Statistical learning of brain structure in DIPY

Statistical learning of human brain structure

Ariel Rokem, University of Washington eScience Institute

Follow along at http://arokem.github.io/2015-10-27-dipy/

My plan for today

Neuroimaging

Statistical learning ideas applied to neuroimaging

Open-source software that does it

Basic science questions

Happy to go off-script!

Normal behavior is supported by brain connectivity

Image from Catani and ffytche (2015)

Not just passive cables

Brain connections change with development

Individual differences account for differences in behaviour

Adapt with learning

This has clinical significance

Magnetic Resonance Imaging (MRI)

Neural activity: functional MRI

Anatomy: structural MRI

...

Brain connectivity: diffusion MRI

Diffusion MRI

Isotropic diffusion

Diffusion MRI

Anisotropic diffusion

Diffusion MRI

Modeling diffusion

Basser, Mattielo and Le Bihan (1994)

Diffusion statistics

Mean diffusivity

Fractional anisotropy

Principal diffusion direction

From diffusion to tracks

From diffusion to tracks

DIPY: Diffusion MRI in Python

Part of the NIPY community

Started in 2009 by Eleftherios Garyfallidis

Contributors from at least six different countries and many different labs

Garyfallidis et al.(2014)

Diffusion MRI: the challenge of validation

Algorithm 1

Algorithm 2

A statistical learning approach

In-vivo validation

Measurement #1

Measurement #2

Test-retest reliability

Model

Cross-validation

Rokem et al. (2015)

Dipy cross-validation API


				gtab = gradient_table(...)
				
				model = ReconstModel(gtab, ...)
				

				fit = model.fit(data, ...) # => ReconstFit
				

				prediction = fit.predict(gtab, ...)

For example


				
				model = dti.TensorModel(gtab)
				

				fit = model.fit(data1)
				

				prediction = fit.predict(gtab)
				

				RMSE = np.sqrt(\ 
 np.mean((prediction - data2) ** 2), -1))

rRMSE = RMSE / np.sqrt(\ np.mean((data1 - data2) ** 2), -1))

Rokem et al. (2015)

Corpus callosum

Corticospinal tract

Superior
longitudinal fasciculus

DTI

Crossing fiber model

Rokem et al. (2015)

When you've only measured once

k-fold cross-validation


				
				# Use a k of 2
				

				dti_pred = kfold_xval(dti_model, data, 2)

csd_pred = kfold_xval(csd_model, data, 2)

Algorithm 1

Algorithm 2

LiFE: Linear Fascicle Evaluation

Forward model from the tracks to the measured signal

Pestilli et al. (2014)

From diffusion to tracks

From tracks to diffusion

...

Pestilli et al. (2014)

Solve for


										>>> X.shape 

										   (10e8, 10e6)

Pestilli et al. (2014)


				
				
					fiber_model = life.FiberModel(gtab)
				

					fit = fiber_model.fit(data, tracks)
			

					prediction = fit.predict(gtab)
				

					optimized_tracks = tracks[fit.beta>0]