|
Our integrated computer-aided detection (CAD) scheme includes three basic modules. The first module detects
whether a microscopic digital image depicts a metaphase chromosome cell. If a cell is detected, the scheme will justify
whether it is analyzable with a decision tree. Once an analyzable cell is detected, the second module is applied to
segment individual chromosomes and to compute two important features. Specifically, the scheme utilizes a modified
thinning algorithm to identify the medial axis of a chromosome. By tracking perpendicular lines along the medial axis,
the scheme computes four feature profiles, identifies centromeres, and assigns polarities of chromosomes based on a set
of pre-optimized rules. The third module is followed to classify chromosomes into 24 types. In this module, each
chromosome is initially represented by a vector of 31 features. A two-layer classifier with 8 artificial neural networks
(ANN) is optimized by a genetic algorithm. A testing chromosome is first classified into one of the seven groups by the
ANN in the first layer. Another ANN is then automatically selected from the seven ANNs in the second layer (one for
each group) to further classify this chromosome into one of 24 types. To test the performance and robustness of this
CAD scheme, we randomly selected and assembled an independent testing dataset. The dataset contains 100 microscopic
digital images including 50 analyzable and 50 un-analyzable metphase cells identified by the experts. The centromere
location, the corresponding polarity, and karyotype for each individual chromosome were recorded in the "truth" file.
The performance of the CAD scheme applied to this image dataset is analyzed and compared with the results in the true
file. The assessment accuracies are 93% for the first module, 90.8% for centromere identification and 93.2% for polarity
assignment in the second module, over 96% for six chromosome groups and 81.8% for one group in the third module,
respectively. These accuracy levels are very comparable with those achieved during our previous studies to develop and
optimize these CAD modules. Hence, the study demonstrates that our automated scheme can achieve high and robust
performance in identification and classification of metaphase chromosomes.
|
