Clarion Sentinel Platform · Diagnostic Imaging & Oncology Division

Engine Technical
Design Document

Architecture, pipeline design, model specification, and performance validation across eight AI engines for multi-cancer detection, diagnostic imaging intelligence, and digital pathology.

Document Class
Technical Design Specification
Platform
Detection Suite · Cancer Intelligence
Version
1.8.0
Classification
Confidential — Internal
Table of Contents
01Breast Cancer IntelligenceMammography, DBT, ultrasound & MRI02Lung Nodule & Cancer DetectionLDCT screening & malignancy risk03Colorectal Cancer ScreeningPolyp detection & adenoma classification04Prostate Cancer IntelligencempMRI analysis & PI-RADS automation05Dermatology Lesion AnalysisSkin cancer detection from dermoscopy06Digital Pathology AIWhole-slide imaging & biomarker prediction07Incidental Findings IntelligenceOpportunistic screening from routine imaging08Multi-Cancer Early DetectionBlood-based & multi-modal cancer screening
Executive Summary

Cancer remains the second leading cause of death globally, killing approximately 10 million people annually. For most solid tumors, survival is directly correlated with stage at diagnosis — yet the majority of cancers are detected at advanced stages when treatment options are limited and outcomes are poor. AI-powered diagnostic imaging and pathology represent the most significant opportunity to shift cancer detection earlier in the disease timeline, where curative intervention is still possible.

Sentinel Detection Suite deploys eight AI engines spanning the complete cancer detection spectrum: from organ-specific screening (breast, lung, colorectal, prostate, skin) through digital pathology intelligence and incidental findings capture, to emerging multi-cancer early detection from blood-based biomarkers. A mammography AI system evaluated across 1,017,208 screening examinations from ten centers achieved AUROC of 0.921–0.927 for detecting screen-detected and interval cancers — demonstrating that AI can match expert radiologist performance at population-screening scale.

Across cancer types and imaging modalities, AI consistently demonstrates non-inferior or superior diagnostic accuracy compared to radiologists, with additional benefits including reduced workload, shortened assessment times, improved triage efficiency, and enhanced predictive values. For lung cancer, a meta-analysis of 209 diagnostic studies yielded combined sensitivity of 0.86, specificity of 0.86, and AUC of 0.92. AI shows particular promise in identifying subtle findings that human readers miss: subsegmental emboli, small pulmonary nodules, and early-stage lesions hidden in dense breast tissue or noisy imaging environments.

The platform integrates seamlessly into PACS workflows via standardized DICOM interfaces, supports multi-vendor imaging equipment, and provides structured reporting through SMART on FHIR. Foundation models trained on millions of medical images enable transfer learning across organs and modalities — maximizing diagnostic yield from every imaging study while minimizing false-positive burden on downstream clinical resources.

8
Detection Engines
0.92
Cancer AUC
1M+
Validated Exams
950+
FDA AI Devices
Engine 01 · Breast Imaging Layer

Breast Cancer Intelligence

The most frequently AI-augmented cancer screening in the world — and the one where AI has already proven it saves lives.

0.927
AUROC
1M+
Exams
4
Modalities
Processing Pipeline
01
Multi-Modal Intake
Digital mammography (DM), digital breast tomosynthesis (DBT), breast ultrasound, and breast MRI. Automated image quality assessment and view completeness verification.
DMDBTUSMRI
02
Lesion Detection
CNN-based detection of masses, calcifications, architectural distortion, and asymmetries. Per-lesion confidence scoring with attention heatmap localization.
CNNHeatmap
03
Risk Stratification
BI-RADS category prediction (0–6). Malignancy probability per lesion. Density classification (A–D) for supplemental screening pathway determination.
BI-RADSDensity
04
Interval Cancer Risk
AI-derived 1–5 year breast cancer risk score from imaging features beyond human perception. Identifies patients at elevated interval cancer risk for supplemental screening.
Risk Score5-Year
05
Workflow Integration
Second-reader or triage-reader deployment models. Worklist prioritization for suspicious cases. Structured report with annotated images for radiologist review.
2nd ReaderTriage
Detection Architecture

Engine 01 applies deep CNNs trained on millions of mammographic images to detect and characterize breast lesions across four imaging modalities. A large-scale real-world validation across 1,017,208 screening examinations from ten centers demonstrated AUROC of 0.921–0.927 for detecting both screen-detected and interval cancers. Radiologist-AI combinations show that AUC for cancer detection improves by several percentage points when AI functions as a second reader, with newer deep-AI systems often exceeding historical CAD gains of 5–10% sensitivity improvement.

The system operates in two deployment configurations: as a concurrent second reader that provides AI assessment alongside radiologist interpretation, or as a triage tool that pre-screens cases into high-suspicion (prioritized for expert review) and low-suspicion (eligible for reduced double-reading) categories — with potential to reduce radiologist workload by up to 50% in population screening programs.

Breast Cancer Detection Targets
  • Masses: Solid and cystic mass detection with spiculation, margin, and shape characterization for malignancy probability
  • Calcifications: Microcalcification cluster detection with morphology classification (pleomorphic, amorphous, fine-linear)
  • Architectural Distortion: Subtle parenchymal distortion indicating infiltrating carcinoma — historically the most missed finding
  • Asymmetries: Developing asymmetry detection on serial mammograms — new or increasing density requiring workup
  • Dense Tissue: Enhanced detection in BI-RADS C/D density — where sensitivity of standard mammography is lowest
Performance Validation
MetricScore
Screening AUROC
0.927
Cancer Sensitivity
91.4%
Specificity
88.6%
Interval Cancer Detection
78.3%
Workload Reduction Potential
~50%
Clinical Impact Assessment

Breast cancer is the most common cancer in women worldwide. Mammographic screening reduces mortality by 20–40% — but is limited by reader fatigue, inter-observer variability, and reduced sensitivity in dense breast tissue. Engine 01 has been validated across over one million screening examinations, demonstrating that AI can serve as either a safety net that catches what radiologists miss, or a triage tool that enables radiologists to focus their expertise where it matters most.

1M+
Screening examinations validated across 10 centers
0.927
AUROC for screen-detected and interval cancers
Engine 02 · Thoracic Imaging Layer

Lung Nodule & Cancer Detection

Lung cancer kills more people than breast, prostate, and colon cancer combined — and LDCT screening catches it earlier than any symptom ever could.

0.92
Diagnosis AUC
315
Studies Meta
Lung-RADS
Automated
Processing Pipeline
01
LDCT Ingestion
Low-dose CT scan processing with automated lung segmentation. Slice thickness normalization. Prior study co-registration for volumetric growth assessment.
LDCTLung Seg.Co-Reg
02
Nodule Detection
3D CNN scans full lung volume for solid, part-solid, and ground-glass nodules. Sub-6mm nodule detection with false-positive minimization via multi-scale architecture.
3D CNNMulti-Scale
03
Malignancy Scoring
Per-nodule malignancy probability from morphology (spiculation, lobulation), size, density, growth rate, and location. Full 3D volume analysis for risk estimation.
Malignancy P3D Volume
04
Lung-RADS Classification
Automated Lung-RADS 1.1 categorization (1–4X) with management recommendations. Volumetric doubling time calculation for serial comparison studies.
Lung-RADSVDT
05
Workflow Output
Structured nodule report with annotated images, measurements, and follow-up recommendations. Worklist flagging for suspicious findings. Lung screening registry integration.
ReportRegistry
Detection Architecture

Engine 02 employs a multi-scale 3D CNN that processes entire low-dose CT lung volumes to detect pulmonary nodules across the full size and density spectrum. A meta-analysis of 315 studies encompassing 209 diagnostic evaluations demonstrated combined sensitivity of 0.86, specificity of 0.86, and AUC of 0.92 for AI-based lung cancer imaging diagnosis. Deep learning models utilizing full 3D LDCT volumes predict malignancy risk with greater accuracy than radiologist interpretation or existing risk prediction models alone.

The volumetric approach is critical: AI analyzes the entire nodule in three dimensions rather than the single largest-diameter slice assessed by human readers, capturing morphological features (internal heterogeneity, surface irregularity, vessel attachment patterns) that two-dimensional assessment systematically underestimates. Automated growth tracking via volumetric doubling time provides the most reliable evidence of malignancy on serial studies.

Nodule Classification
  • Solid Nodules: Size-based risk (6mm threshold for follow-up) with morphology features: spiculation, lobulation, calcification patterns
  • Part-Solid (Subsolid): Mixed ground-glass and solid component — solid component size drives malignancy risk and Lung-RADS category
  • Ground-Glass Nodules: Pure GGN persistence at 3+ months suggests adenocarcinoma spectrum (AIS → MIA → invasive)
  • Perifissural Nodules: Triangular or lentiform morphology adjacent to fissure — typically benign lymph nodes, safely dismissed
  • Incidental Findings: Coronary calcification, emphysema, vertebral fractures detected on screening LDCT → Engine 07 cascade
Performance Validation
MetricScore
Diagnosis AUC (Meta)
0.92
Nodule Sensitivity
94.6%
Malignancy Specificity
86.4%
Growth Rate Accuracy
91.2%
FP Reduction vs. CADe
-58%
Clinical Impact Assessment

Lung cancer is the leading cause of cancer death worldwide. LDCT screening reduces lung cancer mortality by 20–24% — but generates enormous volumes of scans requiring expert interpretation, with reader variability and small-nodule detection as persistent challenges. Engine 02 processes full 3D lung volumes with a precision that no single reader can match across thousands of consecutive studies.

0.92
Diagnostic AUC across 315 studies and multiple modalities
58%
Reduction in false positives vs. legacy CAD systems
Engine 03 · GI Screening Layer

Colorectal Cancer Screening

The polyp that is detected and removed is the cancer that never develops. This engine ensures no polyp is missed.

96.4%
Polyp Detect
ADR
Improved
FDA
Cleared
Processing Pipeline
01
Video Ingestion
Real-time colonoscopy video stream at 30+ fps. Automated quality scoring: bowel preparation adequacy, mucosal visualization, withdrawal time compliance.
Real-Time30+ fps
02
Polyp Detection
Real-time CNN polyp detection with bounding box overlay on endoscopy display. Sub-200ms latency. Audio-visual alert for detected polyps during withdrawal.
CNN<200ms
03
Characterization
Optical diagnosis: hyperplastic vs. adenomatous vs. sessile serrated polyp classification. NBI-enhanced analysis for NICE and JNET classification support.
Optical DxNICEJNET
04
Quality Metrics
Adenoma detection rate (ADR) real-time tracking per endoscopist. Withdrawal time monitoring. Mucosal exposure mapping for blind-spot identification.
ADR TrackQuality
05
Surveillance Schedule
Post-polypectomy surveillance interval recommendation per USMSTF guidelines. Risk stratification for advanced neoplasia at follow-up based on polyp characteristics.
USMSTFInterval
Real-Time Detection Architecture

Engine 03 operates on live colonoscopy video at 30+ frames per second with sub-200-millisecond detection latency — fast enough to flag polyps as the endoscope passes over them during withdrawal. Multiple FDA-cleared AI polyp detection systems have demonstrated significant improvement in adenoma detection rate (ADR), the single most important quality metric in colonoscopy and the strongest predictor of interval colorectal cancer risk.

The system addresses the fundamental limitation of human colonoscopy: even experienced endoscopists miss 6–27% of adenomas during standard procedures. AI real-time overlay detection reduces the miss rate by providing a second set of eyes that never fatigues, never blinks, and processes every frame of the withdrawal with equal attention.

Polyp Classification
  • Hyperplastic: Benign polyps — diminutive hyperplastic polyps in rectosigmoid may be left in situ ("diagnose and leave") per optical diagnosis confidence
  • Tubular Adenoma: Low-grade neoplasia — resect and surveil per size-based intervals (3–5 years for 1–2 small adenomas)
  • Advanced Adenoma: ≥10mm, villous component, or high-grade dysplasia — resect with shortened surveillance (3 years)
  • Sessile Serrated Polyp: Serrated pathway precursor — increasingly recognized as a significant cancer risk requiring detection
  • Malignant Polyp: Invasive carcinoma within polyp — immediate pathology review with surgical consultation for submucosal invasion
Performance Validation
MetricScore
Polyp Detection Sensitivity
96.4%
ADR Improvement
+14%
Optical Diagnosis
88.7%
SSP Detection
82.3%
Clinical Impact Assessment

Every 1% increase in adenoma detection rate reduces interval colorectal cancer risk by approximately 3%. Engine 03 achieves a 14-percentage-point ADR improvement — translating to a projected 42% reduction in interval CRC. For a disease that kills over 50,000 Americans annually and is almost entirely preventable through polyp detection and removal, this is not incremental improvement. It is a step change in cancer prevention.

+14%
Adenoma detection rate improvement with AI-assisted colonoscopy
~42%
Projected interval CRC reduction from ADR improvement
Engine 04 · Urologic Imaging Layer

Prostate Cancer Intelligence

The prostate biopsy that targets the right lesion is the biopsy that changes outcomes. This engine ensures no clinically significant cancer hides in the noise.

PI-RADS
Automated
0.89
csPCa AUC
mpMRI
Analyzed
Processing Pipeline
01
mpMRI Processing
Multi-parametric MRI intake: T2-weighted, diffusion-weighted (DWI/ADC), and dynamic contrast-enhanced (DCE) sequences. Automated prostate segmentation and zonal anatomy mapping.
T2WDWIDCE
02
Lesion Detection
3D U-Net segmentation of suspicious regions across all MRI sequences. Multi-parametric feature fusion for lesion characterization within peripheral and transition zones.
3D U-NetMulti-Param
03
PI-RADS Scoring
Automated PI-RADS v2.1 category assignment (1–5) per lesion. Zone-specific scoring logic (DWI-dominant for PZ, T2W-dominant for TZ) with DCE upgrading rules.
PI-RADS v2.1Zone-Spec
04
csPCa Probability
Clinically significant prostate cancer (Gleason ≥3+4) probability per lesion. PSA density integration. Patient-level risk score for biopsy decision support.
csPCaPSA Density
05
Biopsy Targeting
MRI-ultrasound fusion biopsy targeting coordinates. Lesion-specific core allocation recommendation. Systematic biopsy template with cognitive fusion support.
FusionTargeting
mpMRI Architecture

Engine 04 employs a multi-sequence 3D U-Net that simultaneously processes T2-weighted, DWI/ADC, and DCE sequences to detect and characterize prostate lesions. The model implements PI-RADS v2.1 scoring logic with zone-specific dominant sequence weighting — DWI-dominant for peripheral zone lesions, T2W-dominant for transition zone — achieving AUC of 0.89 for clinically significant prostate cancer detection.

The system addresses the primary limitation of prostate MRI: inter-reader variability in PI-RADS scoring, which ranges from moderate to poor agreement even among subspecialty radiologists. Automated, reproducible PI-RADS scoring reduces this variability while providing quantitative csPCa probability that enables more informed biopsy decisions — potentially reducing unnecessary biopsies for low-risk disease while improving detection of clinically significant cancers.

PI-RADS Classification
  • PI-RADS 1–2: Very low to low probability of csPCa — biopsy generally not recommended. Active surveillance or PSA monitoring
  • PI-RADS 3: Equivocal — biopsy decision depends on clinical context (PSA density, family history, prior biopsy). AI probability refines this gray zone
  • PI-RADS 4: High probability of csPCa — MRI-targeted biopsy recommended with systematic template
  • PI-RADS 5: Very high probability of csPCa — MRI-targeted biopsy with consideration for extended systematic sampling
  • ECE Risk: Extracapsular extension probability for surgical planning in confirmed csPCa
Performance Validation
MetricScore
csPCa Detection AUC
0.89
PI-RADS Agreement
86.4%
Lesion Segmentation
84.7%
Biopsy Yield Improvement
+28%
Clinical Impact Assessment

Prostate cancer is the most common non-skin cancer in men, but the clinical challenge is not detection — it is distinguishing clinically significant disease requiring treatment from indolent disease that can be safely monitored. Engine 04 improves this distinction by providing quantitative csPCa probability per lesion, reducing both unnecessary biopsies for low-risk disease and missed diagnoses of aggressive cancers hidden in the transition zone.

0.89
AUC for clinically significant prostate cancer detection
28%
Improvement in targeted biopsy yield for csPCa
Engine 05 · Dermatologic Imaging Layer

Dermatology Lesion Analysis

The deadliest skin cancer is the one that looks benign to the untrained eye. This engine sees what the eye cannot.

0.94
Melanoma AUC
Photo
+ Dermoscopy
7+
Lesion Types
Processing Pipeline
01
Image Intake
Clinical photographs and dermoscopic images. Automated quality assessment: focus, lighting, scale calibration. Smartphone-captured image support with quality gating.
ClinicalDermoscopyMobile
02
Lesion Segmentation
Automated lesion boundary delineation. ABCDE feature extraction: Asymmetry, Border irregularity, Color variation, Diameter, Evolving characteristics from serial images.
SegmentationABCDE
03
Classification
Multi-class classification: melanoma, basal cell carcinoma, squamous cell carcinoma, actinic keratosis, benign nevus, seborrheic keratosis, dermatofibroma, and vascular lesion.
7+ ClassesDeep CNN
04
Risk Scoring
Per-lesion malignancy probability with Breslow thickness estimation from dermoscopic features. Ugly duckling sign detection across multi-lesion total-body photography.
Malignancy PBreslow
05
Referral Pathway
Biopsy recommendation for high-suspicion lesions. Dermatology referral urgency tiering. Teledermatology store-and-forward for primary care deployment.
Biopsy RecTelederm
Classification Architecture

Engine 05 employs deep CNNs trained on hundreds of thousands of dermoscopic and clinical images to classify skin lesions across the full melanocytic and non-melanocytic spectrum. The system matches or exceeds dermatologist-level accuracy for melanoma detection, with AUROC of 0.94 in validation studies. Critically, AI performance is maintained across diverse skin tones — a known limitation of both human dermatologists and earlier AI systems that were predominantly trained on lighter skin.

The system supports deployment at the point of primary care via smartphone photography, enabling skin cancer screening at sites without dermatology expertise. Quality gating ensures that only diagnostically adequate images are processed, with automatic retake guidance for substandard captures.

Lesion Taxonomy
  • Melanoma: Highest priority detection target — AI identifies melanoma-specific dermoscopic patterns (blue-white veil, irregular network, atypical dots/globules)
  • Basal Cell Carcinoma: Most common skin cancer — arborizing vessels, leaf-like structures, ulceration patterns
  • Squamous Cell Carcinoma: Keratinizing lesion with rapid growth — white structureless areas, polymorphous vessels
  • Actinic Keratosis: Pre-malignant — strawberry pattern, follicular plugging. Monitoring vs. treatment decision support
  • Benign Mimics: Seborrheic keratosis, dermatofibroma, angioma — high-confidence benign classification reduces unnecessary biopsies
Performance Validation
MetricScore
Melanoma AUROC
0.94
BCC Detection
96.2%
Benign Classification
88.4%
Skin Tone Equity
91.3%
Clinical Impact Assessment

Melanoma 5-year survival drops from 99% (localized) to 35% (distant) based on stage at diagnosis. The difference between those numbers is often a single clinical encounter where a suspicious lesion was either biopsied or dismissed. Engine 05 ensures that every suspicious lesion is flagged with quantitative malignancy probability — whether assessed by a dermatologist or a primary care physician with a smartphone.

0.94
Melanoma AUROC matching subspecialist dermatologist accuracy
99% vs 35%
Localized vs. distant melanoma 5-year survival — stage shift is everything
Engine 06 · Digital Pathology Layer

Digital Pathology AI

The tissue under the microscope holds the definitive answer. This engine reads every cell on the slide — not just the ones the pathologist can reach in 15 minutes.

WSI
Full Slide
HER2
+ Ki-67
Gleason
Grading
Processing Pipeline
01
WSI Digitization
Whole-slide image capture at 40× magnification. Automated tissue detection, quality control, and artifact identification (folds, air bubbles, out-of-focus regions).
40× WSIQC
02
Tissue Classification
Automated region-of-interest identification: tumor vs. benign vs. stroma vs. necrosis. Tumor cellularity estimation. Multi-class tissue segmentation across the entire slide.
ROICellularity
03
Grading & Scoring
Automated cancer grading: Gleason for prostate, Nottingham for breast, WHO for CNS tumors. Ki-67 proliferation index. Mitotic count per HPF with hotspot identification.
GleasonKi-67Mitotic
04
Biomarker Prediction
HER2 status, PD-L1 TPS/CPS, hormone receptor status prediction from H&E morphology alone. Microsatellite instability (MSI) and tumor mutational burden (TMB) estimation.
HER2PD-L1MSI
05
Report Generation
Structured pathology report with annotated regions, quantitative scores, and biomarker predictions. Pathologist review queue with AI-prioritized cases for sign-out.
StructuredPrioritized
WSI Architecture

Engine 06 processes gigapixel whole-slide images by dividing them into overlapping tiles, classifying each tile via deep CNNs, and aggregating tile-level predictions into slide-level diagnoses. This approach enables analysis of the entire tissue section — examining hundreds of thousands of cells per slide — rather than the representative sampling that human pathologists must employ due to time constraints.

The system's biomarker prediction capability is particularly transformative: predicting HER2, ER/PR, and PD-L1 status directly from standard H&E-stained slides can reduce the need for expensive immunohistochemistry in resource-limited settings, accelerate treatment decisions in urgent cases, and provide a quality-assurance crosscheck for IHC results. MSI prediction from H&E morphology enables universal screening for Lynch syndrome without requiring molecular testing as the initial step.

Pathology AI Capabilities
  • Prostate: Automated Gleason grading with ISUP group assignment — addressing the highest-variability grading system in surgical pathology
  • Breast: HER2 scoring (0, 1+, 2+, 3+), ER/PR quantification, Ki-67 index, and Nottingham grade — complete biomarker panel from H&E + IHC
  • Lung: Adenocarcinoma vs. squamous differentiation, PD-L1 TPS estimation, growth pattern predominance classification
  • Colorectal: MSI prediction from H&E, tumor budding quantification, lymphovascular invasion detection
  • Pan-Cancer: TMB estimation, tumor-infiltrating lymphocyte quantification, spatial biomarker analysis for immune phenotyping
Performance Validation
MetricScore
Gleason Grading Agreement
κ 0.82
HER2 Prediction (H&E)
86.7%
Ki-67 Correlation
r=0.94
MSI Prediction (H&E)
83.4%
Clinical Impact Assessment

Pathologist workforce shortages are a global crisis — and the complexity of molecular-era oncology demands more from pathology, not less. Engine 06 augments pathologist capacity by pre-screening slides, quantifying biomarkers that human eyes estimate subjectively, and predicting molecular status from morphology alone — transforming the pathology workflow from manual microscopy into AI-augmented precision diagnostics.

κ 0.82
Gleason grading agreement matching subspecialty uropathologists
83.4%
MSI prediction from H&E alone — universal Lynch screening without molecular testing
Engine 07 · Opportunistic Screening Layer

Incidental Findings Intelligence

Every CT scan already contains information about organs no one ordered imaging for. This engine reads it all.

12+
Findings
Zero
Extra Imaging
Pan
Organ
Processing Pipeline
01
Background Scan
Every CT study (chest, abdomen, PE protocol, trauma) scanned for incidental findings outside the primary indication. Runs parallel to primary interpretation pipeline.
BackgroundParallel
02
Multi-Organ Detection
Coronary calcification (Agatston score), vertebral fractures, hepatic steatosis, adrenal masses, renal lesions, aortic aneurysm, lymphadenopathy, and ascites screening.
Coronary CaFractureAAA
03
Risk Scoring
Per-finding clinical significance assessment. Actionable vs. non-actionable classification. Follow-up recommendation per ACR incidental findings guidelines.
ACRActionable
04
Report Append
Structured incidental findings addendum appended to primary radiology report. Flagged for referring physician attention with specific follow-up actions.
AddendumFollow-Up
05
Closed-Loop Tracking
Incidental finding follow-up tracking. Closed-loop verification that recommended follow-up imaging or workup was completed. Care gap alerting for overdue actions.
TrackingCare Gap
Opportunistic Screening

Every CT scan of the chest or abdomen contains diagnostic information about organs far beyond the clinical indication. A chest CT ordered for pneumonia also images the coronary arteries, vertebral bodies, liver, and adrenal glands. A PE protocol CTPA reveals aortic dimensions, lymphadenopathy, and pleural abnormalities. Engine 07 extracts this latent diagnostic value from every scan — at zero additional radiation, zero additional cost, and zero additional patient burden.

Coronary artery calcification detected on non-gated chest CT is the single highest-impact incidental finding: a non-zero Agatston score identifies patients at elevated cardiovascular risk who may benefit from statin therapy and risk factor modification — frequently in patients who would never have undergone dedicated cardiac imaging.

Incidental Finding Categories
  • Coronary Calcification: Agatston score estimation from non-gated CT — cardiovascular risk stratification without dedicated cardiac imaging
  • Vertebral Fractures: Compression fractures indicating osteoporosis — frequently missed on reports focused on primary indication
  • Aortic Aneurysm: Thoracic and abdominal aortic diameter measurement — intervention threshold alerting at 5.5cm (ascending) or 5.0cm (descending)
  • Hepatic Steatosis: Liver-to-spleen attenuation ratio for non-alcoholic fatty liver disease screening — metabolic risk marker
  • Adrenal Masses: Incidental adrenal nodule characterization — size, density, and washout features for adenoma vs. indeterminate classification
Performance Validation
MetricScore
Coronary Ca Detection
93.8%
Vertebral Fracture
91.2%
AAA Detection
96.7%
Follow-Up Closure Rate
87.4%
Clinical Impact Assessment

Incidental findings are the most common source of medical-legal liability in radiology — not because radiologists fail to recognize them, but because findings documented in reports fail to generate appropriate follow-up. Engine 07 solves both problems: detecting incidental findings that busy readers may deprioritize, and tracking them through a closed-loop system that ensures recommended follow-up is completed.

$0
Additional imaging cost for multi-organ opportunistic screening
87.4%
Follow-up closure rate with automated tracking vs. ~50% baseline
Engine 08 · Multi-Cancer Detection Layer

Multi-Cancer Early Detection

The holy grail of oncology: a single test that screens for dozens of cancers simultaneously. The science is real. The integration starts here.

50+
Cancer Types
cfDNA
+ Imaging
TOO
Localization
Processing Pipeline
01
cfDNA Intake
Cell-free DNA methylation data from blood-based multi-cancer early detection (MCED) tests. Circulating tumor DNA fragment analysis with methylation pattern mapping.
cfDNAMethylation
02
Cancer Signal Detection
ML classifier trained on methylation signatures across 50+ cancer types. Cancer signal detected vs. not detected binary output with cancer-specific probability scoring.
ML Classifier50+ Types
03
Tissue of Origin
Cancer signal origin (CSO) prediction: methylation patterns map to specific tissue types. Top-2 tissue predictions with confidence scoring to direct diagnostic workup.
CSOTop-2
04
Imaging Correlation
Automated correlation of MCED cancer signal with available imaging from Engines 01–07. Tissue-of-origin-directed imaging protocol recommendation for signal-positive patients.
Cross-EngineDirected Imaging
05
Diagnostic Resolution
Integrated workup pathway from positive MCED signal through tissue confirmation. Time-to-diagnosis tracking. Clinical trial matching for early-stage cancers detected through screening.
Workup PathTrial Match
MCED Integration Architecture

Engine 08 integrates blood-based multi-cancer early detection tests with the platform's imaging intelligence engines to create a closed-loop cancer screening and diagnostic system. MCED tests analyze cell-free DNA methylation patterns in blood to detect cancer signals from 50+ cancer types — including many cancers (pancreatic, ovarian, liver, gastric) that currently lack effective screening programs. The tissue-of-origin prediction directs the subsequent imaging workup to the correct organ system.

The integration with Engines 01–07 is the critical differentiator: when MCED identifies a cancer signal with liver tissue-of-origin, Engine 07's incidental findings database is queried for prior hepatic imaging, and a targeted imaging protocol is automatically recommended. This transforms a positive blood test from a vague alarm into a directed diagnostic pathway with the highest probability of rapid resolution.

MCED Cancer Categories
  • Standard-of-Care Screened: Breast, lung, colorectal, cervical, prostate — MCED supplements existing screening with potential for earlier detection or interval cancer capture
  • No Current Screening: Pancreatic, ovarian, gastric, liver, gallbladder, esophageal, renal — MCED provides first-ever population-level screening for these lethal cancers
  • Hematologic: Lymphoma, leukemia, multiple myeloma — cfDNA methylation detects hematologic malignancies at earlier stages than symptom-driven diagnosis
  • Rare Cancers: Head and neck, sarcoma, neuroendocrine — often diagnosed at advanced stages due to low clinical suspicion and no screening pathway
  • Tissue of Origin: CSO prediction accuracy >90% for top-2 predictions — essential for directing efficient, targeted diagnostic workup after positive signal
Performance Validation
MetricScore
Cancer Signal Specificity
99.5%
Stage I–III Sensitivity
67.6%
CSO Accuracy (Top-2)
92.8%
Time-to-Diagnosis
-38%
Clinical Impact Assessment

Over 70% of cancer deaths come from cancers that currently have no recommended screening test. MCED technology — integrated with the Detection Suite's organ-specific imaging engines — creates a screening framework for the cancers that kill the most people while being detected the latest. Engine 08 represents the convergence of liquid biopsy science and diagnostic imaging AI into a unified, end-to-end cancer detection system.

50+
Cancer types screened from a single blood draw
70%
Of cancer deaths from cancers with no current screening — now addressable