| | 1 | == Typical problems in vector reconstruction and their solutions == |
| | 2 | |
| | 3 | Most commonly available 2D atlases were not designed to be used as |
| | 4 | sources for 3D models generation. Precision of contour definitions |
| | 5 | and label placement is sufficient for standard use, but not always for |
| | 6 | the automatic reconstruction process. For example, in some published atlases |
| | 7 | one may find a number of deliberate simplifications e.g.~labels are |
| | 8 | omitted because names of some areas are considered obvious or labels |
| | 9 | are literally too large to put them inside the represented area. Such |
| | 10 | omissions have to be fixed in order to get a correct 3D |
| | 11 | reconstruction, indeed for any systematic computer processing. |
| | 12 | |
| | 13 | The vector parser tries to automatically deal with troublesome |
| | 14 | situations. However, if it is unable to correct a problem |
| | 15 | automatically, the program will notify the user and save all the |
| | 16 | important information about the case as a thumbnail image including |
| | 17 | name of the problematic label and its location. These data help the |
| | 18 | user to solve the problem manually by removing ambiguities from |
| | 19 | contour slides. Below we discuss common inconveniences and our |
| | 20 | strategies for fixing them. |
| | 21 | |
| | 22 | The most common defect is gaps in structure outlines. It frequently |
| | 23 | happens when two contour lines, which should touch are drawn so that |
| | 24 | they leave a little space in between. Such an arrangement may not be |
| | 25 | visible in a printed atlas but it greatly disturbs the tracing process |
| | 26 | where every pixel may influence the results. In this case, the |
| | 27 | structure being traced overtakes the space of its neighbour through |
| | 28 | the broken contour which we call ''leaking of the structure'' '''(see |
| | 29 | the spot indicated by a red arrow on Fig.~3A)'''. |
| | 30 | |
| | 31 | It is handled by a heuristic gap filling algorithm ([wiki:barSoftwareGapFillDetails detailed description]). The main idea |
| | 32 | behind this algorithm is to expand the contours by applying a [http://en.wikipedia.org/wiki/Mathematical_morphology dilation] filter ([http://www.imageprocessingplace.com/DIP-3E/dip3e_main_page.htm Gonzalez2008]) until the boundary closes. If a very |
| | 33 | accurate reconstruction is required, the best policy is to prepare a |
| | 34 | precise contour slide so there is no need to apply gap filling. For |
| | 35 | the atlases we tested, the gap filling algorithm performed well, the |
| | 36 | reconstructed structures were adequate and the gains in time were |
| | 37 | tremendous as compared with manual cleaning. |
| | 38 | |
| | 39 | Another common difficulty is an absence of borders drawn within an outline containing multiple labels. |
| | 40 | This can be a consequence of improper drawing but often |
| | 41 | simply follows from lack of sharp transitions between structures |
| | 42 | ('''see cortical areas in Figure...'''). |
| | 43 | When the parser detects two or more overlapping regions that were |
| | 44 | created based on different labels such labels are changed to |
| | 45 | spotlabels in the CAF slide and a notification is logged. Neither |
| | 46 | path is removed as there is no way to determine which one should be |
| | 47 | discarded and which should be kept. After tracing, the user is notified |
| | 48 | about ambiguous delineation and should review the slide in question |
| | 49 | and correct it. |
| | 50 | |
| | 51 | Error correction procedure also looks for labels anchored directly |
| | 52 | above contours which may happen when a structure is defined as a thin |
| | 53 | strip between other structures or placed outside the brain outline. Since |
| | 54 | tracing contours themselves is not allowed, such incidents are |
| | 55 | recorded and a thumbnail image is generated allowing user to examine |
| | 56 | and correct the error. Tracing of a region corresponding to the |
| | 57 | particular label is skipped but in general tracing is not interrupted |
| | 58 | as the whole procedure works non-interactively. |
| | 59 | |
| | 60 | Another problem we encountered during tracing was the lack of labels |
| | 61 | attributed to some areas. 3dBAR vector parser automatically detects |
| | 62 | such unlabeled regions by comparing the sum of traced paths with the |
| | 63 | total area of the ''Brain'' structure. The parser locates all |
| | 64 | contiguous unlabelled regions consisting of more than a given number |
| | 65 | of pixels and traces them. The resulting structures are called |
| | 66 | ''Unlabelled'' and may be indexed and reconstructed similarly to |
| | 67 | the regular structures. |
