Honours Thesis Presentation - Recurrent Convolutional Networks for Prostate Cancer Segmentation in MRI: DCE time series or Ktrans?
Title: Recurrent Convolutional Networks for Prostate Cancer Segmentation in MRI: DCE time series or Ktrans?
Supervisor: Dr. Thomas Trappenberg
Reader: Dr. Nauzer Kalyaniwalla
Dynamic Contrast Enhanced (DCE) Magnetic Resonance Imaging (MRI) is a method of imaging that has applications for diagnosing prostate cancer (PC). The method involves taking a time series of images of the body after the patient is injected with a contrast agent. Typically, machine learning models designed for the segmentation and detection of PC will use a scalar image called Ktrans to summarize the information in the DCE images.
This work proposes a model that combines the U-net fully convolutional network and the convGRU neural network architectures. The U-net is an architecture for segmenting scalar images and the convGRU is an architecture for analyzing time series of images. The proposed model was applied to interpret DCE time series in a temporal and spatial basis for the segmentation of PC. The purpose of this project is to determine whether PC segmentation models that use Ktrans are losing useful information present in the time series that could be extracted by the proposed model. Ultimately, experiments show that the proposed model using the DCE time series can outperform the baseline U-net segmentation model using Ktrans. However, when T2 and ADC, two other types of scalar MR images, are also considered by the models, no significant advantage is observed for the proposed model.
Room 429, Goldberg Computer Science BuildingHonours Thesis Presentation - Recurrent Convolutional Networks for Prostate Cancer Segmentation in MRI: DCE time series or Ktrans?