SPIE Membership Get updates from SPIE Newsroom
  • Newsroom Home
  • Astronomy
  • Biomedical Optics & Medical Imaging
  • Defense & Security
  • Electronic Imaging & Signal Processing
  • Illumination & Displays
  • Lasers & Sources
  • Micro/Nano Lithography
  • Nanotechnology
  • Optical Design & Engineering
  • Optoelectronics & Communications
  • Remote Sensing
  • Sensing & Measurement
  • Solar & Alternative Energy
  • Sign up for Newsroom E-Alerts
  • Information for:
SPIE Photonics West 2019 | Register Today

SPIE Defense + Commercial Sensing 2019 | Call for Papers

2019 SPIE Optics + Photonics | Call for Papers



Print PageEmail PageView PDF

Biomedical Optics & Medical Imaging

Independent component analysis in magnetic resonance imaging

Blindly separating images prior to applying tissue-contrast-enhancing algorithms shows clear improvements over current techniques.
12 May 2009, SPIE Newsroom. DOI: 10.1117/2.1200904.1498

Imaging using magnetic resonance (MR) provides superior soft-tissue contrast resolution over other screening techniques used in medicine. Analysis of the images is generally performed by spatial-domain-analysis-based processing techniques, yet, as with most techniques, improvements to the identification and separation of signal sources are being sought continuously. A recent approach that has demonstrated potential and promise1 considers a set of multiple MR frames as a single multispectral image, where each spectral band in the set is acquired during a particular pulse sequence. Independent component analysis (ICA) is then used to enhance tissue contrast. The technique is capable of blindly separating unknown signal sources into statistically independent components without prior knowledge. For the approach to be effective, however, there can be no more than one gaussian signal source in the data.2 Since noise in MR images is nongaussian,3 ICA seems a very appropriate analysis technique for magnetic resonance imaging (MRI).

Figure 1. (a) Simulated MR and (b) 'ground truth' images of brain-tissue substances shown in (a). Images were generated by the simulator at McGill University (http://www.bic.mni.mcgill.ca/brainweb/). PDW1: Proton-density weighted, T1W1: Spin-lattice weighted, T2W1: Spin-spin weighted.

While ICA has shown its strengths in MRI, it also suffers from a drawback, known as ‘overcomplete ICA.’ It occurs when p, the number of unknown signal sources (generally the number of tissue types), is greater than the number of combined images, L, used in the separation. This condition often occurs in brain imaging using MRI. The number of tissues of interest, such as cerebral spinal fluid (CSF), gray matter (GM), white matter (WM), skull, skin, and muscle, is always greater than the number of MR pulse sequences (e.g., spin-lattice and spin-spin relaxation times, proton density) obtainable. Under such circumstances, the three independent components resulting from ICA must accommodate p>3 signal sources, and thus more than one signal source must be accommodated in a single component. In this case, the full set of signal sources cannot be discriminated and separated in the image analysis.

To resolve this dilemma, we compared two new approaches to a standard classification technique. The first is a spectral-spatial approach, which first implements ICA to blindly separate signal sources and then applies the support-vector machine (SVM) to classify the signal sources in each independent component.4 The other uses the band-expansion process (BEP), which creates additional band images via nonlinear functions so that L becomes greater than p. This then makes ICA and spatial-domain techniques such as SVM effective.5

We used data from axial T1-weighted (spin-lattice), T2-weighted (spin-spin), and proton-density MR brain images with 5mm section thickness, 0% noise, and 0% intensity nonuniformity simulated by the MR imaging simulator at McGill University, Canada (see Figure 1). Using SVM as a spatial technique without prior application of ICA leads to poor classification of the central brain components (see Figure 2). However, the ICA+SVM spectral-spatial technique shows significant improvement over SVM alone. In this approach, the FastICA program (developed at the Helsinki University of Technology, Finland) was first implemented to separate the signal sources into three independent components, to which SVM was subsequently applied. The components relating to CSF, GM, and WM are clearly classified (see Figure 2).

We used BEP to generate three additional images by cross-correlating the original simulated MR images (see Figure 1). We then implemented the SVM and ICA+SVM methods using all six images as the combined image data. The results show once again that using ICA in conjunction with BEP and SVM greatly improves classification performance over either technique alone (see Figure 3).

Figure 2. Test of ICA+SVM applied to images generated by the simulator at McGill University. (a) SVM without prior application of ICA. (b) Independent components identified by applying the FastICA program. (c) Classification of tissues using ICA+SVM applied to the three components shown in (b).
Table 1. Tanimoto indices from the comparison of GM and WM MR images.
MethodTanimoto index (TI)

Figure 3. Classification of tissues using BEP+SVM and BEP+ICA+SVM. (a) Three cross-correlated band images generated from the original simulated images. Classification using (b) BEP+SVM and (c) BEP+ICA+SVM.

To quantitatively assess the four approaches, the Tanimoto index,4


was used as a measure, where A and B are two data sets (in this case the sets of binary information defining each image). If the images are accurate with little to no overlap between them, the intersect of A and B will be near null, and TI will approach 0. The more overlap between images, the higher the intersect and the lower the union, and TI will approach 1. The best performance was obtained using the BEP+ICA+SVM technique (TI = 0.8703), which used BEP to expand the original images, followed by ICA to produce independent components for classification by SVM (see Table 1).

We have thus demonstrated the advantages of using ICA in MR image analysis. Nevertheless, we continue to conduct investigations into improving the technique. In particular, there is a need to find appropriate sets of training samples derived directly from ICA without prior knowledge, to improve upon corrections for overcomplete ICA, and to design techniques to take advantage of ICA for follow-up image analysis. In addition, further work is required to look at how to extend single-slice approaches1,4,5 to a complete set of MR image slices and how to develop approaches to calculate partial volumes of tissues of interest, such as GM and WM.

Jyh-Wen Chai
Department of Radiology
Taichung Veterans General Hospital
Taichung, Taiwan

Jyh-Wen Chai received his MD and PhD degrees in biomedical engineering from the National Yang-Ming Medical School, Taipei, Taiwan, in 1984 and 2002, respectively. He is currently section chief and an assistant professor in the School of Medicine, China Medical University. His research interests include biomedical image processing, computed tomography (CT), MRI, and functional MRI (fMRI).

San-Kan Lee
Taichung Veterans General Hospital
Taichung, Taiwan

San-Kan Lee received his MD degree from the National Defense Medical Center, Taiwan, in 1973. From 1976 to 1989, he served at the Tri-Service General Hospital as a resident, visiting staff, and (eventually) director of the Division of Diagnostic Ultrasound. In 1982, he received a year of clinical fellowship training in the Division of Diagnostic Ultrasound at Thomas Jefferson University Hospital in Philadelphia (PA). From 1990 to 2003, he was the chair of the Department of Radiology at Taichung Veterans General Hospital. He is currently a vice president of the hospital. He is also a professor of Radiology at the National Defense Medical Center and Chung Shan Medical University. His research interests include computer-aided diagnosis (CAD) in x-ray mammograms, ultrasound screening of breast cancer, picture archiving and communication systems, and CT. He has published over 200 peer-reviewed papers and presented more than 50 abstracts at international conferences. The study of CAD has earned him one US and one Taiwanese patent.

Clayton Chi-Chang Chen
Department of Physical Therapy
Hungkung University
Taichung, Taiwan
Department of Radiology
Taichung Veterans General Hospital
Taichung, Taiwan

Clayton Chi-Chang Chen received his MD degree from China Medical College, Taichung, Taiwan, in 1981. He is currently the chair of the Department of Radiology, Veterans General Hospital, Taichung, Taiwan, and an associate professor at both the Department of Radiological Technology, Central Taiwan University of Science and Technology, and the Department of Physical therapy, National Yang-Ming University, Taipei, Taiwan. His research interests include biomedical image analysis, neuroradiology, CT, MRI, and fMRI.

Hsian-Min Chen
Department of Radiology
China Medical University Hospital
Taichung, Taiwan
Department of Medical Research
China Medical University Hospital
Taichung, Taiwan

Hsian-Min Chen received his BS and MS degrees from Huafan University, Taipei, Taiwan, in 1999 and 2001, respectively, and a PhD in electrical engineering from National Chung Hsing University, Taichung, Taiwan. He is currently a postdoctoral fellow. His research interests include digital and biomedical image processing.

Yen-Chieh Ouyang
Department of Electrical Engineering
National Chung Hsing University
Taichung, Taiwan

Yen-Chieh Ouyang received his BSEE degree in 1981 from Feng Chia University, Taiwan, and his MS (1987) and PhD (1992) degrees from the Department of Electrical Engineering, University of Memphis (TN). He is currently a professor and the director of the Multimedia Communication Laboratory. His research interests include hyperspectral image processing, medical imaging, communication networks, and network security.