radiomics: novel paradigm of deep learning for clinical decision support toward plan b using liquid...

12
Technology Trend Analysis through the RI-Biomics Technology Professional Workforce Development Program and Information System Development Vol.04 November 2014 KARA ISSUE PAPER Advanced Research Center for Nuclear Excellence Issue Paper RI-Biomics Based Bio-sensing Application Technology Development

Upload: wookjin-choi

Post on 15-Apr-2017

2.311 views

Category:

Healthcare


0 download

TRANSCRIPT

Page 1: Radiomics: Novel Paradigm of Deep Learning for Clinical Decision Support toward Plan B using Liquid Biopsy

Technology Trend Analysis throughthe RI-Biomics Technology Professional Workforce DevelopmentProgram and Information System Development

Vol.04

November 2014KARA ISSUE PAPERAdvanced Research Center for Nuclear Excellence Issue PaperRI-Biomics Based Bio-sensing Application Technology Development

Page 2: Radiomics: Novel Paradigm of Deep Learning for Clinical Decision Support toward Plan B using Liquid Biopsy

Working Definition

The term, “Radiomics” can be thought of as analogous to “genomics” whereby medical images are broken down into a large number of complex component parts and tagged and made available in a database. This is similar to the process in which patient DNA information is acquired and made available for analysis in a genomic database. The term “radiome” is similarly analogous to the genome. Both the image “radiome” and the patient or tumor genome can be analyzed and combined resulting in paired gene expression data and medical images including diagnostic information from a study cohort. These combined data can be then correlated with a public database of genomic data and image data and this can be utilized to improve clinical outcomes using a closely matched population. This will ultimately result in the ultimate goal of personalized/precision based medical practice.

Methodology

Radiomics (tagged medical images) can be utilized along with genomic data to integrate big data from both imaging informatics and genomic data with 3-dimensional 3D or 4D Computer Aided Detection/Diagnosis(CADe/CADx) for Clinical Decision Support(CDS) through a publically available free, on-line software platform such as ePAD(Stanford university). A super computer(IBM) such as the one currently being introduced at Memorial Sloan-Kettering Cancer Center can read medical records and recommend treatment, particularly for cancer patients as a clinician assistant. The field of Cognitive computing is emerging to overcome such limitations of the current generation of computers, resulting in a new generation of computer applications and technologies that are much better able to “understand” medical data and its relationships.

Perspectives

A technique, referred to as a “liquid biopsy,” (Johns Hopkins) has been applied to find telltale genetic material from a vial of blood for diagnostic purposes. This technique and others have allowed us to delve into the “cockpit of cancer” showing how a series of genetic mutations, adding up silently over decades, turn cells cancerous. Radiomics particularly combined with “liquid biopsy” genomics from these types of blood tests will provide much earlier diagnosis and will allow a new era of mass screening for various diseases through preclinical and clinical stages.

Radiomics : Novel Paradigm of Deep Learningfor Clinical Decision Support02

Advanced Research Center for Nuclear TechnologyKARA Issue Paper Vol.04 Radiomics Novel Paradigm of Deep Learning

for Clinical Decision Support

03 | Introduction04 | Background & Need06 | Overview07 | Cases08 | Perspectives10 | Expectation11 | References

CONTENTSA

bstract

Jongho Kim, MD, PhDDivision of Nuclear Medicine Department of Diagnostic Radiology and Nuclear Medicine

Wook Jin Choi, PhDDepartment of Radiation Oncology

University of Maryland School of Medicine. University of Maryland Medical Center.22 S Greene St, Baltimore, MD, 21201, USA

Page 3: Radiomics: Novel Paradigm of Deep Learning for Clinical Decision Support toward Plan B using Liquid Biopsy

Novel radiomics based approaches can be used to identify new imaging biomarkers when long-term clinical outcomes such as mortality are not known, relying on first acquiring paired gene expression data and medical images at diagnosis from a study cohort, then leveraging the public gene expression data that contain clinical outcomes from a closely matched population. A critical step in this approach involves predicting the image features in the study cohort in terms of gene signatures. The prognostic significance of these gene signatures in public data sets for which gene expression and survival data are available revolutionize and accelerate the application of personalized medicine(Figure 1).

Korean Association for Radiation Application

Introduction

03Advanced Research Center for Nuclear TechnologyKARA Issue Paper Vol.04

Public radiomic data into personalized medicine

Imaging Informatics

Figure 1. Radiomics as public-to-person data communication

Radiomics refers to the comprehensive quantification of disease phenotypes by applying a large number of quantitative image features from medical images such as positron emission tomography(PET), computed tomography(CT), and magnetic resonance imaging(MRI), etc. In the genomic and epigenomic era, personalized medicine aims to tailor medical care to the individual by characterizing tissue heterogeneity through advanced high-throughput molecular technologies. Within this context, the fusion of the radiome and genome would be an essential part of clinical science and practice in a near future.

Page 4: Radiomics: Novel Paradigm of Deep Learning for Clinical Decision Support toward Plan B using Liquid Biopsy

Radiomics : Novel Paradigm of Deep Learningfor Clinical Decision Support04

Korean Association for Radiation Application

Background & Need

By assessing tissue characteristics noninvasively, medical imaging is likely to be a major factor in the advancement of clinical practice and science for human health taking advantage of seamless integration with genomic data particularly from then public domain.

Radiomics data can be extracted from multi-modal imaging data and combined with genomics by leveraging existing already funded public gene expression data containing clinical outcomes such as mortality from a closely matched population, representing 4th generation imaging informatics following anatomic, functional and molecular imaging.

With the advance in computer technologies to deal with exponentially rising data even from a large volume of unstructured data(Giga 109->Tera, 1012->Peta, 1015->Exa, EB, 1018 byte) through high performance processors,

- IBM’s Watson computer vanquished human contests on the TV quiz show named Jeopardy! with its performance at 80 TeraFLOPs and 500 gigabytes, the equivalent of a million books per second.

- To overcome the weakness of the standard way of building computers, von Neumann computers, such as limited comprehension, cognitive computing is emerging. IBM recently unveiled a new processor that uses only a fraction of the energy that today’s processors do, but that can deliver radically greater returns on a brain-like synaptic scale.

- Cognitive computing can be compared to the standard von Neumann computer(Figure 2) as tracked in the red line consisting of three components, a CPU, hard drive and RAM(fast accessible memory). Today the von Neumann plot(solid red) is well ahead on the development curve when compared to the futuristic cognitive(neuromorphic) computer(solid green) but where the computing power of traditional computers and neuromorphic computer intersect will be usher in the dawn of a new computing paradigm.

4th generation imaging informatics

Computer technology

Cognitive Computer

IBM developed a system that can read medical records to guide management, particularly for cancer patients. This has been launched at hospitals in the U.S. including Memorial Sloan-Kettering Cancer Center, New York, functioning as a clinician assistant. Recently, cognitive computing has emerged to overcome such weakness of the standard way of building computers, von Neumann computers, as limitations including comprehension. This has occurred in parallel with new IBM processing capabilities that can deliver radically greater returns using biological mimicking approaches such as a brain-like synaptic scale.

Page 5: Radiomics: Novel Paradigm of Deep Learning for Clinical Decision Support toward Plan B using Liquid Biopsy

05Advanced Research Center for Nuclear TechnologyKARA Issue Paper Vol.04

Figure 2. Neuromorphic vs. standard von Neumann computer

Page 6: Radiomics: Novel Paradigm of Deep Learning for Clinical Decision Support toward Plan B using Liquid Biopsy

Radiomics : Novel Paradigm of Deep Learningfor Clinical Decision Support06

Plan A vs. B

Due to the spatial and temporal heterogeneity of diseases such as tumors, invasive techniques such as biopsies to extract small portions of tissue do not allow for a complete characterization as a whole, therefore a comprehensive view of medical imaging has great potential for monitoring disease progression including response to therapy if associated with genomics or proteomics already available in the public domain. One of the greatest challenges is to incorporate the personalized/precision approach into routine clinical practice.

ePAD, a free software for quantitative imaging informatics, was developed for the National Cancer Institute by researchers at Stanford in the radiology department. It can serve as a model-platform for radiomics for a wide range of imaging-based projects combined with on-line genetic data, by making the content of images(imaging informatics) machine accessible and intelligible to allow electronic correlation with other biomics such as genomics as well as clinical informatics. In parallel, development and translation of basic biomedical informatics will improve radiology practice and decision making tools to efficiently and thoroughly capture the semantic terms radiologists use to describe lesions.

The idea of treating the advanced rather than early diseases is not ideal as a primary approach, but instead it is preferable to reduce cancer deaths using screening to catch cancer early. All these preventive steps represent “Plan B” but unfortunately much less attention and funding has been devoted to early detection of cancer but when prevention works, it has better results than any drug. In the U.S., the chance of dying from colorectal cancer is 40 percent lower than it was in 1975; a decrease mostly due to colonoscopy screening as well as melanoma, and skin cancer, treatable with surgery if caught early. Therefore, the Plan B needs to be Plan A and screening should be the primary approach for treating diseases. This is already the case in screening patients for breast cancer, The National Lung Screening Trial(NLST) is an excellent example of this approach for lung cancer. Radiomics, and imaging informatics combining with genomics from blood test, “liquid biopsy” will propel personalized medicine for prevention, early detection by screening and treatment response of various disease along with preclinical stage of drug development using genetically modified animal models.

Radiomics regarding Heterogeneity

Stanford ePAD Platform for Radiomics

Korean Association for Radiation Application

Overview

Page 7: Radiomics: Novel Paradigm of Deep Learning for Clinical Decision Support toward Plan B using Liquid Biopsy

07Advanced Research Center for Nuclear TechnologyKARA Issue Paper Vol.04

Harvard Radiomics at Dana-Farber CancerInstitute

NationalLung Screening Trial (NLST)

To identify prognostic imaging biomarkers in non-small cell lung cancer(NSCLC) using PET/CT by means of a radiomics strategy, gene expression and medical images were integrated in patients for whom survival outcomes are not available by leveraging survival data in public gene expression data sets. A radiomics strategy for associating image features with clusters of co-expressed genes(metagenes) was defined. First, a radiomics correlation map is created for a pairwise association between image features and metagenes. Next, predictive models of metagenes are built in terms of image features by using sparse linear regression. Similarly, predictive models of image features are built in terms of metagenes. Finally, the prognostic significance of the predicted image features are evaluated in a public gene expression data set with survival outcomes. The predicted image features were mapped to a public gene expression data set with survival outcomes, tumor size, edge shape, and sharpness ranked for prognostic significance.

Human cancers exhibit strong phenotypic differences that can be visualized noninvasively by medical imaging. Radiomics refers to the comprehensive quantification of tumor phenotypes by applying a large number of quantitative image features, which reveal prognostic radiomic signature as well as capture tumor heterogeneity, associated with underlying gene-expression patterns.

The National Lung Screening Trial(NLST) was a randomized trial of lung-cancer-specific mortality among participants in an asymptomatic high-risk cohort who underwent screening with the use of low-dose helical computed tomography(CT) as compared with screening with the use of single-view poster anterior chest radiography. It is possible to perform research analysis of pulmonary nodules(0.3 mm - 3cm) for lung cancer diagnosis, particularly radiomics research and identify the association of a nodule’s radiome(personalized CAD development) and genome(gene signature for early diagnosis), malignancy of nodule, and subsequently predict prognosis and therapy response using NLST metadata. Moreover, the use of NLST data can contribute to personalized medicine for screening in terms of tumor heterogeneity of adenocarcinoma vs. squamous cell cancer with further characterization using PET/CT along with developing genomic biomarkers.

Stanford Radiomics using PET/CT

Korean Association for Radiation Application

Cases

Page 8: Radiomics: Novel Paradigm of Deep Learning for Clinical Decision Support toward Plan B using Liquid Biopsy

Radiomics : Novel Paradigm of Deep Learningfor Clinical Decision Support08

Korean Association for Radiation Application

Perspectives

In order to break the current computing paradigms, neuromorphic chips have the capacity to “learn”. The object recognition software used at Baidu and Google, for example “Google Cat”, is trained on the ImageNet database, boasting thousands of object categories and it’s definitely an achievement to make a state-of-art chip of the scale but it still falls far short of achieving the capabilities of the human brain.

- With support from DARPA(Defense Advanced Research Projects Agency, U.S), IBM’s chip research project, the SyNAPSE(Systems of Neuromorphic Adaptive Plastic Scalable Electronics) aims to create a human brain-like hardware/software system that breaks the von Neumann paradigm. Hence neuromorphic chips such as TrueNorth(IBM) pack the memory, computation and communication components into little modules to proced locally but communicate with each other easily and quickly using a revolutionary new technology inspired by the human brain traditional and cognitive computer as left and right brain respectively(Figure 3).

Can cognitive computing learn to defeat the problemin thereal world ?

Figure 3. Cognitive or synaptic computer

Page 9: Radiomics: Novel Paradigm of Deep Learning for Clinical Decision Support toward Plan B using Liquid Biopsy

09Advanced Research Center for Nuclear TechnologyKARA Issue Paper Vol.04

Radiomics platform

Figure 4. Radiomics platform with ePAD and pCAD

A medical image informatics platform will be developed based on radiomics as shown in Figure 4. The platform process useful information by integration of big data from both imaging and genomic(and epigenomic) along with development of 3-dimensional(3D) or 4D Computer Assisted Diagnosis(CAD) and/or Clinical Decision Support(CDS), and provide information through a free software for quantitative imaging informatics such as ePAD. This radiomics strategy for identifying imaging biomarkers for prognostic phenotype may enable more rapid evaluation and development of novel imaging modalities, thereby accelerating their translation to personalized medicine.

Page 10: Radiomics: Novel Paradigm of Deep Learning for Clinical Decision Support toward Plan B using Liquid Biopsy

01 | Radiomic, novel imaging biomarkers, can be combined digitally with genomics particularly in the public domain as well as clinical informatics in the private domain by leveraging the public gene expression data containing clinical outcomes such as mortality from a closely matched population. This will lead to the advancement of public health through mass screening for early detection of the diseases.

02 | Radiation technology(RT) can be utilized to substantially integrate biotechnology (BT) and information technology(IT) particularly by developing a futuristic neuromorphic chip for cognitive computing as a Korean secure advanced technology, which will lead to new demands and advances in semiconductor technology.

03 | Fusion technology of neuroscience and cognitive computing, as a mutually beneficial platform will advance future Korean science and technology in parallel with U.S. Brain initiatives(Brain Research through Advancing Innovative Neurotechnologies) and E.U. Human Brain Project as well as Japanese Brain/MINDS(Brain Mapping by Integrated Neurotechnologies for Disease Studies).

Korean Association for Radiation Application

Expectation

Radiomics : Novel Paradigm of Deep Learningfor Clinical Decision Support10

Page 11: Radiomics: Novel Paradigm of Deep Learning for Clinical Decision Support toward Plan B using Liquid Biopsy

⊙ Decoding tumour phenotype by noninvasive imaging using a quantitative radiomics approach, Nature Communications 2014

⊙ Non-Small Cell Lung Cancer: Identifying Prognostic Imaging Biomarkers by Leveraging Public Gene Expression Microarray Data-Methods and Preliminary Results, Radiology 2012

⊙ Results of the Two Incidence Screenings in the National Lung Screening Trial, New England Journal of Medicine 2013

⊙ Detection of circulating tumor DNA in early- and late-stage human malignancies, Science Translational Medicine 2014

Korean Association for Radiation Application

References

※Please contact the following regarding inquiries about this report.

Korean Association for Radiation ApplicationProject supervisor : Tai-Jin ParkPhone : +82-2-3490-7120Email : [email protected]: Sol-Ah JangPhone : +82-2-3490-7106Email : [email protected]

Korea Atomic Energy Research InstituteProject supervisor : Sang-Hyun ParkPhone: +82-63-570-3370Email : [email protected]

11Advanced Research Center for Nuclear TechnologyKARA Issue Paper Vol.04

Page 12: Radiomics: Novel Paradigm of Deep Learning for Clinical Decision Support toward Plan B using Liquid Biopsy

18F, Seoul-forest IT Valley, 77, Seongsuil-ro,Seongdong-gu, Seoul, Korea, 133-822Tel | +82-2-3490-7114 Fax | +82-2-445-1014www.ri.or.kr