Abstract
Purpose / Objective: Manual labeling of individual {sup 192}Ir seeds in localization images for dosimetry of multi-strand low-dose-rate (LDR) implants is labor intensive, tedious and prone to error. The objective of this investigation is to develop computer-based methods that analyze digitized localization images, improve dosimetric efficiency, and reduce labeling errors. Materials and Methods: {sup 192}Ir localization films were digitized with a scanned-laser system and analyzed using Multiscale, Geometric, Statistical Pattern Recognition (MGSPR), a technique that recognizes and classifies pixels in gray-scale images based on their surrounding, neighborhood geometry. To 'teach' MGSPR how to recognize specific objects, a Gaussian-based mathematical filter set is applied to training images containing user-labeled examples of the desired objects. The filters capture a broad range of descriptive geometric information at multiple spatial scales. Principled mathematical analysis is used to determine the linear combination of filters from a large base set that yields the best discrimination between object types. Thus the sensitivity of the filters can be 'tuned' to detect specific objects such as{sup 192} Ir seeds. For a given pixel, the output of the filter is a multi-component feature vector that uniquely describes the pixel's geometric characteristics. Pixels with similar geometric attributes have feature vectors that
More>>
Citation Formats
Bird, William F, Chaney, Edward L, and Coggins, James M.
Identification of {sup 192}Ir seeds in localization images using a novel statistical pattern recognition approach and a priori information.
United States: N. p.,
1995.
Web.
doi:10.1016/0360-3016(95)97970-C.
Bird, William F, Chaney, Edward L, & Coggins, James M.
Identification of {sup 192}Ir seeds in localization images using a novel statistical pattern recognition approach and a priori information.
United States.
https://doi.org/10.1016/0360-3016(95)97970-C
Bird, William F, Chaney, Edward L, and Coggins, James M.
1995.
"Identification of {sup 192}Ir seeds in localization images using a novel statistical pattern recognition approach and a priori information."
United States.
https://doi.org/10.1016/0360-3016(95)97970-C.
@misc{etde_20409464,
title = {Identification of {sup 192}Ir seeds in localization images using a novel statistical pattern recognition approach and a priori information}
author = {Bird, William F, Chaney, Edward L, and Coggins, James M}
abstractNote = {Purpose / Objective: Manual labeling of individual {sup 192}Ir seeds in localization images for dosimetry of multi-strand low-dose-rate (LDR) implants is labor intensive, tedious and prone to error. The objective of this investigation is to develop computer-based methods that analyze digitized localization images, improve dosimetric efficiency, and reduce labeling errors. Materials and Methods: {sup 192}Ir localization films were digitized with a scanned-laser system and analyzed using Multiscale, Geometric, Statistical Pattern Recognition (MGSPR), a technique that recognizes and classifies pixels in gray-scale images based on their surrounding, neighborhood geometry. To 'teach' MGSPR how to recognize specific objects, a Gaussian-based mathematical filter set is applied to training images containing user-labeled examples of the desired objects. The filters capture a broad range of descriptive geometric information at multiple spatial scales. Principled mathematical analysis is used to determine the linear combination of filters from a large base set that yields the best discrimination between object types. Thus the sensitivity of the filters can be 'tuned' to detect specific objects such as{sup 192} Ir seeds. For a given pixel, the output of the filter is a multi-component feature vector that uniquely describes the pixel's geometric characteristics. Pixels with similar geometric attributes have feature vectors that naturally 'cluster', or group, in the multidimensional space called 'feature space'. After statistically quantifying the training-set clusters in feature space, pixels found in new images are automatically labeled by correlation with the nearest cluster, e.g., the cluster representing {sup 192}Ir seeds. One of the greatest challenges in statistical pattern recognition is to determine which filters result in the best labeling. Good discrimination is achieved when clusters are compact and well isolated from one another in feature space. The filters used in this study are unique in their ability to extract multiscale geometric information that discriminates one object from another object in gray-scale images. The labeling from MGSPR analysis can be refined using a post processing technique that incorporates a priori information including seed shape, intrastrand seed spacing, and the number of seeds per strand. The current implementation of post processing is user-guided but can be automated. The user interaction consists of pointing and clicking on a known seed in a strand. The algorithm then sequentially locates the remaining seeds belonging to that strand by fitting a predefined seed template to the seed pixels labeled by MGSPR. During the seed location process, the search is guided by the orientation of the template fit at each seed, the known seed spacing, and possible candidates for seeds further along the strand. The latter strategy permits correct labeling of poorly imaged seeds. Crossing and intertwining strands create ambiguous branching points that are resolved by minimizing an energy function that is correlated to the amount of strand bending, or flexion. Results: Challenging localization images from actual {sup 192}Ir implants were used in our study. The images contain up to {approx}20 strands, many of which intersect, and up to {approx}200 seeds, many of which partially or completely overlap with other seeds. Most seeds ({approx}95-100%) were labeled successfully. Non-seed pixels with seed-like geometric properties were sometimes labeled incorrectly. The post processing technique successfully resolved ambiguous overlapping seeds, extrapolated poorly visualized seeds that were not labeled, and eliminated incorrectly labeled non-seed pixels. User-guided post processing also labeled seeds that were not visualized at all, e.g., due to obscuration by contrast media or a radiopaque applicator. Conclusion: Our MGSPR approach and complementary post processing technique can be used to successfully label {sup 192}Ir seeds in digitized localization images. Overall user time and the opportunities for errors are great ly reduced.}
doi = {10.1016/0360-3016(95)97970-C}
journal = []
issue = {971}
volume = {32}
journal type = {AC}
place = {United States}
year = {1995}
month = {Jul}
}
title = {Identification of {sup 192}Ir seeds in localization images using a novel statistical pattern recognition approach and a priori information}
author = {Bird, William F, Chaney, Edward L, and Coggins, James M}
abstractNote = {Purpose / Objective: Manual labeling of individual {sup 192}Ir seeds in localization images for dosimetry of multi-strand low-dose-rate (LDR) implants is labor intensive, tedious and prone to error. The objective of this investigation is to develop computer-based methods that analyze digitized localization images, improve dosimetric efficiency, and reduce labeling errors. Materials and Methods: {sup 192}Ir localization films were digitized with a scanned-laser system and analyzed using Multiscale, Geometric, Statistical Pattern Recognition (MGSPR), a technique that recognizes and classifies pixels in gray-scale images based on their surrounding, neighborhood geometry. To 'teach' MGSPR how to recognize specific objects, a Gaussian-based mathematical filter set is applied to training images containing user-labeled examples of the desired objects. The filters capture a broad range of descriptive geometric information at multiple spatial scales. Principled mathematical analysis is used to determine the linear combination of filters from a large base set that yields the best discrimination between object types. Thus the sensitivity of the filters can be 'tuned' to detect specific objects such as{sup 192} Ir seeds. For a given pixel, the output of the filter is a multi-component feature vector that uniquely describes the pixel's geometric characteristics. Pixels with similar geometric attributes have feature vectors that naturally 'cluster', or group, in the multidimensional space called 'feature space'. After statistically quantifying the training-set clusters in feature space, pixels found in new images are automatically labeled by correlation with the nearest cluster, e.g., the cluster representing {sup 192}Ir seeds. One of the greatest challenges in statistical pattern recognition is to determine which filters result in the best labeling. Good discrimination is achieved when clusters are compact and well isolated from one another in feature space. The filters used in this study are unique in their ability to extract multiscale geometric information that discriminates one object from another object in gray-scale images. The labeling from MGSPR analysis can be refined using a post processing technique that incorporates a priori information including seed shape, intrastrand seed spacing, and the number of seeds per strand. The current implementation of post processing is user-guided but can be automated. The user interaction consists of pointing and clicking on a known seed in a strand. The algorithm then sequentially locates the remaining seeds belonging to that strand by fitting a predefined seed template to the seed pixels labeled by MGSPR. During the seed location process, the search is guided by the orientation of the template fit at each seed, the known seed spacing, and possible candidates for seeds further along the strand. The latter strategy permits correct labeling of poorly imaged seeds. Crossing and intertwining strands create ambiguous branching points that are resolved by minimizing an energy function that is correlated to the amount of strand bending, or flexion. Results: Challenging localization images from actual {sup 192}Ir implants were used in our study. The images contain up to {approx}20 strands, many of which intersect, and up to {approx}200 seeds, many of which partially or completely overlap with other seeds. Most seeds ({approx}95-100%) were labeled successfully. Non-seed pixels with seed-like geometric properties were sometimes labeled incorrectly. The post processing technique successfully resolved ambiguous overlapping seeds, extrapolated poorly visualized seeds that were not labeled, and eliminated incorrectly labeled non-seed pixels. User-guided post processing also labeled seeds that were not visualized at all, e.g., due to obscuration by contrast media or a radiopaque applicator. Conclusion: Our MGSPR approach and complementary post processing technique can be used to successfully label {sup 192}Ir seeds in digitized localization images. Overall user time and the opportunities for errors are great ly reduced.}
doi = {10.1016/0360-3016(95)97970-C}
journal = []
issue = {971}
volume = {32}
journal type = {AC}
place = {United States}
year = {1995}
month = {Jul}
}