MEDICAL_IMAGING_AI
X-ray&CT AI
Production-grade segmentation and classification with MONAI Label, MONAI Deploy, and OHIF v3.11
2024–2025 Stack
•
DICOM-Native
•
Viewer-Ready
LIVE_PROCESSING
1847
studies processed
104
anatomies segmented
4.2s
segmentation time
MONAI Label
v2024.11
MONAI Deploy SDK
v3.2.0
OHIF Viewer
v3.11
Research AUCs without DICOM I/O don't make it to the reading room
Hospitals need workflow-integrated AI, not one-shot demos
1
Ingest DICOM from PACS/VNA
[Required]
2
Fast, reproducible on-prem inference
[Required]
3
Reviewable overlays (SEG/RTSTRUCT)
[Required]
4
Provenance & audit logging
[Required]
5
Scale to thousands of studies/day
[Required]
Production stack with MONAI Label for serving, MONAI Deploy SDK for packaging, and OHIF v3.11 for in-viewer overlays
STACK
2024–2025 Releases
Open, Current, Clinic-Ready
1
MONAI Label
v2024.11 (Nov 22, 2024)
Whole-Body CT (104 anatomies in ~4s)
Keycloak multi-user RBAC
2
MONAI Deploy App SDK
v3.0.0 (Apr 2025) / v3.2.0 (Sep 2025)
Holoscan v3 alignment
PyTorch 2.7–2.8 / CUDA ≥12.6
3
OHIF Viewer
v3.10 / v3.11 (2025)
WebGPU GrowCut (one-click 3D)
DICOM Labelmap Segmentations IOD
Clinic-Ready Flow
PACS/VNA
→
Informatics Gateway
C-STORE/C-MOVE, FHIR links
↓
MONAI Deploy Orchestrator
App SDK v3.x • MAP routing
↓
MONAI Label Server
GPU • RBAC via Keycloak
• CT bundles (Whole-Body 104 anatomies)
• X-ray classifiers / 2D segmenters
↓
DICOM SEG / RTSTRUCT / Labelmap
+ JSON metadata
→
OHIF v3.11 VIEWER
Gateway handles DICOM I/O • App SDK binds series → inputs/outputs • OHIF renders overlays with WebGPU tools
Model Portfolio
Curated for X-ray & CT
1
MONAI Label Whole-Body CT
CLASSES
104 anatomies
SPEED/INFO
~4 seconds
USE_CASE
Default coverage model for multi-organ CT
2
TotalSegmentator v2
CLASSES
≥117 classes
SPEED/INFO
SNOMED mappings
USE_CASE
Exhaustive anatomy list with clinical codes
3
nnU-Net v2
CLASSES
Site-specific
SPEED/INFO
Auto-config
USE_CASE
Task-adaptive fine-tuning for local protocols
4
SAM-Med2D
CLASSES
31 organs
SPEED/INFO
4.6M images trained
USE_CASE
Promptable 2D segmentation for X-ray & QA
Modalities & I/O
🏥
CT (3D)
PIPELINE
Import via Gateway → resample/normalize → model → DICOM SEG (preferred) or RTSTRUCT
VIEWER
OHIF 3.11 can load Labelmap for speed on big volumes
📷
X-ray (2D)
PIPELINE
Classify/triage → heatmap/contour to Secondary Capture or SEG for key regions
VIEWER
SAM-Med2D for promptable masks and reader QA
OHIF WebGPU GrowCut (v3.10)
Rapid edits directly in the viewer for one-click 3D segmentation
Case Blueprints
A
Chest CT — Lungs/Lobes/Airways + Nodule Assist
Speed quantification & review
01
PIPELINE
Series selection → Whole-Body CT or TotalSegmentator → lobe split, airway pruning → lung/lobe volumes → nodule candidates → SEG + JSON
VIEWER
OHIF shows masks and radiomics table; GrowCut for edits
TARGETS
Lung/lobe Dice ≥ 0.95/0.90 • P95 < 25s • >95% overlay acceptance
B
Abdomen CT — Liver/Pancreas/Spleen + Vessels
Planning support; consistent volumetry
02
PIPELINE
Multi-organ bundle + vessel centerlines → organ volumes, aortic diameter map → export RTSTRUCT for planning
VIEWER
Treatment planning system import
TARGETS
Organ Dice ≥ 0.90 • Centerline continuity > 98% • P95 < 40s
C
Chest X-ray — Triage + Explainability
Flag likely abnormal CXRs; interpretable regions
03
PIPELINE
Classifier → heatmap → prompt SAM-Med2D for ROI mask → SC or SEG back
VIEWER
OHIF for overlay review
TARGETS
AUC ≥ 0.92 on site validation • Overlay render < 2s
Performance & Rollout
PERFORMANCE_TARGETS
CT multi-organ
≤60s/study (batch)
Scale via GPU pools
X-ray triage
<1s classify
Heatmaps return first
Viewer overlay
Near-instant
Labelmap reduces latency
PACS→AI→PACS
≥99.5% success
Retries & back-pressure
ROLLOUT_TIMELINE
1–2
Connect PACS↔Gateway • 100-case golden set • Baseline Dice
3–4
Site tuning (spacing, windowing) • nnU-Net fine-tunes • OHIF workflow
5
Shadow deploy • Nightly batch • Reader feedback instrumented
6–8
Canary 10% • KPI gates (success, latency, Dice) • Scale
Packaging
Orchestration
Monitoring
Promotion
Not just models
the operational substrate
104
Anatomies Segmented
<4s
Whole-Body CT
99.5%
Success Rate
v3.11
OHIF Viewer
MONAI_LABEL_v2024.11
•
MONAI_DEPLOY_SDK_v3.2.0
•
OHIF_v3.11
Frequently Asked Questions
Which imaging modalities does TensorBlue support?
CT, X-ray, MRI, ultrasound, mammography, and digital pathology WSI. We deploy DICOM-native pipelines that integrate with PACS and reporting workflows.
How accurate are your medical imaging AI models?
Production accuracy depends on the task and dataset, but our deployed radiology and pathology systems consistently meet or exceed published clinical benchmarks (sensitivity 0.90+, specificity 0.92+ on independent test sets).