Using Deep Learning for Earlier Detection of Acute ... · Algorithm 2 Performance on Stroke: 62% Sensitivity (CI95%, 38-82%). 96% Specificity (CI95%, 82-100%), with AUC = 0.81 True

Confidential │ Trade Secret │ Proprietary │ Do Not Copy

1

Confidential │ Trade Secret │ Proprietary │ Do Not Copy Strategy and Planning │The Ohio State University Wexner Medical Center © 2017

Using Deep Learning for Earlier Detection of Acute

Infarction of the Brain

Barbaros S. Erdal, Ph.D.

Department of Radiology

The Ohio State University Wexner Medical Center

May, 2017


2


2What is New in the Field: Radiology

Positive

(Gold standard)

Negative

(Gold Standard)

Total

Positive

(AI algorithm)

38 5 43

Negative

(AI Algorithm)

2 35 37

Total 40 40 80

Neurologic Disease Medical Imaging Informatics

(e.g., Artificial Intelligence)

Cardiovascular DiseaseFast MRI

(e.g., MRE, 4D Flow)

Cancer Low-Dose MolecuIar Imaging

(e.g., Digital PET)

Standard 10 x Reduction

NCI R01CA195513

RSNA Medical Student

RSNA Molecular Imaging

Ohio Third Frontier


3


Explore “Deep Learning” for Pattern Recognition inImages, Digital Pathology, and Genomic Data:


4


Radiology Workflow

EMR System

AgeGenderReason

CPOE for Radiology Exam

Physician requests imagingstudy

Imaging Exam Requesting

Facility

REMIX

HIS/RIS

Exam Scheduled and Performed

PACSReconstructed images are searchable, and ready foradvanced Image analysis

All study relevant data and Images are linkable andmineable

Clinical Reasearchers

REMIX receives and de-identifiesimage data and related metadata

Data Warehouse

Patients have already beenConsented for TCCP


5




Facility

REMIX Desktop

PACS

Reconstructed images are searchable, and ready foradvanced Image analysis

All study relevant data, Images are linkable andmineable

Clinical Reasearchers / Researchers


Scanner

Enterprise Data Warehouse

Patients have already beenConsented based on study involved

Diagnostic Workstation

EMR\HIS\RISEnterprise Viewer

VNA

REMIX Recon REMIX BIREMIX AI

REMIX Server

Clinical\Operational User

Clinical and Operationaluses are uninterrupted

• REMIX (Radiology and Enterprise Medical Imaging Extensions)


6


• REMIX: De-Identification (Operated by OSUWMC Imaging Informatics)

Patient Search for Research Dataset Preparation

Clinical Trial Support

Honest Broker Compliant Batch Image Processing for Large datasets

Data verification supported by OSUWMC Imaging Informatics

CD Burning and Image sending to custom folders and\or destinations


7


REMIX Data-Mining Capabilities


8


REMIX Data-Mining Capabilities


9


REMIX: Quantitative Capabilities


10


Developed Texture-Analysis Capabilities


11


Texture Analysis Capabilities


12


12Quality Metrics: Radiology - Clinical Service Efficiency

Source: IHIS

Neuro MRI: Routine vs Stat

Stat

Role for Artificial Intelligence in Imaging


13




Facility

REMIX Desktop

PACS

Reconstructed images are searchable, and ready foradvanced Image analysis

All study relevant data, Images are linkable andmineable

Clinical Reasearchers / Researchers


Scanner

Enterprise Data Warehouse

Patients have already beenConsented based on study involved

Diagnostic Workstation

EMR\HIS\RISEnterprise Viewer

VNA

REMIX Recon REMIX BIREMIX AI

REMIX Server

Clinical\Operational User

Clinical and Operationaluses are uninterrupted

• REMIX


14


Positive

(Gold standard)

Negative

(Gold Standard)

Total

Positive

(AI algorithm)

38 5 43

Negative

(AI Algorithm)

2 35 37

Total 40 40 80

Neurologic Disease Medical Imaging Informatics

(e.g., Artificial Intelligence)

REMIX -AI


15


REMIX AI specs

Processor 1x intel Core i7-5930 K Processor (15M Cache,

3.50 GHz)

Memory 64GB DDR4

GPUs 4 x NVIDIA GeForce GTX Titan X GPUs (7

Teraflops of single precision, 336.5 GB/s of

memory bandwidth, 12 GB memory per GPU)

Operating System (OS) Ubuntu 14.04

Storage 2x 256 GB SSD disk for OS and software libraries

and 3x3TB standard disk on RAID 5 for data

storage

Connecting to REMIX AI:1) REMIX AI web interface, allowing users to upload their data into the

system2) REMIX Desktop, permitting users to directly save their image data

into shared disk drives of REMIX AI3) Python-based client libraries, so that users can make Restful API calls

to REMIX PACS


16


REMIX BI (Dataset Queries)

1,015

611

104 101 89

980

582

75 68 60

951

564

58 52 450

200

400

600

800

1,000

1,200

2008 2010 2014 2015 2016

Order to Finalize in Minutes Begin to Finalize in Minutes

Complete to Finalize in Minutes Expon. (Complete to Finalize in Minutes)

Query 1: Find all “non-contrast head CT Exams” where in the clinical system patients has been associated with “critical findings” (e.g., hemorrhage, mass effect and hydrocephalus) for a given month (under 12 seconds)

Query 2: Find all “non-contrast Head CT Exams” where in the clinical system patients have been associated with “Stroke” (under 8 seconds)


17


REMIX - Desktop


18


REMIX - Recon


19


Pitfalls


20


REMIX Desktop: Pre-Processing DICOM to JPEG

Installed on a desktop system: Intel Xeon E3-1270 v5 @ 3.60 GHz, 4 Cores CPU, 32 GB installed system memory, and an NVIDIA Quadro K1200 GPU with 4 GB graphics memory.

DICOM to color jpeg images using color lookup tables for each respective window and level.

Query 1, a “Brain Window” setting was utilized (W90/C40).

Query 2 a “Stroke Window’ setting was utilized (W30/C30).

The results were written into shared folders where they could be accessed by the REMIX AI module.

243 minutes.


21


Data Flow for AI


22


Results

True Positive example for Algorithm 1

Algorithm 1 Performance on Hemorrhage, Mass effect, or hydrocephalus: 90% Sensitivity (CI95%, 78-97%). 85% Specificity (CI95%, 76-92%). AUC = 0.91

Algorithm 2 Performance on Stroke: 62% Sensitivity (CI95%, 38-82%). 96% Specificity (CI95%, 82-100%), with AUC = 0.81

True Positive example for Algorithm 2

False Negative example for Algorithm 1

False Positive example for Algorithm 2


23


REMIX AI Performance

Images to 256x256 matrix and processed with GoogLeNet convolutional network running on Caffe. 60 training epochs used.

Model creation with the first dataset (from Query 1, with 2,583 images) was 6 minutes and 19 seconds

Model creation for the second dataset (from Query 2, with 646 images), total processing took 97 seconds

Once image-classification models were created, batch image classifications performed at approximately 25 images per second.


24


Questions?

Barbaros S. Erdal, Ph.D.

Department of RadiologyThe Ohio State University Wexner Medical Center

Contact: [email protected]

mailto:[email protected]

Using Deep Learning for Earlier Detection of Acute ... · Algorithm 2 Performance on Stroke: 62% Sensitivity (CI95%, 38-82%). 96% Specificity (CI95%, 82-100%), with AUC = 0.81 True

Documents