Quantitative 3-D surface reconstruction in fluorescence microscopy of multicellular specimens

This thesis provides a reconstruction technique for the analysis of cell surface shape and surface protein distributions from three dimensional wide-field fluorescence images. To attain accurate surface estimation, it is necessary to recover the loss of axial information inherent in wide-field image acquisition. Commonly used deconvolution techniques to estimate the 3-D intensity distribution in the specimen do not overcome the loss of z-spatial frequencies resulting in significant inaccuracies in any subsequent segmentation and analysis. Here we report that surface estimation and signal quantitation is improved using a new surface-constrained reconstruction (SCR) approach compared with standard deconvolution techniques. Briefly, the SCR is composed of two steps. First, an initial surface is estimated from the acquired data with a deformable surface model. This model, which requires no a-priori knowledge of the object, expands under the influence of four separate forces to lock on to the surface signal in the image data. Second, the surface estimate is further deformed to eliminate the axial extension in the data and to minimize the error between the acquired data and the image of the surface. To achieve this, the mathematical model of the surface is rendered to a 3D-voxel representation that can be convolved with the microscope point spread function. Forces that act on the surface position and signal concentration are then derived from the difference between the estimated and the actual image stacks. An iterative procedure jointly minimizes this difference and changes in surface curvature, converging to a surface that conforms to the image data while retaining smoothness and completeness. This reconstruction technique is tested in simulations first and subsequently in biological systems. Both cases demonstrate that axial extension is reduced and the likelihood of the final surface estimate is increased compared with standard deconvolution techniques. In simulation the final error in surface position is reduced to less than one voxel, and the error in signal concentration to +/-15%. Besides surface signals, many biological images are acquired with volume labels, for example DNA stains in developmental biology. The SCR algorithm is extended to provide quantitation of volume data within the surface boundary. The RMS error of integrated volume signals within the surface boundary is within 3%.