A review of segmentation and deformable registration methods applied to adaptive cervical cancer radiation therapy treatment planning

Abstract

A c c e p t e d M a n u s c r i p t Highlights  This paper reviews the current state-of-the-art segmentation and deformable registration methods applied to cervical cancer adaptive radiation therapy planning. Strength and weaknesses of the registration and the segmentation methods are studied and analysed. Use of shape prior constraints can significantly reduce segmentation and registration errors.  Use of tissue specific classification of tumour may reduce tumour segmentation error. *Highlights (for review)Page 2 of 30 A c c e p t e d M a n u s c r i p t Abstract ObjectiveManual contouring and registration for radiotherapy treatment planning and online adaptation for cervical cancer radiation therapy in computed tomography (CT) and magnetic resonance images (MRI) is often necessary. However manual intervention is time consuming and may suffer from inter or intra rater variability. In recent years a number of computer-guided automatic or semi-automatic segmentation and registration methods have been proposed. Segmentation and registration in CT and MRI for this purpose is a challenging task due to soft tissue deformation, inter-patient shape and appearance variation and anatomical changes over the course of treatment. The objective of this work is to provide a state-of-the-art review of computer-aided methods developed for adaptive treatment planning and radiation therapy planning for cervical cancer radiation therapy. MethodsSegmentation and registration methods published with the goal of cervical cancer treatment planning and adaptation have been identified from the literature (PubMed and Google Scholar). A comprehensive description of each method is provided. Similarities and differences of these methods are highlighted and the strengths and weaknesses of these methods are discussed. A discussion about choice of an appropriate method for a given modality is provided. ResultsIn the reviewed papers a Dice similarity coefficient of around 0.85 along with mean absolute surface distance of 2-4 mm for the clinically treated volume were reported for transfer of contours from planning day to the treatment day. ConclusionsMost segmentation and non-rigid registration methods have been primarily designed for adaptive re-planning for the transfer of contours from planning day to the treatment day. The use of shape priors significantly improved segmen- 1Page 3 of 30 A c c e p t e d M a n u s c r i p t tation and registration accuracy compared to other models.

Autodelineation of cervical cancers using multiparametric magnetic resonance imaging and machine learning

“…As cervical cancers tend not to be constrained to the cervix, this technique is less relevant here. A variety of methods for voxel classification and image pre-processing has been used in the literature [11]. In our study, we saw only minor potential benefits of using the nonlinear QDA classifier instead of LDA.…”

“…No automatic segmentation methods for cervical cancers have been published [11]. For prostate cancer, however, several autodelineation algorithms have been presented [15,24].…”

“…No fully automatic delineation methods have been published for cervical cancers. Most studies, as recently reviewed by Ghose et al [11], focus on delineating organs at risk, or use the manual delineation from treatment planning to autodelineate tumours in images acquired during treatment using registration algorithms [12]. Methods proposed for identifying tumours in the cervix tend to have a high degree of manual interaction [11,13].…”

MR‐based treatment planning in radiation therapy using a deep learning approach

“…Specifically, after training the deep learning network, the acquired MR input images require minimal preprocessing, and pseudo CT images can be generated in less than a minute which is highly compatible with clinical workflows. Compared to conventional approaches where the soft tissue target (provided by MRI) is spatially coregistered to a planning CT using software tools, the utilization of deepMTP would enable planning to occur in the pixel space of the MR imaging data without the need for image coregistration. Note that geometric accuracy for the utilized MR scanners must be properly calibrated to ensure minimal distortion in the acquired MR images .…”

CBCT‐guided evolutive library for cervical adaptive IMRT

“…Such an implementation could be performed either through auto-segmentation 51 or with a nonrigid contour propagation with model constraint. [52][53][54] Moreover, delineation and dose calculation for replanning based on the CBCT is challenging due to limitations related to Hounsfield unit inconsistency, low image quality, and a limited field of view. In clinical practice, replanning could thus be performed by acquiring a new CT scan after treatment delivery.…”