Temporal Analysis of Tumor Heterogeneity and Volume for Cervical Cancer Treatment Outcome Prediction: Preliminary Evaluation |
| |
Authors: | Jeffrey W Prescott Dongqing Zhang Jian Z Wang Nina A Mayr William TC Yuh Joel Saltz Metin Gurcan |
| |
Institution: | (1) Department of Biomedical Informatics, The Ohio State University, 333 W. 10th Ave., Columbus, OH 43210, USA;(2) Department of Radiation Medicine, The Ohio State University Medical Center, 300 W. 10th Ave, Columbus, OH 43210, USA;(3) Department of Radiology, The Ohio State University Medical Center, 607 Means Hall, 1654 Upham Dr., Columbus, OH 43210, USA; |
| |
Abstract: | In this paper, we present a method of quantifying the heterogeneity of cervical cancer tumors for use in radiation treatment
outcome prediction. Features based on the distribution of masked wavelet decomposition coefficients in the tumor region of
interest (ROI) of temporal dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) studies were used along with the
imaged tumor volume to assess the response of the tumors to treatment. The wavelet decomposition combined with ROI masking
was used to extract local intensity variations in the tumor. The developed method was tested on a data set consisting of 23
patients with advanced cervical cancer who underwent radiation therapy; 18 of these patients had local control of the tumor,
and five had local recurrence. Each patient participated in two DCE-MRI studies: one prior to treatment and another early
into treatment (2–4 weeks). An outcome of local control or local recurrence of the tumor was assigned to each patient based
on a posttherapy follow-up at least 2 years after the end of treatment. Three different supervised classifiers were trained
on combinational subsets of the full wavelet and volume feature set. The best-performing linear discriminant analysis (LDA)
and support vector machine (SVM) classifiers each had mean prediction accuracies of 95.7%, with the LDA classifier being more
sensitive (100% vs. 80%) and the SVM classifier being more specific (100% vs. 94.4%) in those cases. The K-nearest neighbor
classifier performed the best out of all three classifiers, having multiple feature sets that were used to achieve 100% prediction
accuracy. The use of distribution measures of the masked wavelet coefficients as features resulted in much better predictive
performance than those of previous approaches based on tumor intensity values and their distributions or tumor volume alone. |
| |
Keywords: | |
本文献已被 SpringerLink 等数据库收录! |
|