CT · PET/CT · MRI
LymphNode AI
Detect, segment, and measure lymph nodes across the body — for staging and follow-up.
510(k) pathway in progress
OncoAI Suite · AI-Native Oncology Imaging Suite
Detection, segmentation, and longitudinal tracking across cancer types — on one workstation
Six AI engines under a single viewer: BrainTumor in clinical use today, LymphNode on the 510(k) pathway, LungNodule in clinical research use, and kidney / liver / whole-body PET/CT in development. CT, MRI, and PET/CT — same UI, same pipeline.
The suite
Six AI engines under one workstation
Multi-parametric MRI
BrainTumor AI
Glioma, brain metastasis, and vestibular schwannoma — segmented with volumetric tracking across visits.
Validated · in clinical use
Chest CT
LungNodule AI
Detection, segmentation, and Lung-RADS scoring on chest CT — semi-automatic or fully automatic.
In clinical research use
CT (contrast)
KidneyTumor AI
Automatic segmentation of renal tumors and cysts — for staging, surveillance, and surgical planning.
In development
CT (contrast)
LiverTumor AI
Automatic segmentation of liver tumors from contrast-enhanced CT for response assessment.
In development
PET/CT
PET/CT Whole-body AI
Whole-body lesion segmentation on PET/CT — coregistered against CT anatomy.
In development
How it works
A three-stage AI pipeline tuned for cancer-imaging realities
Cancer targets are small, low-contrast, and unpredictably located. A single segmentation model misses them or floods the report with false positives. OncoAI separates the problem into three networks — each tuned to a different objective — and the LymphNode model is the most-validated example.
01
Detect — maximize sensitivity
RetinaUNet3D scans patch-by-patch to surface every candidate. Recall is prioritized so nothing is missed downstream.
02
Triage — maximize AUC
EfficientNetV2S with 3-panel augmentation rules out false positives without sacrificing the sensitivity from stage 1.
03
Segment — voxel-precise masks
UNet3D produces masks for each confirmed finding. Long/short-axis and volume measurements come directly from the masks.
666
Scans · LymphNode validation
625 CT + 41 MRI — head & neck, thorax, abdomen, pelvis. Train 444 / val 222.
80%
True-positive rate (distance-hit)
79.93% sensitivity at 1.27 false positives per true positive on the 222-scan validation set.
0.69
Dice (node-level segmentation)
0.6922 ± 0.2283 mean ± SD on 212 scans with ground-truth masks.
Publications
Cancer-imaging AI methods, peer-reviewed
-
EXACT-Net: Framework for EHR-guided lung tumor auto-segmentation for non-small cell lung cancer radiotherapy
-
Brain tumor segmentation for multi-modal MRI with missing information
-
Head and neck tumor segmentation in PET/CT: The HECKTOR challenge
-
Deep learning-based automatic tumor burden assessment of pediatric high-grade gliomas, medulloblastomas, and other leptomeningeal seeding tumors
-
Prediction of lung malignancy progression and survival with machine learning based on pre-treatment FDG-PET/CT
-
Lung nodule malignancy prediction in sequential CT scans: Summary of ISBI 2018 Challenge
See OncoAI on your modality
On-premise web service · DICOM C-FIND / push / upload · NIfTI / NRRD · DICOM Seg / CSV / report out · Batch CLI for cohort runs.