Skip to main content

Model reference — AI Medical · Radiology

AI models for CT analysis

This is the fastest-moving group on the site. Representation models produce a general CT embedding and support triage, retrieval and segmentation from one pass. Vision-language models pair volumes with reports and return findings, prognosis and answers. Organ-specific diagnostic models cover one indication at higher accuracy.

Two reported results are worth noting: an abdominal CT diagnosis model with external-centre AUC 0.895 across eight centres, and a pan-organ model reported best on 319 of 366 findings in its own study. Both are single-study figures, which is exactly why we benchmark before sizing.

Every model on this page runs as part of a managed AI pipeline in our GPU clusters, with a dedicated private cluster in our cloud or an on-premise installation where medical governance requires it. Output is decision support for a qualified professional to review, not a diagnosis.

CT analysis service AI Medical services Pricing

Input type — CT volumes

Models in this group take a 3D CT volume, some with report or question text alongside. Each table gives three hardware tiers — Minimum, the smallest configuration on which the model runs correctly; Medium, the usual production configuration; and High, a configuration sized for peak volume. The rate is what one server of that tier processes per hour. Use these figures for initial sizing only. Before production we benchmark your own data to confirm accuracy, latency, throughput and cost.

CT-FM

Vendor: Project Lighter

What it does: produces a general-purpose representation of a CT volume that triage, retrieval and segmentation heads all reuse — so one pass over the archive serves several endpoints instead of one.

Task-dependent across segmentation, triage and retrieval benchmarks.

RequirementMinimumMediumHigh
GPU typeRTX 3090A100 80 GBH100 80 GB
VRAM24 GB80 GB80 GB
vCPUs122448
RAM64 GB128 GB256 GB
Server1× RTX 3090 24 GB1× A100 SXM 80 GB2× H100 SXM 80 GB
Rate (volumes/hour)≈ 60≈ 210≈ 540

Merlin

Vendor: Stanford MIMI

What it does: reads a CT volume alongside the clinical text that accompanies it, and returns findings, masks, report text and prognosis. The broadest single deployment on this page.

Task-dependent across CT classification, segmentation, retrieval, report generation and prognosis.

RequirementMinimumMediumHigh
GPU typeRTX 3090A100 80 GBH100 80 GB
VRAM24 GB80 GB80 GB
vCPUs122448
RAM64 GB128 GB256 GB
Server1× RTX 3090 24 GB1× A100 SXM 80 GB2× H100 SXM 80 GB
Rate (volumes/hour)≈ 60≈ 210≈ 540

CT-CLIP

Vendor: CT-CLIP authors

What it does: indexes CT volumes against report text so an archive can be searched by description, and a new finding scored without a training set for it.

Task-dependent across CT classification and retrieval benchmarks.

RequirementMinimumMediumHigh
GPU typeRTX 3090A100 80 GBH100 80 GB
VRAM24 GB80 GB80 GB
vCPUs122448
RAM64 GB128 GB256 GB
Server1× RTX 3090 24 GB1× A100 SXM 80 GB2× H100 SXM 80 GB
Rate (volumes/hour)≈ 60≈ 210≈ 540

CT-CHAT

Vendor: CT-CLIP authors

What it does: answers questions about a CT volume in plain language, which suits review and teaching rather than routine reporting.

Task-dependent across CT visual question answering.

RequirementMinimumMediumHigh
GPU typeL40S 48 GBA100 80 GBH100 80 GB ×2
VRAM48 GB80 GB160 GB
vCPUs163264
RAM64 GB128 GB256 GB
Server1× L40S 48 GB1× A100 SXM 80 GB2× H100 SXM 80 GB
Rate (volumes/hour)≈ 40≈ 140≈ 360

RADAR

Vendor: Alibaba DAMO Academy and collaborators

What it does: diagnoses across abdominal anatomy from a contrast-enhanced CT, and is one of the few models here with multi-centre external validation — AUC 0.895 across eight sites.

External-centre AUC 0.895 across eight centres; emergency CT AUC 0.904 in the reported Science study.

RequirementMinimumMediumHigh
GPU typeRTX 3090A100 80 GBH100 80 GB
VRAM24 GB80 GB80 GB
vCPUs122448
RAM64 GB128 GB256 GB
Server1× RTX 3090 24 GB1× A100 SXM 80 GB2× H100 SXM 80 GB
Rate (volumes/hour)≈ 60≈ 210≈ 540

LiON

Vendor: Alibaba DAMO Academy and collaborators

What it does: classifies liver malignancy and outlines the lesion, validated in a multicentre liver CT study. A single-indication model, and stronger inside that indication than any general one.

Task-specific performance reported in a multicentre liver-CT study; no single universal metric.

RequirementMinimumMediumHigh
GPU typeRTX 3090A100 80 GBH100 80 GB
VRAM24 GB80 GB80 GB
vCPUs122448
RAM64 GB128 GB256 GB
Server1× RTX 3090 24 GB1× A100 SXM 80 GB2× H100 SXM 80 GB
Rate (volumes/hour)≈ 60≈ 210≈ 540

Percival

Vendor: Percival authors

What it does: pairs a CT volume with its report to support retrieval, classification and prognosis from the same index.

Task-dependent across retrieval, classification and prognosis benchmarks.

RequirementMinimumMediumHigh
GPU typeRTX 3090A100 80 GBH100 80 GB
VRAM24 GB80 GB80 GB
vCPUs122448
RAM64 GB128 GB256 GB
Server1× RTX 3090 24 GB1× A100 SXM 80 GB2× H100 SXM 80 GB
Rate (volumes/hour)≈ 60≈ 210≈ 540

Pillar-0

Vendor: Yala Lab

What it does: covers an unusually wide finding list across CT and MRI — reported best-performing on 319 of 366 findings in its own study, which is a claim worth testing on your data before it is relied on.

Reported as best-performing model on 319 of 366 RATE findings in its study; accuracy varies by finding.

RequirementMinimumMediumHigh
GPU typeRTX 3090A100 80 GBH100 80 GB
VRAM24 GB80 GB80 GB
vCPUs122448
RAM64 GB128 GB256 GB
Server1× RTX 3090 24 GB1× A100 SXM 80 GB2× H100 SXM 80 GB
Rate (volumes/hour)≈ 60≈ 210≈ 540

SPECTRE

Vendor: SPECTRE authors

What it does: returns embeddings, segmentation and retrieval scores from a CT volume with optional text, covering three task types from one model.

Task-dependent across CT classification, segmentation and retrieval.

RequirementMinimumMediumHigh
GPU typeRTX 3090A100 80 GBH100 80 GB
VRAM24 GB80 GB80 GB
vCPUs122448
RAM64 GB128 GB256 GB
Server1× RTX 3090 24 GB1× A100 SXM 80 GB2× H100 SXM 80 GB
Rate (volumes/hour)≈ 60≈ 210≈ 540

TAP-CT-B-3D

Vendor: TAP-CT authors

What it does: produces a task-agnostic 3D CT representation, intended as the base layer under your own segmentation or classification head.

Task-dependent across CT segmentation and classification benchmarks.

RequirementMinimumMediumHigh
GPU typeRTX 3090A100 80 GBH100 80 GB
VRAM24 GB80 GB80 GB
vCPUs122448
RAM64 GB128 GB256 GB
Server1× RTX 3090 24 GB1× A100 SXM 80 GB2× H100 SXM 80 GB
Rate (volumes/hour)≈ 60≈ 210≈ 540

3DINO

Vendor: AICONS Lab

What it does: handles both CT and MRI volumes in one 3D representation model, which avoids maintaining a separate encoder per modality.

Task-dependent across 3D classification and segmentation datasets.

RequirementMinimumMediumHigh
GPU typeRTX 3090A100 80 GBH100 80 GB
VRAM24 GB80 GB80 GB
vCPUs122448
RAM64 GB128 GB256 GB
Server1× RTX 3090 24 GB1× A100 SXM 80 GB2× H100 SXM 80 GB
Rate (volumes/hour)≈ 60≈ 210≈ 540

Curia-B

Vendor: Raidium

What it does: reads CT and MRI DICOM directly and returns general radiology embeddings — a practical base where your inputs arrive straight from PACS.

Task-dependent; general-purpose radiology representation model.

RequirementMinimumMediumHigh
GPU typeRTX 3090A100 80 GBH100 80 GB
VRAM24 GB80 GB80 GB
vCPUs122448
RAM64 GB128 GB256 GB
Server1× RTX 3090 24 GB1× A100 SXM 80 GB2× H100 SXM 80 GB
Rate (volumes/hour)≈ 60≈ 210≈ 540

CT Foundation

Vendor: Google Health

What it does: compresses a CT volume into a 1408-dimension embedding, so a classifier trained on a few hundred labelled studies becomes viable.

Data-efficient 3D CT classification from embeddings; research-endpoint origin, now downloadable (developer-reported).

RequirementMinimumMediumHigh
GPU typeRTX 3090A100 80 GBH100 80 GB
VRAM24 GB80 GB80 GB
vCPUs122448
RAM64 GB128 GB256 GB
Server1× RTX 3090 24 GB1× A100 SXM 80 GB2× H100 SXM 80 GB
Rate (volumes/hour)≈ 60≈ 210≈ 540

Choosing between them

If you want one indication, an organ-specific model is the strongest option and the easiest to validate. If you want breadth or retrieval, a representation or vision-language model is the base. 3D CT is memory-bound rather than compute-bound, so the sizing below is driven by volume dimensions more than by parameter count.

Accuracy figures above are those the producers and independent evaluations report, on their own test sets. They are a shortlist tool, not a prediction of what you will see. At the start of a project we run a short proof of concept on a sample of your own data, which replaces them with real figures — so the cost and the schedule for the full engagement are known before anything is committed.

CT analysis service AI Medical services Pricing

From benchmark to production

Send a representative sample, your expected volume and your latency target for ct analysis. We benchmark the shortlisted models, recommend the lowest-cost GPU configuration that meets the target, and scale it from pilot capacity to a dedicated production cluster — with the cost per unit of work known before you commit.