An open framework for parameter mapping and image processing

An open framework for parameter mapping and image processing

Apr 8, 2013Jason Su

Motivation

• Mapping methods share many common tasks and features– Fitting of a tissue model to data with a known signal equation– Multi-subject processing– Residuals and statistics analysis to evaluate fit– Pause and resume, recovery from crash

• We should promote code sharing in the community, so groups around the world can experiment with promising new mapping methods

Goals

• There are 2 audiences, coders and users• For developers of novel methods:– Provide a framework that fosters readability and code re-use– Easy to convert a bare-bones implementation into

production-level functionality, avoid code re-writes

– Provide easy ways to speed up code: embarrassing parallelization (across slices or subjects) and caching

– What else do developers need?• Data/error visualization• ?

Goals

• For clinicians and users:– A consistent workflow and usage pattern for producing

a multitude of mapping methods– Support for all platforms– Scalability to large-subject studies and diverse

computing resources• Amazon EC2 has built-in support for Python

• A community– Open source repository to share and improve

The Framework: Protocol class• Like when we sit down at

the scanner, we start with a protocol

• Input is readable JSON, not an inscrutable stream of command-line switches

• Instantiates a bunch of Series and manages access to their image data– I’ve assumed voxel-wise

fits

{ "Mask": { "filename": “brainMask_mask.nii.gz" }, "SPGR": [ { "filename”: "spgr_fa3000.nii.gz", "flip_angle": 3.0, "repetition_time": 6.74, "weight": 2 },

... { "filename": "spgr_fa8000.nii.gz", "flip_angle": 8.0, "repetition_time": 6.74 } ], "SSFP": [ { "filename": "ssfp_fa11000_ph0.nii.gz", "flip_angle": 11.0, "repetition_time": 3.6, "rf_phase_cycle": 0.0, "weight": 2 },

... ]}

Series class• Defines important scan

parameters, can pull from DICOM header

• Loads image from disk

• SPGR and SSFP are implemented– Flexible for any

combination of phase cycles, TRs, flip angles

• B0, B1/Kappa, ROI?

Utilities module• Memoization/caching of

expensive results

• pandas data structures inspired by R– Vectors and table data

frames with named access

• Notification of other functions when data changes

>> @memodef stats(x):

return (mean(x), std(x))

>> x = DataFrame(d, index=['a', 'b', 'c', 'd'])

P001 P002Types 'RRMS' 'SPMS'EDSS 3.5 6.0>> x['Bob']

>> y = notify_container_factory(DataFrame)(stats,data=x)

>> y.ix['EDSS'] = [4.0 7.0]-> stats() function is called and

recomputes mean and std. of EDSS

Implementing a signal equation• Define a TissueModel

– RelaxationTissueModel• Take parameters from TissueModel and compute expensive

intermediate parts of the signal equation:– SPGRSystemArray and SSFPSystemArray

• Finally, incorporate sequence parameters from a Series and compute the intensity at a voxel– SPGRSignal and SSFPSignal

• Series are able to discover the modules that model their signal behavior– Automatically looks for the module based on its name– SPGRSeries is tied to the sequence.spgr module

TissueModel class• A bare-bones base class that provides a flexible dictionary

hierarchy and notification> model.parameters['myelin']['fraction'] = 0.2

• Methods:– add_notifiee– del_notifiee– desync – do something when the model is changed– sync – call notifiees, propagate the changes

• Notifiees are called in the order they are added, duplicates are ignored

• Acts like a state machine

• Classes group/organize functions and the data they manipulate under common umbrellas

RelaxationTissueModel• Converts parameters to

pandas DataFrame for vectorized operations

• N-component relaxation model– mcDESPOT is usually 2-

component– Pure tissue, does not include

terms like RF phase cycle• desync enforces model

assumptions– Chemical equilibrium and unit

total volume

component_list = [ { 'name': 'intraextra', 'T1': 1000, 'T2': 70, 'off-resonance': 0.1, 'fraction': 0.75, 'exchange_rate': {'myelin': .1, 'csf': 0 } }, { 'name': 'myelin', 'T1': 800.0, 'T2': 30.0, 'off-resonance': 0.2, 'fraction': 0.2, 'exchange_rate': {'intraextra': .1, 'csf': 0 } }, { 'name': 'csf', 'T1': 5000.0, 'T2': 300.0, 'off-resonance': 0.3, 'fraction': 0.05, 'exchange_rate': {'intraextra': 0, 'myelin': 0 } }]

model = RelaxationTissueModel(component_list)

Simplifying Assumptions

• 2 component model

– Only need to find fF (the fast volume fraction)

• Chemical equilibrium

– Allows us to eliminate finding kSF=1/τS

• Both components are on the same resonance

€

fFkFS = fSkSF€

fS =1− fF

€

ΔωF = ΔωS

SPGR/SSFPSystemArray

• Subclassing promotes code re-use– SPGR is the simpler

form, so SPGR extends it• Computes shared

intermediate quantities• A shared cache

between different objects

• Notified by TissueModel

€

ASSFP =

− 1T2,F

− kFS kSF θRF + ΔωF 0 0 0

kFS − 1T2,S

− kSF 0 θRF + ΔωS 0 0

− θRF + ΔωF( ) 0 − 1T2,F

− kFS kSF 0 0

0 − θRF + ΔωS( ) kFS − 1T2,S

− kSF 0 0

0 0 0 0 − 1T1,F

− kFS kSF

0 0 0 0 kFS − 1T1,F

− kFS

⎡

⎣

⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢ ⎢

⎤

⎦

⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥ ⎥

SPGR/SSFPSignal• With phase cycle

separated out of the tissue model matrix, matrix multiply instead of expm– Adding more phase cycles

has minimal impact on processing time

• Current C implementation calculates all quantities for every image

Solvers

• Input a TissueModel, constraints, and initial guess to fit– Each of these are like TissueModel

• Stochastic Region Contraction, in progress– Store summary about each contraction step for

later viewing– Idea: chain a solver like Newton’s or gradient

descent once SRC has narrowed in

What is the advantage of all this structure?

• Thinking about code re-use and class structure helps to expose similarities in signal equations– Can lead to speed-ups as with SPGR/SSFP

• Improved debugging via unit testing and exceptions– e.g. Series raises exceptions giving valuable feedback

about what sequence parameters are missing in the input JSON

• Dynamic typing can allow for unexpected clever use by others

Open Source: Python vs. Matlab• Matlab Central is a nice central repository for interesting

code– But for the most part it is single contributor– Limits the possibility of large scale improvements to MATLAB

functionality• Python Package Index (PyPI)

– Has single line install for most everything– Just in the category of nearly drop-in acceleration of Python:

• Theano – GPU and generated C• Numba(Pro) – LLVM• PyPy/NumPyPy – total rewrite of CPython core interpreter• Numexpr – optimizes evaluation of math expressions including

considerations like CPU cache

Disadvantages

• Need to learn a new language• Loss of existing tools in MATLAB, including:– read_pfile– Extended phase graph analysis

• Development environment is different– Spyder tries to imitate MATLAB IDE

• Too many packages to install?– There are bundles that have all of the best packages

for Python with a double-click

Possible Future Features

• Access to super-large files with HDF5 and PyTables, important for high-res 32ch data

• Extending to handle more exotic cases– QSM– ?

• P-files and raw data?– Reconstruction does share some similar desired functions:

• Multi-subject processing, pause and resume

• Auto-suggest of possible fits given a Protocol• GUI for users?