ON THE EFFECTIVENESS OF DOUBLY ADAPTIVE ESTIMATION FOR DYNAMIC MRI SEQUENCE ACQUISITIONS W. Scott Hoge * , Lawrence P. Panych Brigham and Women’s Hospital and Harvard Medical School Department of Radiology 75 Francis Street, Boston, MA 02115 Dana H. Brooks, Eric L. Miller, and Hanoch Lev-Ari Northeastern University ECE Department 360 Huntington Ave., Boston, MA 02115 ABSTRACT We present experimentally-acquired MR image sequence reconstruction results and a review of the recently proposed doubly adaptive temporal update method (DATUM) for the acquisition of dynamic MRI sequences. The DATUM al- gorithm is novel in providing an estimation and tracking framework for both image reconstruction and the image ac- quisition inputs. We discuss the difficulty in choosing viable system in- puts adaptively and compare DATUM to other minimal data MRI acquisition techniques. New results of image estimates constructed from data acquired on a standard production MRI scanner show that adapting system inputs to the se- quence provides substantial image quality improvement to dynamic sequence estimation. 1. INTRODUCTION While the clinical use of magnetic resonance imaging (MRI) is widespread today, physicians continue to push for better temporal and spatial resolution in dynamic MR im- age sequences. Historically, there have been two comple- mentary approaches to meeting this need. One approach is to enhance the physical hardware, for example by using an array of coils to improve temporal resolution. The comple- mentary approach is to develop algorithms which exploit the data redundancy in standard MRI data acquisition methods. Central to the algorithmic approach of dynamic se- quence acquisition are the two issues of image estimation and input selection. Previous algorithmic approaches have concentrated primarily on image estimation utilizing a static set of data acquisition inputs. Building on this approach, our group recently presented the theoretical fundamentals for a new MRI data acquisition algorithm [1] that provides a * Research supported by NIH NRSA Training Grant PA-00-103, and CenSSIS (the Center for Subsurface Sensing and Imaging Systems) under the Engineering Research Centers Program of the National Science Foun- dation (Award Number EEC-9986821). mechanism to adaptively determine both the image estimate and an appropriate set of inputs in tandem. By dynamically varying the inputs for each acquired image in the sequence, very high quality estimates of the dynamic sequence can be achieved. We present here a brief description of the DATUM method, a discussion on the difficulty of choosing new input from previous estimates, and results acquired on a standard production MRI scanner. 2. PROBLEM FORMULATION The DATUM method builds on the linear MRI acquisition system model described by Panych, et. al., in [2]. Under the assumption of low tip angles and rf encoding, the data ac- quisition process can be described as a matrix-vector prod- uct Y n = A n X n (1) where the columns of X n and Y n represent samples of the input and output rf waveforms, respectively, and A n repre- sents data for a single image slice at a given time n. For many image acquisition protocols, the acquisition time is proportional to the number of inputs used. Thus for mini- mal data acquisition and reconstruction, X n and Y n should be tall-thin matrices with r columns. For an image data ma- trix of size M × N , X n and Y n are of size N × r and M × r respectively. We choose to restrict the columns of the in- put matrix X n to be orthonormal, both to follow historical precedent and because it greatly simplifies the analysis. Due to the fundamental principles of MRI data acquisi- tion, A n is typically the frequency encoded, or k-space, rep- resentation of the image data. To view the image, this data can be transformed to the spatial domain through appropri- ate use of the unitary Fourier transform matrix [3, Chp. 5].