TY - JOUR
T1 - Automated Cell Selection Using Support Vector Machine for Application to Spectral Nanocytology
AU - Miao, Qin
AU - Derbas, Justin
AU - Eid, Aya
AU - Subramanian, Hariharan
AU - Backman, Vadim
N1 - Publisher Copyright:
© 2016 Qin Miao et al.
PY - 2016
Y1 - 2016
N2 - Partial wave spectroscopy (PWS) enables quantification of the statistical properties of cell structures at the nanoscale, which has been used to identify patients harboring premalignant tumors by interrogating easily accessible sites distant from location of the lesion. Due to its high sensitivity, cells that are well preserved need to be selected from the smear images for further analysis. To date, such cell selection has been done manually. This is time-consuming, is labor-intensive, is vulnerable to bias, and has considerable inter- and intraoperator variability. In this study, we developed a classification scheme to identify and remove the corrupted cells or debris that are of no diagnostic value from raw smear images. The slide of smear sample is digitized by acquiring and stitching low-magnification transmission. Objects are then extracted from these images through segmentation algorithms. A training-set is created by manually classifying objects as suitable or unsuitable. A feature-set is created by quantifying a large number of features for each object. The training-set and feature-set are used to train a selection algorithm using Support Vector Machine (SVM) classifiers. We show that the selection algorithm achieves an error rate of 93% with a sensitivity of 95%.
AB - Partial wave spectroscopy (PWS) enables quantification of the statistical properties of cell structures at the nanoscale, which has been used to identify patients harboring premalignant tumors by interrogating easily accessible sites distant from location of the lesion. Due to its high sensitivity, cells that are well preserved need to be selected from the smear images for further analysis. To date, such cell selection has been done manually. This is time-consuming, is labor-intensive, is vulnerable to bias, and has considerable inter- and intraoperator variability. In this study, we developed a classification scheme to identify and remove the corrupted cells or debris that are of no diagnostic value from raw smear images. The slide of smear sample is digitized by acquiring and stitching low-magnification transmission. Objects are then extracted from these images through segmentation algorithms. A training-set is created by manually classifying objects as suitable or unsuitable. A feature-set is created by quantifying a large number of features for each object. The training-set and feature-set are used to train a selection algorithm using Support Vector Machine (SVM) classifiers. We show that the selection algorithm achieves an error rate of 93% with a sensitivity of 95%.
UR - http://www.scopus.com/inward/record.url?scp=84958568589&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=84958568589&partnerID=8YFLogxK
U2 - 10.1155/2016/6090912
DO - 10.1155/2016/6090912
M3 - Article
C2 - 26904682
AN - SCOPUS:84958568589
SN - 2314-6133
VL - 2016
JO - BioMed Research International
JF - BioMed Research International
M1 - 6090912
ER -