Multi-object segmentation using shape particles
IPMI'05 Proceedings of the 19th international conference on Information Processing in Medical Imaging
| Hi-index | 0.00 |
The parenchymal and skeletal structure as recorded on chest radiographs can vary significantly from person to person. The person's height, width, age, gender, and other factors will result in significant variations in the presentation of these structures. As a result, the application of an active shape model (ASM) for segmentation can be problematic. The segmentation task can be made easier if, before the creation of the model for the ASM, the chest x-rays are classified according to some measure of similarity. The ASM rib-parenchyma model is defined using the 14 chest radiographs from the International Labor Organization(ILO) set of standard chest x-rays. Groundtruth was established by manually segmenting the ribs for the standard x-rays. A "leave one out" procedure was used to perform the tests on the ILO set. The measure of success was defined by the sensitivity of the segmentation in correctly classifying a pixel as parenchyma, i.e. not rib (Sensitivity = True Positive / (True Positive + False Negative). In one experiment, the chest radiographs were first pre-classified according to the number of parenchymal regions visible in each lung, and consecutive ASM model training and testing was performed over each class. In a second, similar experiment, the x-rays were first pre-classified by the height of the lungs, e.g. "tall" and "short" x-rays. It was found that the sensitivity improved an average of 0.20 over the baseline test (no pre-classification), when pre-classifying the lung by the number of parenchymal regions (8 or 9). Sensitivity improved an average of 0.19 over the baseline, when pre-classifying the lung by height (tall or short).