deploying ai into the clinical radiology workflow

23
Dr. Judy W Gichoya Emory University Department of Radiology & Imaging Sciences @judywawira Deploying AI into the clinical radiology workflow Important considerations for radiologists and informaticists

Upload: others

Post on 15-Apr-2022

12 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Deploying AI into the clinical radiology workflow

Dr. Judy W Gichoya

Emory University Department of Radiology & Imaging Sciences

@judywawira

Deploying AI into the clinical radiology workflow – Important considerations for

radiologists and informaticists

Page 2: Deploying AI into the clinical radiology workflow

• No personal disclosures

• During infrastructure review may mention some of the equipment we have -> Not an endorsement

Page 3: Deploying AI into the clinical radiology workflow

• To describe pilot deployment of the ACR AI-Lab tool at Emory University

• To describe our architecture from development to production for inhouse and commercially available AI models

• Review ML data pipelines for working with AI in production

Page 4: Deploying AI into the clinical radiology workflow

Emory Radiology

Clinical Research Education

Page 5: Deploying AI into the clinical radiology workflow

Emory Department of Radiology and Imaging Sciences

Informatics Team

Revised 8/27/19

Imaging Workflow Special ists

Brenda Hall

Radiology

EUH

Jacqueline McCarty

Radiology

ESJH

Candace Moczarski

Radiology

EJCH

Wilbert Pope

Radiology

ESJH

Steve St. Louis

Radiology

EUH

Imaging Workflow Administrators

Karen Boles

Dir. of Informatics

and Tech Services

April Carter

Enter. Sol. Arch III

Project Manager

Kelli Miller

Project Manager

Project Management Team

Nabile Safdar, MD

Director

Imaging Informatics

Willie Arnold

Administrator,

Clinic Operations

Denise Fennell

Administrative

Assistant

Peter Harri, MD

Imaging

Informatics

Asst. Professor

Adam Prater, MD

Imaging

Informatics

Asst. Professor

Mercy Mutahi

Business Analyst

Judy Gichoya, MD

Imaging

Informatics

Asst. Professor

Marjin Brummer

Staff Scientist

Havi Trivedi, MD

Imaging

Informatics

Asst. Professor

Imaging Informatics

Starla Longfellow

Asst. Director

Imaging Srvcs.

Stacey Adams

Department System

Specialist II

Clinical Bus. Support

Chris Braithwaite

Desktop Support

EU IT Support

Wendy Lybrand

Enterprise Trainer

Trainer

Roslyn Baitey

EUHM

Mammo

Stacey Walker

EUHM

Evenings

Judy Graham

EUHM

Paulette Jackson

EUMHLennex Annor

EUHM

Evenings

Wangail Assamenew

Executive Park

MSK

Latonia Smith

Telerad

Nina Stephenss

EUMHErika Petty

Breast Imaging

(Winship)

Sharon Cain

TeleRadLisa Kappel

Supervisor

IWS

Lynn Coram-Allen

EUH

Interventional

Inez Dupree

EUH

Nuclear Medicine

Henok Abate

EUH

Neuroradiology

Dexter Bostic

EUH

Cardio

Gerald “Rick”

Foster

EUH

Brenda Stokes

TeleRad

Veena Amin

App Solutions

Analyst III

Jeff Trott

Technical Apps

Specialist III

Daria Miller

Enterprise

Solutions Arch II

RadNet Team

Jason Jacob

Dept. Systems

Specialist

Lee Taylor

Technical Apps

Specialist III

Lance Manley

Technical Apps

Specialist IV

Chris Vant

Technical Apps

Specialist III

Dario Rodriguez

Technical Apps

Specialist II

Rick Cobb

Technical Apps

Specialist III

Brian Goertemiller

Imaging System

Software Spec. Lead

Dongqing Shi

Imaging Application

Specialist

Luke Wademan

Enterprise

Solutions Arch III

Imaging Applications Support

Geo Eapen

Enterprise

Solutions Arch III

Collette Erickson

EJCH

Valerie

Fowler-Robinson

ESJH - Mammo

Ursula Handy-Evans

ESJH

Mary Greer Thomas

EJCH

Mammo

Lola Oceanty

ESJH

Outpatient

Sabrina Baida

Intern

Page 6: Deploying AI into the clinical radiology workflow

IVC filter case study

Page 7: Deploying AI into the clinical radiology workflow

IVC filter case study

• Complications if delayed removal• Venous thromboembolism

• Stent fracture

• IVC damage

• Death

• Frequent delayed removal:• Lack of patient awareness

• Improper identification on imaging

• Absence of follow up Cleveland Clinic Journal of Medicine. 2018 November;85(11):835-83

Page 8: Deploying AI into the clinical radiology workflow

Methods - Dataset

828 Radiographs : abdominal, thoracoabdominal, lumbar

Positive: With IVCF348

Negative: Without IVCF480

Strongly labeled with bounding boxesIR Fellowship in-training radiologist

Randomly divided without patient overlap

Training501

Validation127

Test200

Page 9: Deploying AI into the clinical radiology workflow

Methods - Technique

Retinanet architecture Encoder : Resnet-50

Pretrained : COCO dataset

Batch size : 1 Learning rate : 0.00001

15 epochs

Resolution : smallest side > 800 pixels, largest side < 1333 pixels No data augmentation : preserve high resolution IVCF spatial representation

Focal loss: γ=2 to compensate for pixel class imbalance

Metrics: Classification : AUC, Sensitivity, Specificity

Object Localization : Mean Average precision at IOU > 0.5 (mAP-0.5)

Lin T-Y, Goyal P, Girshick R, He K, Dollar P. Focal loss for dense object detection.

2017 IEEE International Conference on Computer Vision (ICCV). IEEE; 2017. p. 2999–3007.

Page 10: Deploying AI into the clinical radiology workflow

Methods – Baseline classifier

EfficientNet architecture B0 baseline subtype

Pretrained : Imagenet

Batch size : 8

Learning rate : 0.001 to 0.00001

30 epochs

Image resolution : 512 x 512 pixels

Standard affine data augmentation

BCE loss

Metrics: Classification : AUC, Sensitivity, Specificity

Mingxing Tan and Quoc V Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning, 2019.

Page 11: Deploying AI into the clinical radiology workflow

Findings - Performance

Page 12: Deploying AI into the clinical radiology workflow

Findings – Confusion Matrix

Page 13: Deploying AI into the clinical radiology workflow

What next?

Page 14: Deploying AI into the clinical radiology workflow

Metadata

• Started project over 1 year ago

• IRB for data – Retrospective versus umbrella versus prospective ?• Clinical trial ?

• Quality improvement project

• Exempt status

• Obtaining data • Existing IVC filter DB for different project – Accession No -> Manual extraction via

Osirix

• Deidentification pipeline with manual scrubbing -> and then manually verified

• Training – Single GPU 1080 Ti• Collaboration across institutions ?

Page 15: Deploying AI into the clinical radiology workflow

Metadata

• Operationalizing aka “Lets save lives !”• Research PACS – nightly build

• DICOM Headers – CXR, KUB, Lumbar and Thoracic radiographs

• Containers • .py file VS Jupyter notebook

• Docker for containers

• Anonymization – Needs to be identifiable

• Output / Monitoring • JPEG with bounding box -> Classification tasks ? What output - ? Probability, heat

maps ? segmentation output

• Clinical data integration -> Anticoagulants, Upcoming appts, Labs

Page 16: Deploying AI into the clinical radiology workflow

Background : Hardware

• 1 Titan X

• 3 Titan Vs

• Lambda

Page 17: Deploying AI into the clinical radiology workflow

Background: Network

DMZ

Healthcare

Academic

Outside Traffic

Page 18: Deploying AI into the clinical radiology workflow

Background : Software

• Annotation – Vision tasks, NLP

• Data science – Jupyter Hub + Jupyter notebooks . Pandas • Support ease of sharing + versioning

• DL tools – Tensorflow, Pytorch, Keras, FastAI (python shop)

• Front end – Web based (Flask apps, Bootstrap)

• Code – Versioning (private repos) , Tickets/ Sprints for work management, Parameters tracking for multiple engineers

• Deployment – Containers

• Not cloud interactive

• Open source

Page 19: Deploying AI into the clinical radiology workflow

Success looks like

IVC filter detector PACS stream HL7 clinical feed Communication

- CXR /KUB/ Spine Xrays

- Follow up imaging

- IVC filter present / removed

Labs Medication list

Problem list Clinic referral

Page 20: Deploying AI into the clinical radiology workflow

ACR AI-Lab

Page 21: Deploying AI into the clinical radiology workflow

ACR AI-Lab at Emory

• IRB submission – Umbrella

• Aim for access to duplicate/research PACS

• Authentication

• Anonymization challenge

• Hardware – Inference GPU

Page 22: Deploying AI into the clinical radiology workflow

Summary

• Avoid #FOMO ….. Manage the hype

• Build team

• T vs N

• Metadata aggregation • New data streams

• Access integration engines

• Resources – Leadership, Align with CMIO , Business case • Build or buy ?

• Don’t forget to “pay the rent”

• Southern AI club?

Page 23: Deploying AI into the clinical radiology workflow

Contact / Feedback

[email protected]

@judywawira