Transforming images into meaningful
data through AI and imaging biomarkersAngel Alberich-Bayarri, PhD
[email protected] Biomedical Imaging Research Group (GIBI230)
La Fe Polytechnics and University Hospital2 QUIBIM SL
Content
1. Introduction
2. Steps to meaningfulness1. Acquisition
2. Processing
3. Analysis
4. Interpretation - Radiomics
3. Implementation
4. Topics for discussion
Introduction
• Current information systems do not allow the appropriate
management of image annotations, segmentation masks
and quantitative data, among others.
• There still exists a high variability between imaging
vendors, both in the image quality characteristics and in
the measurements generated in different workstations.
• AI is usually proposed as the solution for everything but it
is really efficient when designed to solve highly specific
problems
Introduction
• Challenges of AI and Quantitative imaging solutions in
Medicine
Objective and reproducibleEfficacy in the evaluation of the disease and treatment
Cost-efficientProviding surrogate results or
close to clinical endpoints
Image analysis methods
Meaningfulness
“The quality of having great value or
significance”
Steps to meaningfulness
1Martí Bonmatí L, Alberich-Bayarri A, et al. [Imaging biomarkers, quantitative imaging, and
bioengineering]. Radiologia. 2012 May-Jun;54(3):269-78.
• Stepwise strategy for the extraction of quantitative
information from medical images
Steps to meaningfulness
European Society of Radiology. ESR statement on the stepwise
development of Imaging Biomarkers. Insights Imaging. 2013 Apr;4(2):147-52.
Steps to meaningfulness
IMAGE ACQUISITION OPTIMIZATION
IMAGE PROCESSINGSEGMENTATION
REGISTRATION
IMAGE ANALYSIS
INTERPRETATION, RADIOMICS
Steps to meaningfulness
Our approach:
AI is not the focus, but the tool to
optimize our workflows when
needed
METROLOGY
• Acquisition
Steps to meaningfulness
Steps to meaningfulness
Calibration and phantoms are a must
• Acquisition
CT - HU MR – T1 & T2 CT and MR, porous
structures
Steps to meaningfulness
• Processing
Beyond image classification and
semantics, segmentation is the
main part of the quantitative
imaging workflows that will benefit
from AI
• Computing model inspired by
biological neural networks.
• Key elements:
o Input Layer
o Hidden Layers
o Output Layer
o Connections (Weights)
• Modify the intensity of its
connections (weights) during
training with labeled data
Steps to meaningfulness
Neural Networks
• Special type of neural networks which present an outstanding
performance in Computer Vision problems.
• Key elements:
o Convolutional Layers
o Pooling Layers
o Fully Connected Layers
• In addition of fully connected weights, convolutional filters are learnt
during training, which are able to recognize patterns in images.
Steps to meaningfulness
Convolutional Neural Networks (CNN)
True positive
False negative
False positive
DICE COEFFICIENT:
Training: 91,25%
Validation: 82,47%
Test: 81,83%
Steps to meaningfulness
• Prostate segmentation
50 cases
True positive
False positive
False negative
DICE COEFFICIENT:
Training: 96.59%
Validation: 96.60%
Test: 87.37%
Steps to meaningfulness
• Liver segmentation
35 cases
• Analysis: Prostate cancer
Steps to meaningfulness
Source images
Size
(~3)
Shape
(~10)
Intensity
(~5)
Textures
(~30)
Function
(~200-500)Massive analysis of Imaging
Biomarkers
Imaging Data Mining
• Analysis: Prostate cancer
– Combine infromation from
cellularity (MR diffusion) and
tumoral angiogenesis (MR
perfusion) to determine regions
with higher tumoral
aggresiveness
– Registration:• Afine + B-Splines (elastic)
• Mutual information
T2 Diffusion(register)
Perfusion(register)
Deformation
field
Steps to meaningfulness
• Analysis: Prostate cancer– Diffusion and perfusion analysis
– Voxelwise data clustering
– Hierarchy aglomerative clustering
– Z-score
Ktr
ans
[min
-1]
ADC [mm2/s]
Ktr
ans
[z-z
sco
re]
ADC [z-score]
Steps to meaningfulness
• Reports
Steps to meaningfulness
• Analysis: Prostate cancerG
lea
so
n 6
Gle
as
on
7G
lea
so
n 9
Steps to meaningfulness
• Data reduction
Baseline Follow-up
• Experimental treatment (40 patients)
Publication pending
Steps to meaningfulness
• Data reduction
• Experimental treatment (40 patients)
• Prognostic value:– Relapse
– D2D at baseline (p=0.05)
Publication pending
0 1
1.0
1.1
1.2
1.3
1.4
1.5
2D Fractal Dimension vs. Relapse of Rectal Cancer
Relapse
2D
Fra
cta
l D
imensio
n
Steps to meaningfulness
Implementation
• Storage, imaging biomarkers analysis and radiomics.
Local and Cloud solutions.
• 7500 analysis/year and growing
• More than 250 registered users (radiologists,
engineers, biomedical imaging researchers)
Implementation
Neuro
MSK
Oncology
Abdomen
Lung
Brain volumetry with parcellationWhite matter lesions detection
State of art
Brain volumetry with parcellation – detailed regions
Longitudinal white matter lesions
Advanced
Diffusion ADC, IVIMPerfusion – Semiquantitative analysis
Texture analysis
Diffusion Kurtosis, StretchedPerfusion – Pharmacokinetics modeling
2D Bone microarchitecture, BMD from CT
3D Bone microarchitecture3D Bone microarchitecture + FE analysis
Cartilage morphologyCartilage T1 mapping
Liver T1 and T2 mapping Liver simultaneous fat & iron quantification
Lung Emphysema Lung low functional densities
Implementation
Implementation
...
Study folder
Series folder
Series folder
info.json
header.json
DICOM files
NIFTI files
biomarker folder
analysis.json
PRECISION CLI APP
Code folder
*.bat
Code files
*.jade
*.bat call with analysis.json path as input
INPUT
CODE
ANALYSIS OUTPUT
Study folder
biomarker folder
Precision CLI call to retrieve results
Report folder
Results folder
Results retrieval
Images folder
report.xml
Other images
result.xml
report.jade
Series folder
Input/code filesOutput files
analysis.json
Introduce your plugin!
Implementation
AI classification applications
• Chest X-Ray screening
• Deep Learning
• 7470 labeled images
• Normal vs. Abnormal
• AUC=0.91
B-0402 B. Fos Guarinos et al. SS604, ECR 2017
Research agreement
with NVIDIA
Acknowledgements
POST-DOCAlejandro Torreño, PhD - Technology DevelopmentAlejandro Rodríguez, PhD - Image Analysis Engineer
PhD STUDENTSAmadeo Ten - Image Analysis EngineerSara Carratalá - CNS Analysis
CLINICAL TRIALS AND PREBISandra Pérez - Data ManagerJuan Ramón Terrén - Data ManagerRebeca Maldonado - Technician & PREBI
ADMINISTRATIONAna Penadés - Economic & Financial Manager
GIBI230
QUIBIMIMAGE ANALYSIS SCIENTISTSFabio García Castro - Chief Image Analysis ScientistBelén Fos Guarinos - Image Analysis ScientistAna María Jiménez Pastor - Image Analysis ScientistRafael López González - Image Analysis Scientist
DEVELOPMENTRafael Hernández Navarro - Chief Technology OfficerAlejandro Mañas García - Full Stack Senior DeveloperEduardo Camacho Ramos - Front-End Developer
CLINICAL TRIALSIrene Mayorga Ruíz - Clinical Trials CoordinatorRaúl Yébana Huertas - Image Analysis Technician
MARKETING AND COMMUNICATION Katherine Wilisch Ramírez - Marketing Manager
MANAGEMENTIsabel Montero Valle – Team CoordinatorEncarna Sánchez Bernabé - Chief Operating OfficerDaniel Iordanov López - Assistant to Business Development
Luis Martí Bonmatí MD, PhD. GIBI Principal Investigator
QUIBIM Founder
Ángel Alberich Bayarri, PhD. GIBI Scientific-Technical DirectorQUIBIM CEO & Founder
Acknowledgements
Topics for discussion
• The ‘training datasets’ issue: How can we stimulate and
encourage the participation of experts (radiologists,
residents) in segmentation and classification tasks? • Publications, meetings, fee?
• How can we break the data silos?• Experiences in building new biobanks (Valencia – LaFe - Euro-
Bioimaging, Parenchima COST, Italian biobank)
• How can we integrate these solutions in hospitals?• Buy all start-up solutions?
• Through big players? Siemens, Philips, GE, TeraRecon(EnvoyAI)
Transforming images into meaningful
data through AI and imaging biomarkersAngel Alberich-Bayarri, PhD
[email protected] Biomedical Imaging Research Group (GIBI230)
La Fe Polytechnics and University Hospital2 QUIBIM SL