Medical Imaging Research

Advancing diagnostics through intelligent imaging

Porcupine is a medical imaging research company founded by engineers from Microsoft, Meta, and Google alongside clinical researchers. We build open, reproducible processing pipelines that help radiologists and pathologists work more accurately at scale.

Radiology is drowning in data

Global imaging volumes are surging while the workforce to interpret them is shrinking. The diagnostic gap is widening every year.

images per CT exam over time
1999 82 images / exam 2010 679 images / exam (8×)

Over 3.6 billion imaging exams are performed worldwide each year (WHO 2016). A single CT averaged 82 images in 1999; by 2010 that had surged to 679—an 8× increase per exam.

McDonald et al., Academic Radiology 2015
images per radiologist per year
2.6M 1.3M 0 467K 2.6M 2005 2020

A 561% increase in images per radiologist over 15 years. On-call workloads have quadrupled. To keep pace, a radiologist must evaluate one image every 3–4 seconds across an 8-hour shift.

Peng et al., European J. of Radiology 2022; Bruls & Kwee, Insights into Imaging 2020
U.S. radiology workforce crisis

More than 53% of the ~21,000 radiologists in active patient care are 55 or older, and burnout rates exceed 50%. Half of all U.S. radiology positions went unfilled in 2023. AAMC projects up to 42,000 specialist shortfall by 2033.

ACR Congressional Testimony 2023; AAPPR Benchmarking Report 2024
diagnostic error rate (persistent since 1949)
Retrospective errors ~30% unchanged since 1949 Real-time clinical 3–5% ~40M mistakes / year

The retrospective error rate has held near 30% since Garland's landmark 1949 study. Real-time clinically significant errors run 3–5%, translating to ~40 million diagnostic mistakes per year worldwide.

Garland, Radiology 1949; Brady, RadioGraphics 2018

10–30% of breast cancers are missed on screening mammography. 19% of lung cancers are overlooked on initial chest radiographs. Diagnostic error is the leading cause of malpractice claims in radiology. Across all conditions, diagnostic errors cost the U.S. healthcare system an estimated $100 billion annually, and contribute to roughly 800,000 deaths or permanent disabilities each year.

Ekpo et al., Lancet Digital Health 2020; Quekel et al., Chest 1999; Newman-Toker et al., BMJ Quality & Safety 2023

AI-assisted imaging is inflecting now

Clinical evidence, regulatory momentum, and economic pressure are converging to make intelligent imaging infrastructure a necessity, not a luxury.

FDA-authorized AI/ML devices per year
300 150 0 '19 '20 '21 '22 '23 '24 295 '25

Over 1,300 AI/ML devices authorized to date, more than 75% in radiology. A record 295 clearances in 2025 with a median review of just 142 days.

FDA AI-Enabled Device List; Innolitics 2025 Year in Review
AI-assisted sensitivity gain
Lung-nodule: without AI 72.8% sensitivity Lung-nodule: with AI 83.5% +10.7 pp gain

Multi-reader trials show +10.7 pp sensitivity gain in lung-nodule detection without increasing false positives. In prostate MRI (10,207 exams, 62 radiologists), AI AUROC of 0.91 vs. 0.86.

Robert et al., Academic Radiology 2024; Saha et al., Lancet Oncology 2024
AI-in-medical-imaging market ($B)
$8B $4B $0 $1.5B ~$7.5B 2020 2024 2030

Market reached $1.5 billion in 2024. Projected 30–35% CAGR through 2030, driven by workforce shortages, rising imaging volumes, and value-based reimbursement demands.

Nova One Advisor; Grand View Research; Mordor Intelligence
reading time with AI assistance
Brain-tumor: without AI 100% reading time Brain-tumor: with AI −31% reading time

AI triage and pre-annotation cut interpretation time by up to a third. In brain-tumor detection, AI reduced reading time by 31% while boosting sensitivity from 82.6% to 91.3%.

Lu et al., Neuro-Oncology 2021; RSNA Radiology 2023

End-to-end medical imaging intelligence

From raw DICOM ingestion to clinical-grade inference, our platform handles the full lifecycle of medical image analysis.

Intelligent Acquisition

Real-time quality assessment during image capture. Our models flag suboptimal scans before the patient leaves, reducing repeat visits.

Multi-Modal Fusion

Combine CT, MRI, PET, and histopathology data into unified 3D representations for holistic analysis of patient anatomy.

Anomaly Detection

Proprietary transformer architectures trained on 50M+ annotated studies to surface subtle pathological patterns early.

HIPAA-Grade Security

End-to-end encryption, de-identification pipelines, and SOC 2 Type II compliance built into every layer of the stack.

Pipeline Orchestration

Visual DAG builder for constructing reproducible analysis workflows. Version every model, dataset, and parameter.

Clinical Analytics

Dashboards for tracking model performance, drift detection, and cohort analysis across your entire imaging fleet.

Who we are

Our team sits at the intersection of large-scale systems engineering and clinical imaging science.

MK

Michael

CEO · Research Engineer

Former senior engineer at Microsoft Research and Google. Leads company strategy and oversees the core imaging research agenda, bridging production engineering with clinical validation.

AE

Arthur

Engineering

Ex-Meta and ex-Microsoft infrastructure engineer. Designs and builds the distributed processing backbone—DICOM ingestion, model serving, and HIPAA-compliant data pipelines at scale.

AH

Aaron

Engineering

Previously at Google Cloud. Focuses on pipeline orchestration, reproducible ML workflows, and the platform tooling that connects research prototypes to clinical deployment.

AG

Anupam

Clinical Research

Background in clinical radiology and epidemiology. Leads multi-site validation studies and ensures that every model meets clinical evidence standards before deployment.

NS

Nitin

AI Research

Deep learning researcher specializing in medical foundation models. Develops the anomaly detection and multi-modal fusion architectures at the core of the Porcupine platform.

Interested in collaborating?

We work with hospital systems, academic labs, and biotech companies. Reach out to discuss research partnerships, platform access, or clinical validation studies.

[email protected]

We typically respond within one business day.