TY - GEN
T1 - Evolution Strategy Classification Utilizing Meta Features and Domain-specific Statistical A Priori Models for Fully-automated and Entire Segmentation of Medical Datasets in 3D Radiology
AU - Zwettler, Gerald Adam
AU - Backfrieder, Werner
PY - 2015/10/5
Y1 - 2015/10/5
N2 - The employment of modern machine learning algorithms marks a huge advance towards automated and generalized segmentation in medical image analysis. Entire radiological datasets are classified, leading to a meaningful morphological interpretation, clearly distinguishing pathologies. After standard pre-processing, e.g. smoothing the input image data, the entire volume is partitioned into a large number of sub-regions utilizing watershed transform. These fragments are atomic and fused together building contiguous structures representing organs and typical morphology. This fusion is driven by similarity of regions. The relevant similarity measures respond to statistical a-priori models, derived from training datasets. In this work, the applicability of evolution strategy as classifier for a generic image segmentation approach is evaluated. Furthermore, it is analyzed if accuracy and robustness of the segmentation are improved by incorporation of meta features evaluated on the entire classification solution besides local features evaluated for the pre-fragmented regions to classify. The proposed generic strategy has a high potential in new segmentation domains, relying only on a small set of reference segmentations, as evaluated for different imaging modalities and diagnostic domains, such as brain MRI or abdominal CT. Comparison with results from other machine learning approaches, e.g. neural networks or genetic programming, proves that the newly developed evolution strategy is highly applicable for this classification domain and can best incorporate meta features for evaluation of solution fitness.
AB - The employment of modern machine learning algorithms marks a huge advance towards automated and generalized segmentation in medical image analysis. Entire radiological datasets are classified, leading to a meaningful morphological interpretation, clearly distinguishing pathologies. After standard pre-processing, e.g. smoothing the input image data, the entire volume is partitioned into a large number of sub-regions utilizing watershed transform. These fragments are atomic and fused together building contiguous structures representing organs and typical morphology. This fusion is driven by similarity of regions. The relevant similarity measures respond to statistical a-priori models, derived from training datasets. In this work, the applicability of evolution strategy as classifier for a generic image segmentation approach is evaluated. Furthermore, it is analyzed if accuracy and robustness of the segmentation are improved by incorporation of meta features evaluated on the entire classification solution besides local features evaluated for the pre-fragmented regions to classify. The proposed generic strategy has a high potential in new segmentation domains, relying only on a small set of reference segmentations, as evaluated for different imaging modalities and diagnostic domains, such as brain MRI or abdominal CT. Comparison with results from other machine learning approaches, e.g. neural networks or genetic programming, proves that the newly developed evolution strategy is highly applicable for this classification domain and can best incorporate meta features for evaluation of solution fitness.
UR - http://www.scopus.com/inward/record.url?scp=84961838972&partnerID=8YFLogxK
U2 - 10.1109/ICCCT2.2015.7292712
DO - 10.1109/ICCCT2.2015.7292712
M3 - Conference contribution
T3 - Proceedings of the International Conference on Computing and Communications Technologies, ICCCT 2015
SP - 12
EP - 18
BT - Proceedings of the International Conference on Computing and Communications Technologies, ICCCT 2015
PB - IEEE
T2 - 2015 IEEE International conference on Computing and Communications Technologies (ICCCT'15)
Y2 - 26 February 2015 through 27 February 2015
ER -