LaConte Lab

Supervised learning of fMRI with 3dsvm can be used for predicting brain states to enhance our understanding of brain systems, complementing the conventional emphasis on spatial mapping. The figure below illustrates the basic classification formalism used by 3dsvm. For each time point, the brain voxel intensities can be represented in a high-dimensional vector space. During an fMRI experiment, each image is a point in the vector space. Based on the locations of these points, a classification model can be estimated to distinguish between experimental states. After the model is determined, independent data can be assigned a class membership.

SVM brain state classification: (A) Each image is represented as an N voxel-dimensional vector. (B) A block design experiment and intensity changes for two voxels. Based on the distribution of intensities, a classification model (decision line) can be estimated. After the model is determined, it can be used to assign a brain state to new images. (C) Illustration of the SVM. The dotted box represents any (general) classifier. Inside the dotted box is a conceptual representation of the SVM approach.

AFNI began in the mid 1990's and is still an actively developing project written primarily in C (requires Unix, X11, and Motif). It provides great flexibility to the user through both the Unix shell and a gui interface. AFNI is used world-wide, open-source, maintained and developed by the Scientific and Statistical Computing Core at the National Institute of Mental Health (NIH, DHHS, USA, Earth), and provides great flexibility to the user through both the Unix shell and a gui interface.

SVM-Light is the core of 3dsvm. To be more compatible with fMRI data, 3dsvm reads AFNI-supported binary image formats, including the NIfTI format. It also allows for region of interest (ROI) 'masking' of variables (brain pixels), discarding training samples, and visualizing alphas and linear weight vectors. It can also handle multiple category classes. Currently 3dsvm uses SVM-Light V5.00.

We would like to thank Thorsten Joachims for providing SVM-Light to the scientific community, and for his enthusiastic support of this project. Similarly we would like to thank Bob Cox for AFNI and it's plugin support and Ziad Saad for his amazing support. Thanks also to Giuseppe Pagnoni and Stephen Strother for their comments and advice and Kyle Simmons and Michael Beauchamp for testing early versions of the plugin.

3dsvm is an active project in the LaConte Lab in the Department of Neuroscience at Baylor College of Medicine, Houston, TX (see developers). Stephen LaConte and Jeffrey Prescott developed early versions of 3dsvm at the Biomedical Imaging Technology Center (Xiaoping Hu, director) Department of Biomedical Engineering at Georgia Tech and Emory University. This inital work was supported by the National Institute of Health (NIH) National Institute of Neurological Disorders and Stroke (NINDS) R21NS050183.

interactive screen shot

Move mouse over regions of the interface below for further description.

download info

3dsvm is distributed with AFNI. You can go directly to the AFNI download page.

data and use

Usage:
------
	 3dsvm [options] 

Examples:
---------
1. Training: basic options require a training run, category (class) labels 
   for each timepoint, and an output model. In general, it usually makes 
   sense to include a mask file to exclude at least non-brain voxels

	 3dsvm -trainvol run1+orig \ 
	       -trainlabels run1_categories.1D \ 
	       -mask mask+orig \ 
	       -model model_run1

2. Training: obtain model alphas (a_run1.1D) and 
   model weights (fim: run1_fim+orig)

	 3dsvm -alpha a_run1 \
	       -trainvol run1+orig \ 
	       -trainlabels run1_categories.1D \ 
	       -mask mask+orig \ 
	       -model model_run1
	       -bucket run1_fim

3. Training: exclude some time points using a censor file 

	 3dsvm -alpha a_run1 \
	       -trainvol run1+orig \ 
	       -trainlabels run1_categories.1D \ 
	       -censor censor.1D \ 
	       -mask mask+orig \ 
	       -model model_run1
	       -bucket run1_fim

4. Training: control svm model complexity (C value)

	 3dsvm -c 100.0 \
	       -alpha a_run1 \
	       -trainvol run1+orig \ 
	       -trainlabels run1_categories.1D \ 
	       -censor censor.1D \ 
	       -mask mask+orig \ 
	       -model model_run1
	       -bucket run1_fim

5. Testing: basic options require a testing run, a model, and an output
   predictions file

	 3dsvm -testvol run2+orig \
	       -model model_run1+orig \
	       -predictions pred2_model1

6. Testing: compare predictions with 'truth' 

	 3dsvm -testvol run2+orig \
	       -model model_run1+orig \
	       -testlabels run2_categories.1D \
	       -predictions pred2_model1

7. Testing: use -classout to output integer thresholded class predictions
   (rather than continuous valued output)

	 3dsvm -classout \
	       -testvol run2+orig \
	       -model model_run1+orig \
	       -testlabels run2_categories.1D \
	       -predictions pred2_model1

todo

Improvements:

• clean up source code:
  1. update function definitions
  2. comment style is inconsistent
  3. error messages are not consistent
  4. re-check each malloc and free pair for error checking, etc.
• command line history should be in the model field - even when generated by the gui version. This was working at one point, and should be a minor fix.
• Try to conform completely to to svm-light structure (queryid and costfactor)
• Institute usual AFNI checks for overwrite protection for command line (hint: check_output_files and/or check_for_valid_inputs)
• The current approach to generating alpha files is not elegant

Extending Capabilities:

• enable afni flexibility for reading .1D files (e.g. labels.1D'[1]')
• enable multiple methods for doing multi-class classification
• enable gui based visualization of discriminant boundary (any 2D at a time)
• handle event-related data
• add the flexibility to use stim times
• allow multiple ways of visualizing models and adding them to the training output bucket.
• endow 3dsvm with utility functionality like extracting alphas from model headers, and converting data back and forth from SVM-Light's data structure.

developers

• Jeff W. Presscott (now in medical school)
• Will Curtis (now in medical school)

• Stephen M. LaConte
• Jonathan Lisinski
• Jason White
• Ziad Saad
• You? Help us catch bugs and extend 3dsvm! Contact Stephen LaConte ( ).

reference

• LaConte, S., Strother, S., Cherkassky, V. and Hu, X. 2005. Support vector machines for temporal classification of block design fMRI data. NeuroImage, 26, 317-329. pdf

BCM Home BCM Intranet Privacy Notices Contact BCM BCM Site Map (c)2007 Baylor College of Medicine Stephen M. LaConte - One Baylor Plaza, T-107, Houston TX 77030 Phone: 713-798-8499 | Fax: 713-798-4488 E-mail: