates
  • Cohort metric tables across structures
  • Per-cohort reports and CSV export
PLATFORM
  • Self-hosted in your own environment
  • Admin, editor, and viewer roles
  • Multi-institution separation

Table stakes, done properly.

Start with the story →
Why a platform, not scripts

Your script ran.
Can you trust the dataset?

Physicists have been scripting data out of the TPS and OIS for years; we’ve written our share, and good scripts get real answers. The hard part starts after the script runs. The platform case, in five lines:

Curation is the hard part

Cord vs SpinalCord vs SC_PRV03; which of six plans was treated. Radial standardizes deterministically, with review queues: errors surfaced, not absorbed.

Ingest once

Every new question sends a script back through the whole archive, file by file. Radial ingests into a database built for cohort-wide queries; after that, answers take seconds.

Any TPS

Scripts target one vendor’s API and often need rework when it changes. Radial is DICOM-native: collaborators, archives, and partner sites included.

Reproducible

In-house scripts tend to leave with their author. Radial is the same validated engine for everyone, benchmarked against dicompyler-core; the methods section becomes one sentence.

Alive after the paper

A script produces one CSV, once. Radial keeps scoring, monitoring, and assigning as new cases accrue.

We don’t replace your science. We replace the wiring.

Seeing is believing

See it live,
on a real cohort.

Every panel on this page is drawn from real screens in the product. Book a booth demo at ASTRO 2026 and we’ll run the ones you care about.