On active contour models and balloons
CVGIP: Image Understanding
Machine Learning
International Journal of Computer Vision
Computer Vision and Image Understanding
Geodesic Active Regions and Level Set Methods for Supervised Texture Segmentation
International Journal of Computer Vision
Laplacian Eigenmaps for dimensionality reduction and data representation
Neural Computation
Level Set Evolution without Re-Initialization: A New Variational Formulation
CVPR '05 Proceedings of the 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05) - Volume 1 - Volume 01
MICCAI '09 Proceedings of the 12th International Conference on Medical Image Computing and Computer-Assisted Intervention: Part I
MICCAI'10 Proceedings of the 13th international conference on Medical image computing and computer-assisted intervention: Part III
Hi-index | 0.00 |
We present a high-throughput computer-aided system for the segmentation and classification of glands in high resolution digitized images of needle core biopsy samples of the prostate. It will allow for rapid and accurate identification of suspicious regions on these samples. The system includes the following three modules: 1) a hierarchical frequency weighted mean shift normalized cut (HNCut) for initial detection of glands; 2) a geodesic active contour (GAC) model for gland segmentation; and 3) a diffeomorphic based similarity (DBS) feature extraction for classification of glands as benign or cancerous. HNCut is a minimally supervised color based detection scheme that combines the frequency weighted mean shift and normalized cuts algorithms to detect the lumen region of candidate glands. A GAC model, initialized using the results of HNCut, uses a color gradient based edge detection function for accurate gland segmentation. Lastly, DBS features are a set of morpho-metric features derived from the nonlinear dimensionality reduction of a dissimilarity metric between shape models. The system integrates these modules to enable the rapid detection, segmentation, and classification of glands on prostate biopsy images. Across 23 H & E stained prostate studies of whole-slides, 105 regions of interests (ROIs) were selected for the evaluation of segmentation and classification. The segmentation results were evaluated on 10 ROIs and compared to manual segmentation in terms of mean distance (2.6 ± 0.2 pixels), overlap (62 ± 0.07%), sensitivity (85 ± 0.01%), specificity (94 ± 0.003%) and positive predictive value (68 ± 0.08%). Over 105 ROIs, the classification accuracy for glands automatically segmented was (82.5 ± 9.10%) while the accuracy for glands manually segmented was (82.89 ± 3.97%); no statistically significant differences were identified between the classification results.