Click here to


Are you sure ?

Yes, do it No, cancel

Prediction of Uterus Volume Shrinkage for Cervical Cancer Patients During Radiotherapy Using Machine-Learning Approach with Treatment Planning-CT Radiomic Features

M Nakano1*, T Nakamoto2, Y Kumai1, Y Koizumi1, M Sumi1, K Nawa2, T Imae2, Y Yoshioka1, M Oguchi1, (1) Cancer Institute Hospital of JFCR, Koto-ku, Tokyo, JP, (2) The University of Tokyo Hospital, Bunkyo-ku, Tokyo, JP


(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: In some patients of cervical cancer, uterus commonly shows large shrinkage during concurrent chemo-radiotherapy (CCRT), and shrinkage prediction might be a clue of adaptive planning of radiotherapy (RT) and dose (de-)escalation. This study aims to predict uterus volume shrinkage for one month using machine-learning (ML) approach with radiomic features extracted from treatment planning-CT (pCT) for external beam radiotherapy (EBRT).
Methods: Pairs of pCT image series of 39 cervical cancer patients, who had pelvic or para-aortic nodal metastases, were acquired before whole pelvic EBRT (1st pCT) and before nodal boost EBRT (2nd pCT), and there were approximately one-month intervals. Volume ratio was calculated as the ratio of uterus volumes on 1st pCT and 2nd pCT. Shrunk group was defined as the ratio of 0.8 or less, and the others were classified as non-shrunk group. ML methods used in this study was logistic regression with L1 regularization term (LRL1) and L2 term (LRL2), and support vector classifier (SVC), provided in Scikit-Learn (Ver. 0.21.3) ML library. Prediction models were validated in leave-one-out cross-validation (LOOCV) framework. Radiomic features were extracted from uterus structures on 1st pCTs and 8 kinds of its wavelet-transformed images, using PyRadiomics Ver. 2.2.0. Those features were filtered using Pearson correlation coefficient with the threshold of 0.95, then 20 and 40 features were selected using Welch’s t-test in each LOOCV loop.
Results: Accuracy and AUC value of SVC with 40 features were 87.2% and 0.914, respectively, and LRL2 with 20 features was 71.8% and 0.693, respectively. The results of LRL1 were worse than others, especially in terms of specificity.
Conclusion: The model with support vector classifier showed the best results. The results imply that ML approaches with CT radiomic features to predict shrinkage of uterus is feasible, and this approach might help decisions of dose (de-)escalation in radiotherapy.

Funding Support, Disclosures, and Conflict of Interest: The present study was supported by JSPS KAKENHI Grant Number 18K15569.


Image Analysis, Texture Analysis, Feature Selection


IM/TH- Informatics: Informatics in Imaging (general)

Contact Email