7/6/20

1

RADIATION ONCOLOGY

Radiomics and Radiogenomics Modeling with Machine LearningIssam El Naqa, PhD, DABR, FAAPM

Department of Machine LearningDivision of Quantitative Science

Moffitt Cancer Center July 16th, 2020

1

RADIATION ONCOLOGY Courtesy Randy Ten Haken, Andy Jackson , and Larry Marks

QUANTEC Symptomatic

pneumonitis vs. mean lung dose

DP/DMLD < 2 %/Gy

Why ‘omics outcomes modeling?

2

RADIATION ONCOLOGY

Population-based dose response

Courtesy Randy Ten Haken

Too shallow!

3

7/6/20

2

RADIATION ONCOLOGY

The Pan-Omics of Oncology

El Naqa et al, PMB, 2017

4

RADIATION ONCOLOGY

Outcome Modeling

Outcomes Modeling Schemes

Bottom-up

Clinical endpoint (phenotype)

Biophysical interaction(genotype)

Top-down

El Naqa, Wires, 2014

First principle mechanisms

Integrative models

5

RADIATION ONCOLOGY

= f }{ {

Damage metric

Integrative radiobiological modelingTCP/NTCP are multi-factorial and depend on: radiation dose and patients’ genomic (radiogenomics) and imaging (radiomics) characteristics before & during radiotherapy

El Naqa, Methods, 2016

6

7/6/20

3

RADIATION ONCOLOGY

Radiogenomics NTCP Modeling: Dose + genomics

Coates et al, RO, 2015

CNV

V10, V20, D50 and XRCC1 CNV

Rectal bleeding in prostate cancer

Lyman-Kutcher-Burman model

Logistic regression model

7

RADIATION ONCOLOGY

0 0.2 0.4 0.6 0.8 1Damage Fraction

0

0.2

0.4

0.6

0.8

1

NTC

P

f50=0.515 (0.459,0.571)

50=1.05 (0.0564,2.05)

f50=0.217 (-0.275,0.709)

50=0.651 (-0.998,2.3) =0.959 (-2,3.92)

with TGF

Dose only

MRI-DCE perfusion

Panomics NTCP Modeling: Dose+ biologics + imaging

El Naqa et al, IJROBP, 2018 (CME article)

ALBI changes in liver cancerParallel architecture model

8

RADIATION ONCOLOGY

Outcome modeling by Machine learning (ML)

–Generativemodels• Model class-conditional PDFs and prior probabilities

(Bayesian networks, Markov models)– To predict you need to know the system

–Discriminant models• Directly estimate posterior probabilities (logistic regression,

neural networks, CNN, random forests, SVM)– Predict without knowing the system

Tseng, Oncology, 2018

9

7/6/20

4

RADIATION ONCOLOGY

Deep vs conventional machine learning

10

RADIATION ONCOLOGY

Cui, Med Phys, 2020

Applications of ML/DL in Medical Physics and Radiation Oncology

11

RADIATION ONCOLOGY

Machine Learning Outcome Modeling

(medicalphysicsweb.org highlights, 2017)

Step

1:

Blan

ket S

elec

tion

Step

2:

Stru

ctur

e le

arni

ng

Dee

p Le

arni

ng

12

7/6/20

5

RADIATION ONCOLOGY

A MO-BN can be used to predict multiple radiation outcomes simultaneously, which provides opportunitiesof finding appropriate treatment plans to solve the trade-off between competing risks.

Luo et al, Med Phys, 2018 (Editor’s Choice)

Multi-Objective Generative Models

13

RADIATION ONCOLOGY

Human-in-the loop: Pre/During Treatment BNs for LC and RP2 Prediction

Luo, ASTRO 2019; under review

14

RADIATION ONCOLOGY

Composite neural network architecture

Cui et al, IEEE TRPMS, 2018 (Best of Physics, ASTRO)

Prediction of tumor local in lung cancer using DNN

15

7/6/20

6

RADIATION ONCOLOGY

Deep learning architectures for joint actuarial prediction of LC and RP2

Cui et al, AAPM 2020

16

RADIATION ONCOLOGY

How to build an ML/DL model?

Medical Physics Special Issue on Machine Learning

17

RADIATION ONCOLOGY

ML validation

• Depending on the level of evidence– Selection appropriate learning

algorithms– Validation and evaluation (TRIPOD

criteria)• Internally (cross-validation

schemes)• Externally (independent datasets)

– Provide interpretation of machine learning prediction

Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD)

18

7/6/20

7

RADIATION ONCOLOGY Luo, BJR-O, 2019

ML Accuracy versus interpretabilityFeature maps or Proxy models

Ens

embl

es

19

RADIATION ONCOLOGY

Interpretation of deep learning architectures for prediction of LC and RP2

Cui et al, AAPM 2020

20

RADIATION ONCOLOGY

Conclusions• Radiomics and radiogenomics offer new opportunities to develop better

TCP/NTCP models and for personalizing radiotherapy• Machine/deep learning techniques can improve feature selection and

statistical learning in radiomics/radiogenomics analytics and modeling radiotherapy outcomes

• Main challenges for radiomics/radiogenomics modeling– Harmonization and optimization of data integration methods– Uncertainties in data and model building schemes – Proper validation (TRIPOD criteria) and robustness for clinical decision

support– Better interpretation of radiomics/radiogenomics models is still lagging

21

7/6/20

8

RADIATION ONCOLOGY

THANK YOU!

May 2020 | Volume 47, Issue 5

Medical Physics is an official journal of the AAPM, the International Organization for Medical Physics (IOMP), and

the Canadian Organization of Medical Physicists (COMP).

Collage of ! gures from articles in the Special Issue on “The Role Machine Learning in Modern Medical Physics”, pp e125–e126 (online only, free available)

mp_v47_i5_cover.indd 1mp_v47_i5_cover.indd 1 5/15/2020 7:33:10 AM5/15/2020 7:33:10 AM

James Balter. PhDYue Cao, PhDPaul Carson, PhDKyle Cuneo, MDJoseph Deasy, PhDYuni Dewaraja, PhDIvo Dinov, PhDRobert G illies, PhDMike Green, MD, PhDJudy Jin, PhDShruti Jolly, MDTheodore Lawrence, MD PhD Marc Kessler, PhDDaniel L. McShan, PhD Martha M. Matuszak, PhD Chuck Mayo, PhDMichelle M ierzwa, MDJean Moran, PhDEduardo Moros, PhDMorand Piert, MDDipankar Ray, PhDAl Rehemtulla, PhDBen Rosen, PhDJan Seuntjens, PhDRandall K. Ten Haken, PhDXueding Wang, PhD

• Noora Ba Sunbul• James Coates, PhD• Sunan Cui, PhD• Jamalina Jamaluddin • Yi Luo, PhD • Dipesh Niraula, PhD• Abe Oraiqat, PhD• Julia Pakela• Wenbo Sun, PhD• Huan-Hsin Tseng, PhD• Lise Wei, PhD• Wei Zhang, PhD

22