17
Center for Biomedical Image Computing and Analytics https://www.med.upenn.edu/cbica/ The Cancer Imaging Phenomics Toolkit (CaPTk) Christos Davatzikos and Despina Kontos (on behalf of the team)

The Cancer Imaging PhenomicsToolkit ( CaPTk) › groups › itcr › File:Davatzikos_Kontos...Deep Learning in CaPTk Any user can: •download DeepMedicand train any model. •load

  • Upload
    others

  • View
    2

  • Download
    0

Embed Size (px)

Citation preview

Page 1: The Cancer Imaging PhenomicsToolkit ( CaPTk) › groups › itcr › File:Davatzikos_Kontos...Deep Learning in CaPTk Any user can: •download DeepMedicand train any model. •load

Center for Biomedical Image Computing and Analytics

https://www.med.upenn.edu/cbica/

The Cancer Imaging Phenomics Toolkit (CaPTk)

Christos Davatzikos and Despina Kontos (on behalf of the team)

Page 2: The Cancer Imaging PhenomicsToolkit ( CaPTk) › groups › itcr › File:Davatzikos_Kontos...Deep Learning in CaPTk Any user can: •download DeepMedicand train any model. •load

[email protected] #ITCR19

Cancer Imaging Phenomics Toolkit (CaPTk)

O p e n - S o u r c e

E a s y t o u s e

E x t e n d a b l e

W e b - a c c e s s i b l e

ipp.cbica.upenn.edu(HPC-shared resources for

computationally demanding pipelines)

www.cbica.upenn.edu/captkgithub.com/cbica @CBICAannounce @CBICAannounce

Page 3: The Cancer Imaging PhenomicsToolkit ( CaPTk) › groups › itcr › File:Davatzikos_Kontos...Deep Learning in CaPTk Any user can: •download DeepMedicand train any model. •load

Feature Synthesis and Integrationvia Machine Learning

Second Level

Segmentation:

Regions of Interest (ROIs)

Registration:- Measure change with time:

- Population atlases:

A B BàA

Mutationvs.Wildetype

Image Operations:

• DICOM access• Format conversion• Intensity normalization• Co-registration• Noise Reduction• ROI annotation• Seed-point initialization

Feature Extraction:Texture, histogram,

dynamics, spatial patternWavelet-DP1

Image Harmonization:

Image Analysis Algorithms

CaPTk Radiomic Panel

Open-CV

Output Modules and OutcomesPersonalized Treatment:

Peri-tumoral Infiltration

Predictive models:

Before After

RadiogenomicsImaging signatures of molecular characteristics:

Breast MRI Phenotypes vs. Oncotype DX

Imaging Signatures of GBM mutations

First Level

Precision Diagnosis, Risk Estimation:

Kaplan-Meier EstimatorBreast Density FactorvIII+vIII-

EGFRvIII+ EGFRvIII-

ITK

Two-Level Architecture

Page 4: The Cancer Imaging PhenomicsToolkit ( CaPTk) › groups › itcr › File:Davatzikos_Kontos...Deep Learning in CaPTk Any user can: •download DeepMedicand train any model. •load

What is New Since Last May?

Page 5: The Cancer Imaging PhenomicsToolkit ( CaPTk) › groups › itcr › File:Davatzikos_Kontos...Deep Learning in CaPTk Any user can: •download DeepMedicand train any model. •load

[email protected] #ITCR19

Feature Extraction

• Participation and compliance with the International Imaging Biomarker Standardization Initiative (IBSI)

• Validation based on• Digital Phantom• Strict Preprocessing Steps• Comparing with 21 other software packages

Page 6: The Cancer Imaging PhenomicsToolkit ( CaPTk) › groups › itcr › File:Davatzikos_Kontos...Deep Learning in CaPTk Any user can: •download DeepMedicand train any model. •load

[email protected] #ITCR19

Deep Learning in CaPTk

Any user can:• download DeepMedic and train any model.• load on CaPTk for direct use• share back to the CaPTk community via GitHub• Active collaboration with Imperial College London (Ben Glocker and Kostas Kamnitsas)

Allowing inclusion of custom DL models

Page 7: The Cancer Imaging PhenomicsToolkit ( CaPTk) › groups › itcr › File:Davatzikos_Kontos...Deep Learning in CaPTk Any user can: •download DeepMedicand train any model. •load

[email protected] #ITCR19

Interactive ML-assisted segmentation

Page 8: The Cancer Imaging PhenomicsToolkit ( CaPTk) › groups › itcr › File:Davatzikos_Kontos...Deep Learning in CaPTk Any user can: •download DeepMedicand train any model. •load

[email protected] #ITCR19

Our contribution to the CWL community

• A fully compatible implementation of the CommonWorkflow standard

• A mechanism to construct easy-to-use command line APIs for C++ developers (github.com/cbica/cmdparser)

parser.addOptionalParameter("i", "inImage", cbica::Parameter::STRING, "File", "Input Image");

• Automatically creating the CWL interface file, from the same API

• Enabling easier construction of pipelines with different applications

• Maximizing pipeline utilization to supported cloud providers

https://www.commonwl.org/

Page 9: The Cancer Imaging PhenomicsToolkit ( CaPTk) › groups › itcr › File:Davatzikos_Kontos...Deep Learning in CaPTk Any user can: •download DeepMedicand train any model. •load

[email protected] #ITCR19

Analyzing the IvyGAP data (Spyros Bakas et.al.)• Manually-revised Tumor Segmentation Labels• Comprehensive Panel of Extracted Features• Spatial Distribution Atlases

EGFR

vIII-

EGFR

vIII+

UPenn subjects (N=129)

EGFR

vIII-

EGFR

vIII+

IvyGAP subjects (N=33)

Page 10: The Cancer Imaging PhenomicsToolkit ( CaPTk) › groups › itcr › File:Davatzikos_Kontos...Deep Learning in CaPTk Any user can: •download DeepMedicand train any model. •load

[email protected] #ITCR19

Clinical Applications: Brain GBM Progression• Clinical Trial using CaPTk to guide elevated radiation in peri-tumoral

tissue; Prospective NRG clinical trial on extensive resection

• Assessing Pseudo-progression vs. true progression in GBM

Penn Case WesternCa

PTk s

core

Page 11: The Cancer Imaging PhenomicsToolkit ( CaPTk) › groups › itcr › File:Davatzikos_Kontos...Deep Learning in CaPTk Any user can: •download DeepMedicand train any model. •load

[email protected] #ITCR19

Parenchymal Texture Analysis

Spatial Lattice

Clinical Applications: Breast Cancer Risk Assessment

Entropy Inhomogeneity

>45,000 screening digital mammograms

UnadjustedOR (95% CI) P-value

Adjusted*

OR (95% CI) P-valuePhenotype

Green 3.49 (1.76 – 6.91)

0.0002

2.54 (1.09 – 6.05)

0.001Dark Green 1.53 (0.70 – 3.33) 2.33 (0.82 – 6.75)Dark Red Reference ReferenceRed 0.40 (0.13 – 1.24) 0.06 (0.00 – 0.51)

Percent density (per std. dev.)

1.35 (1.04 – 1.78) 0.027 2.69 (1.73 – 4.43) <0.0001

*Complete case-analysis, adjusted for age and BMI

Kontos et al. Radiology 2018

Page 12: The Cancer Imaging PhenomicsToolkit ( CaPTk) › groups › itcr › File:Davatzikos_Kontos...Deep Learning in CaPTk Any user can: •download DeepMedicand train any model. •load

[email protected] #ITCR19

Radiomic phenotypes of breast cancer tumor heterogeneity

Chitalia et al. RSNA 2018, Chitalia et al. Clinical Cancer Research (under review)

Recurrence free survival probabilities for patient groups assigned to heterogeneity phenotypes

Cox proportional hazards survival analysis

Clinical Applications: Breast Cancer Prognosis

Page 13: The Cancer Imaging PhenomicsToolkit ( CaPTk) › groups › itcr › File:Davatzikos_Kontos...Deep Learning in CaPTk Any user can: •download DeepMedicand train any model. •load

[email protected] #ITCR19

Clinical Applications: Lung Cancer RadiotherapySBRT Response Prediction

18F-FDG-PET dataset of 100 consecutive patients treated with SBRT

H Li, et al. Radiotherapy and Oncology 129 (2), 218-226 2018

Page 14: The Cancer Imaging PhenomicsToolkit ( CaPTk) › groups › itcr › File:Davatzikos_Kontos...Deep Learning in CaPTk Any user can: •download DeepMedicand train any model. •load

[email protected] #ITCR19

Collaborations• Penn-UCSD: work on EGFR A289 published in Cancer Cell relating CapTk imaging features

to tissue analysis and mouse experiments

• Intel AI: work on federated learning for segmentation

• Imperial College London: work on deep learning-based segmentation

• Inter-institutional collaboration providing 3,000 Brain MRIs for training CaPTk ML modules

• Collaboration with EORTC and adoption of CaPTk into Pipelines for the GLASS consortium

• Penn/Mayo Clinic/UCSF: Consortium with >45,000 digital screening mammograms to be analyzed to establish novel breast cancer risk imaging phenotypes

• Penn/UNC: collaboration >50,000 digital screening mammograms to establish relationship between breast cancer risk imaging phenotypes and tissue-based markers

• Lung screening NIH/PROSPR 2: >2,000 LDCT scans of COPD for lung cancer risk estimation

Page 15: The Cancer Imaging PhenomicsToolkit ( CaPTk) › groups › itcr › File:Davatzikos_Kontos...Deep Learning in CaPTk Any user can: •download DeepMedicand train any model. •load

[email protected] #ITCR19

Future directions (Cohort-based analysis)

Page 16: The Cancer Imaging PhenomicsToolkit ( CaPTk) › groups › itcr › File:Davatzikos_Kontos...Deep Learning in CaPTk Any user can: •download DeepMedicand train any model. •load

Integrate imaging, EHR, histopathology, and molecular markers in multi-parametric prognostic and predictive models

+

Precision Screening, Prognosis, and Treatment of Cancer

Future Directions: Integrated Diagnostics

+ +

Page 17: The Cancer Imaging PhenomicsToolkit ( CaPTk) › groups › itcr › File:Davatzikos_Kontos...Deep Learning in CaPTk Any user can: •download DeepMedicand train any model. •load

[email protected] #ITCR19

Thank you!

http://www.cbica.upenn.edu