Program Information
A 2.5D Assembly Framework with 2D CNNs to Segment Medical Images of Arbitrary Dimension
T Zhao*, D Ruan , UCLA School of Medicine, Los Angeles, CA
Presentations
MO-F-205-1 (Monday, July 31, 2017) 4:30 PM - 6:00 PM Room: 205
Purpose: Deep neural network has been widely applied in generic natural/synthetic image processing, but its application in high-dimensional medical imaging is challenged by the high memory/computation cost and specific medical considerations. This study aims to develop a learning framework to segment high-dimensional medical images in an economical way.
Methods: We design a 2.5D assembly framework based on native 2D convolutional neural networks (CNNs) for image segmentation. Specifically, independent 2D CNNs are used as weak classifiers along different dimensions, their classification results are aggregated using a Bayesian rationale to generate a preliminary field for segmentation label. This segmentation can be further polished by geometric regularization. The proposed framework naturally applies to segmentation tasks of various dimensionality without exponential cost growth, and is intrinsically amicable to parallel processing. Performance/cost assessment and comparison with the state-of-the-art 3D CNN are performed using clinical volumetric liver CT images.
Results: Compared to the 3D CNN based segmentation, the proposed method results in comparable performance in terms of Dice similarity coefficient (DSC). In addition, the computation cost of the proposed 2.5D scheme with patch size of mxm is reduced to O(m²) flops, compared to O(m³) flops for the 3D CNN method with patch size of mxmxm. In our implementation with m=25, the proposed 2.5D scheme was shown to increase the segmentation speed by a factor of ~35.
Conclusion: This work provides a 2.5D assembly framework based on 2D CNNs to segment arbitrarily high-dimensional medical images, and demonstrates similar performance but with less memory/computation demand, compared to the state-of-the-art 3D method. Such cost-saving benefit will be more pronounced as the problem dimensionality increases.
Contact Email: