Joint tumor growth prediction and tumor segmentation on therapeutic follow-up PET images

Med Image Anal. 2015 Jul;23(1):84-91. doi: 10.1016/j.media.2015.04.016. Epub 2015 Apr 30.

Abstract

Tumor response to treatment varies among patients. Patient-specific prediction of tumor evolution based on medical images during the treatment can help to build and adapt patient's treatment planning in a non-invasive way. Personalized tumor growth modeling allows patient-specific prediction by estimating model parameters based on individual's images. The model parameters are often estimated by optimizing a cost function constructed based on the tumor delineations. In this paper, we propose a joint framework for tumor growth prediction and tumor segmentation in the context of patient's therapeutic follow ups. Throughout the treatment, a series of sequential positron emission tomography (PET) images are acquired for tumor response monitoring. We propose to take into account the predicted information, which is used in combination with the random walks (RW) algorithm, to develop an automatic tumor segmentation method on PET images. Moreover, we propose an iterative scheme of RW, making the segmentation more performant. Furthermore, the obtained segmentation is applied to the process of model parameter estimation so as to get the model based prediction of tumor evolution. We evaluate our methods on 7 lung tumor patients, totaling 29 PET exams, under radiotherapy by comparing the obtained tumor prediction and tumor segmentation with manual tumor delineation by expert. Our system produces promising results when compared to the state-of-the-art methods.

Keywords: Iterative random walks; Lung tumor; Tumor growth prediction; Tumor segmentation.

Publication types

  • Research Support, Non-U.S. Gov't

MeSH terms

  • Algorithms
  • Disease Progression
  • Humans
  • Imaging, Three-Dimensional / methods
  • Lung Neoplasms / diagnostic imaging*
  • Lung Neoplasms / pathology
  • Lung Neoplasms / radiotherapy
  • Positron-Emission Tomography / methods*
  • Predictive Value of Tests