Principal Investigator/Program Director Williams,Robert W.
Assessment of feasibility
The major tests of feasibility to be performed are the accuracy of the reconstructed atlases and their cross alignment (Aim 1), the performance of 2.5DBA (Aim 2), and the intrinsic fiducialguided spline interpolation (Aim 3). In this section we explain how these test data will be evaluated.
Aim 1: Construction of mouse 3D brain atlases and navigation software
3D atlas reconstruction
To evaluate atlas reconstruction accuracy, the contiguous section faces on consecutive atlas sections will be examined. High-resolution images will be collected; bisected cells will be identified manually and their misregistration error measured. Figure 5 demonstrates convincingly that such a match between sections can be achieved. We estimate the accuracy of matching fiducial point setscentroids of cell nucleito be better than 4 mm (including error in converting slide coordinates to aligned-image coordinates). With the aid of the iScope (Project 2), we will collect 25 well-dispersed fiducials in each of 15 contiguous section pairs from all the different atlas animals. For this purpose, iScope accuracy will be enhanced by calibrating the position using intrinsic fiducials. The three section pairs to be examined from each animal will be collected from the ventral, middle, and dorsal regions. Our objective is to reduce misalignment to below 10 mm as determined by the error of the 14th worst measure of the 75 measures on each animal (i.e., ~ median + 1 standard deviation).Cross-atlas alignment
the remaining knots,
evaluating the error on the remaining knot, and cycling over the points
(leave-one-out paradigm). We will seek to achieve an accuracy of 20 m.
2: Linear alignment of the MBL into the standard coordinate system.
Performance of 2.5DBA will be assessed using two methods. In the first, we will employ 30mm-isotropic data sets (male C57BL/6J) similar to the one shown in Figure 3. These will be aligned to the male C57BL/6J atlas (cresyl violetstained) using 3DBA and 2.5DBA. For the latter, we will use every fifth section (i.e., the same unimodality sampling as for the MBL data). Taking the result of the former alignment as our gold standard, we will assess misregistration of the 2.5DBA. The placement error can be reduced to the maximum distance (the maximum distance error on the outer section contour) between the position of a given section under the two alignment regimes. Our objective is to obtain a mean of the absolute value of that error across the entire brain of less than 200 mm (i.e., a mean rotational error of less than 1.5, or a displacement of less than 7 pixels).
In a second performance task, we will compare matching accuracy between 2.5DBA and experts. As described below, human matching accuracy in the framework of rigid body is about 225 mm (using the same measure as described above). We are seeking to roughly match this level with the automatic system.
AIM 3: Segmentation of the MBL and extraction of quantitative traits
Implicit fiducialguided spline alignment.To test the virtual fiducial algorithm, we will use 45 MBL brains with 100 manually defined fiducials (due to the coarser z-plane of these relative to the atlas, delineation will take only about 1 month). Thirty-two of the brains will be aligned to the atlas using manual VOIs. Stable fiducials will then be automatically generated by examining the correlation at fixed standard coordinate positions across the data set. The stable set will than be used to align the remaining 13 brains, and the accuracy of the alignment will be assessed. Our objective is positional accuracy of 36 mm (the displacement of a 250-mm-diameter circular ROI leading to a 20% false positive and 20% false negative rate).