purpose machine learning as new tool for predicting ... · machine learning as new tool for...

7
7/14/10 1 Machine Learning as New Tool for Predicting Radiotherapy Response Issam El Naqa, PhD Medical Physics Unit McGill University Machine Learning Symposium Purpose Motivation – Radiotherapy outcomes are determined by complex interactions between heterogeneous variables (treatment techniques, cancer pathology, and patient related physiological and biological factors) – Methods based on machine learning can Identify data patterns, variable interactions, and higher order relationships among prognostic variables Generalize to unseen data before • Objectives – Overview of current role of machine learning methods for predicting radiotherapy outcomes Normal tissue toxicities (NTCP) Tumor control probability (TCP) • Objective – Modeling (understanding) of post-RT effects Prediction generalizability • Types – Analytical (mechanistic models) such as (CV, LQ, LKB, EUD,…) (Moiseenko et al, ‘05) – Data-driven models Multivariate (multi-metric) analysis (Articles in IJROB by Levegrun et al ‘01, Marks ‘02, Tucker et al ‘04, Bradley et al ‘04, Blanco et al ’05, El Naqa ‘06) Learning methods (Munley et al PMB ’99, Su et al Med Phys ’05, Lennernas et al IJROB ’04, Gulliford et al Radiother Oncol ’04, Dawson et al IJROB ’05, etc) Proposed Methods for Radiobiological Modeling Radiobiological Modeling

Upload: others

Post on 18-Jul-2020

13 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Purpose Machine Learning as New Tool for Predicting ... · Machine Learning as New Tool for Predicting Radiotherapy Response Issam El Naqa, PhD Medical Physics Unit McGill University

7/14/10

1

Machine Learning as New Tool for Predicting Radiotherapy Response

Issam El Naqa, PhD

Medical Physics Unit McGill University

Machine Learning Symposium

Purpose •  Motivation

– Radiotherapy outcomes are determined by complex interactions between heterogeneous variables (treatment techniques, cancer pathology, and patient related physiological and biological factors)

– Methods based on machine learning can •  Identify data patterns, variable interactions, and higher order

relationships among prognostic variables •  Generalize to unseen data before

•  Objectives – Overview of current role of machine learning methods

for predicting radiotherapy outcomes •  Normal tissue toxicities (NTCP) •  Tumor control probability (TCP)

•  Objective – Modeling (understanding) of post-RT effects

⇒ Prediction generalizability •  Types

– Analytical (mechanistic models) such as (CV, LQ, LKB, EUD,…) (Moiseenko et al, ‘05)

– Data-driven models •  Multivariate (multi-metric) analysis (Articles in IJROB by

Levegrun et al ‘01, Marks ‘02, Tucker et al ‘04, Bradley et al ‘04, Blanco et al ’05, El Naqa ‘06)

•  Learning methods (Munley et al PMB ’99, Su et al Med Phys ’05, Lennernas et al IJROB ’04, Gulliford et al Radiother Oncol ’04, Dawson et al IJROB ’05, etc)

Proposed Methods for Radiobiological Modeling Radiobiological Modeling

Page 2: Purpose Machine Learning as New Tool for Predicting ... · Machine Learning as New Tool for Predicting Radiotherapy Response Issam El Naqa, PhD Medical Physics Unit McGill University

7/14/10

2

Machine learning for outcomes modeling concepts •  Data: refers to clinical outcomes or observations •  Learning: discovering patterns in the data to make predictions about future

observations •  Assumptions on the Data

–  Same source –  Redundancy-> compressibility (entropy) –  Predictability (hidden relations in the data)

•  Pattern Types •  Exact relations , e.g. E=mc2 •  Approximate relations •  Probabilistic relations

•  Properties of a good model –  Computational efficiency –  Robustness (handling of ‘noisy’ data) –  Statistical stability (insensitive to particular datasets)

•  Learning tasks –  Supervised: data with known labels (classification, regression) –  Unsupervised: data without labels (clustering, novelty detection, dimensionality

reduction

Statistical Learning (VC)

Variance Bias

Pre

dict

ion

Err

or

Model Complexity

Training curves

Testing curve

Unsupervised learning: Principle Component Analysis

xxxxx xxx

x xx xxx xx x

x

PCA1 PC

A2

Normal tissue toxicity in lung/HNC

Page 3: Purpose Machine Learning as New Tool for Predicting ... · Machine Learning as New Tool for Predicting Radiotherapy Response Issam El Naqa, PhD Medical Physics Unit McGill University

7/14/10

3

Analysis of Parotid/Liver toxicity

Dawson et al., IJROBP, 2005

PC1 vs. PC2 and PC1 vs. PC3 for 90 single parotid not shown). cumulative dose volume histograms. Bubble size represents saliva flow at 12 months after irradiation. Light bubbles represent complications ( 25% saliva flow).

PC1 vs. PC2 and PC1 vs. PC3 for 90 single parotid not shown). cumulative dose volume histograms. Bubble size represents saliva flow at 12 months after irradiation. Light bubbles represent complications ( 25% saliva flow).

Supervised Learning: Neural Networks Synapses of Network

f(x)

Inpu

t Dat

a

FFNN

GRNN

RBF

Pros: Flexible

Cons: Use heuristics

Hagan et al ‘96 Neural Network Design

Example: Radiation Pneumonitis (RP)

NNall NNdose

Chen et al., Med Phys, 2007

BPNN was successfully used to predict Tdelay from tumor ADC values obtained from

HT29 xenografts undergoing fractionated chemoradiation therapy

Kakar et al., IJROBP, 2009

Page 4: Purpose Machine Learning as New Tool for Predicting ... · Machine Learning as New Tool for Predicting Radiotherapy Response Issam El Naqa, PhD Medical Physics Unit McGill University

7/14/10

4

Supervised Learning II: Kernel-based methods (Support Vector Machines)

(Vapnik ’98, Nature of statistical learning)

SVM RP

SVMall SVMdose

Chen et al., Med Phys, 2007

Software tools for modeling: (DREES)

Bradley et al., IJROBP ‘07

Multivariate approach: Lung radiation pneumonitis

Page 5: Purpose Machine Learning as New Tool for Predicting ... · Machine Learning as New Tool for Predicting Radiotherapy Response Issam El Naqa, PhD Medical Physics Unit McGill University

7/14/10

5

El Naqa et al., PMB, 2009

Results I: NTCP (RP)

Linear/nonlinear modeling comparison

Dose-volume modeling

Mu et al, ASTRO, 2008

Nonlinear TCP modeling

El Naqa et al., AO, 2010

Page 6: Purpose Machine Learning as New Tool for Predicting ... · Machine Learning as New Tool for Predicting Radiotherapy Response Issam El Naqa, PhD Medical Physics Unit McGill University

7/14/10

6

Results II: TCP visulaization

Radial basis kernel Linear logistic regression

Zhang et al., IJROBP, 2009

Generative Models

X1

X4 X3

X2

X5 X6

Y

Bayesian Networks

Roweis et al, NC ‘99

Example I: Interaction analysis

Jung Hun Oh, ICMLA ’09

Page 7: Purpose Machine Learning as New Tool for Predicting ... · Machine Learning as New Tool for Predicting Radiotherapy Response Issam El Naqa, PhD Medical Physics Unit McGill University

7/14/10

7

BN vs SVM

Jayasurya et al., Med Phys, 2010

Combined RP

Das et al., Med Phys, 2007

Conclusions •  Approaches based on machine learning are useful

for modeling complex radiotherapy treatment outcomes

•  Different methods based on supervised and unsupervised learning methods have been used to model NTCP and TCP

•  Unsupervised methods can provide guidance about the complexity of the data and the type of modeling required

•  Better understanding of what machine learning based methods mean is still lagging