Panel Discussion Part I Methodology Ideas from adult MR brain segmentation are used in neonatal MR brain segmentation. However, additional challenges.

Panel Discussion Part I Methodology

Ideas from adult MR brain segmentation are used in neonatal MR brain segmentation. However, additional challenges in neonatal brain

segmentation require more advanced methods.

Could these methods be used again in adult MR brain segmentation to improve performance

and increase robustness?

Adult MR Brain Segmentation

Neonatal MR Brain Segmentation

Is an all-in-one approach for MR brain tissue segmentation in data from neonates to adults (and

everything in between) desirable?

Neonates AdultsMRI at different time points

pediatrics adolescents

Method development in medical image analysis often depends on the available data.

Are we limited by tuning algorithms on small datasets?

Do we focus enough on handling datasets with pathologies (which are of more clinical interest)?

Method Development

Evaluation

Could we develop more advanced and more robust algorithms if a wide variety of data would be available?

MR

Acqu

isiti

on

Age PathologyM

R Ac

quis

ition

Could we develop more advanced and more robust algorithms if a wide variety of data would be available?

MR

Acqu

isiti

on

Age PathologyM

R Ac

quis

ition

Do we need to fill this space before we can succeed?

Adult MR brain segmentation is one of the oldest medical image analysis problems. However, there is still

no method that is used in clinical practice?

Would more extensive evaluation help here?What is the role of researchers?What about the role of industry?

Decisions in the setting of clinical neurology mostly rely on qualitative (ie, visual) interpretation of structural brain images. On visual assessment, for example, global brain atrophy is recognized by an increase of cerebrospinal fluid (CSF) spaces and shrinkage of parenchymal structures.

Giorgio, A., De Stefano, N. (2013), “Clinical Use of Brain Volumetry”, JMRI, 37(1), 1–14.

Are automated segmentation tools really needed?

Do we need to join forces?

Combine expertise?

Combine methods?

Combine results of different methods?

Share data?

Panel Discussion Part IIData Sharing

Reasons for Sharing DataEconomic

• Cost of repeated data acquisition without new insights

• May lead to community efforts for development and

validation of image processing tools

• Avoid replication of existing data and tools

Maximize investment from funders

Reasons for Sharing Data

Engagement with patients and general public

• Volunteers and patients give their consent for

contributing to new knowledge and discovery

• Maximize participation of patients and volunteers

while keeping their right to opt out

Reasons for Sharing DataScientific

• Prevents selective and biased data analysis and

reporting

• Possibility to replicate results, to identifiy sub-

optimal or even wrong solutions

Reasons for Sharing Data

Credits

• Sharing data can be rewarding for funding and

publishing.

Advantages• Journal paper about

the data (credits, citations, clear description)

• Peer reviewed• …

Sharing Data via dedicated data journal (e.g. Nature Scientific Data)

Disadvantages• No direct comparison

of methods• Risk that only special

datasets will be accepted

• …

Sharing Data

Is it our responsibility as medical image analysis researchers?

Is it important for us to share data? Or rather to share evaluation data (e.g. manual segmentations)?

We actually do it with challenges… Could we take this one step further?

Advantages• Direct comparison of

methods• Opportunity to write a

journal paper• …

Sharing Data: Challenges

Disadvantages• Organizing a challenge

is time consuming• After a challenge is set

up, you cannot add data anymore

• …

How to Share Data?• Do we need standardization?

• Do we need quality control?– Ratio cost / added value

– Automatic vs. Human Quality Control

– Crowdsourcing: Let people rate the data?

• Infrastructure Issues– Data sharing repositories:

Big Data Centers vs Distributed Storage vs Peer to Peer storage (bioTorrent)?

– Code sharing: Is it different from data sharing?

How to Share Data?

There are no technological barriers for data sharing!

But sharing code is more difficult (and perhaps less desirable/acceptable)?

How to Share Data?

How do we reward data sharing?

Should we pay for access to data, especially if data has been annotated?

How to Share Data?

How can funding agencies/institutions help to promote data sharing?

Should funders decide which direction our field develops?

Or should we as a technical community decide what we need?

Closing Remarks