A general way to strengthen the reliability of a system is to use multiple
algorithms. Such techniques are very popular in the machine learning
society, and also in medical image processing. As applications, we can
mention atlas-based segmentations in three-dimensional confocal microscopy
images, breast tumour detection by applying common ensemble creation
techniques, e.g. bagging and boosting. Random subspace ensembles are
proved to be better than individual approaches in the case of brain
classification in functional magnetic resonance imaging. Combining
multiple algorithms to detect the optic disc in digital fundus images are
also reported to be efficient. A recent initiation is to organize
competitions for medical detectors, which also justifies the creation of
ensembles. For instance, in the case of liver segmentation, pulmonary
nodule detection, and microaneurysm detection, the combination of the
submitted results outperformed the individual approaches. As a more
specific field, we will focus on retinal imaging to see the efficiency of
ensemble-based methods. We will deal with the detection of several
anatomical parts of the retina: optic disc, macula, lesions
(microaneurysms, exudates), where an ensemble can be created considering
different detector algorithms. We will check how pointwise and regionwise
algorithms can be combined. Applying standard techniques to ensemble
creation e.g. in microaneurysm detection is not straightforward, since the
accuracy of the individual detectors are generally lower than it is
required. In such cases, we can successfully combine detector algorithms
with other complementary techniques, like different preprocessing methods.
When we are about to build up the ensemble from several individual
algorithms, usually a large search space should be scanned through to find
the final participant of such a system. The scan of large spaces needs
special search algorithms (like simulated annealing) which we will also
study.
|