Investigating the Feasibility of MRI Auto-segmentation for Image Guided Brachytherapy

A feasibility study has been performed to investigate the viability of applying auto-segmentation methods to the delineation of regions of interest (ROIs) in the treatment of cervical cancer using Image Guided Brachytherapy (IGBT). The introduction of auto-segmentation in IGBT aims to improve outlining consistency while improving patient experience by reducing the time taken to plan treatments. An anonymised database of MRI images and corresponding clinical ROI outlines was curated, categorised by brachytherapy treatment applicator type. This database was then used to train and test an autosegmentation model to contour the Bladder using three established algorithms, U-Net, SegNet and PSPNet. Quantitatively the U-Net model was found to produce contours geometrically closest to the original manual contours with a mean Dice Similarity Coefficient (DSC) of 0.942 compared to 0.919 and 0.879 for SegNet and PSPNet respectively and a mean Mean Distance to Agreement (mDTA) value of 0.46mm compared to 0.66mm and 0.89mm for SegNet and PSPNet. Visual assessment of the resulting contours demonstrated good agreement for the U-Net and SegNet produced outlines, particularly in the region of clinical significance, with greater variations seen at the extremities of the contour. In conclusion this feasibility study has shown that auto-segmentation methods can be applied to MRI IGBT contour delineation with a method established to facilitate further investigations in the application to all clinical ROIs and brachytherapy applicator types.