A registration algorithm-guided framework for more accurate adaptive radiotherapy segmentation

Abstract

Purpose

Adaptive Radiotherapy (ART) adjusts treatment plans based on changes in the patient's anatomy as seen in re-planning Computed Tomography (CT) scans. The standard approach is to use alignment algorithms and segmentation models trained on planning CT scans, but these often do not yield good results on re-planning CTs. We propose a better segmentation framework that integrates prior anatomical information from planning CTs to enhance re-planning CT segmentation accuracy.

Methods

We propose a two-stage segmentation model that guides deep learning with image registration. This model has two main components: a registration algorithm and a deep learning segmentation model (use nnUNet framework). The registration algorithm calculates how to align planning CTs with re-planning CTs to create a consistent spatial map. Then, the adjusted planning CTs and their segmentations are fed into the deep learning model to help it segment the re-planning CTs more accurately. We've gathered 82 sets of CT data for this project, using 68 for training our model and 14 for testing its performance.

Results

The framework achieved more accurate segmentation on re-planning CTs, with an average Dice similarity coefficient (DSC) of 70.7 % across 20 OARs. This is higher than the DL model and registration alone by 4.9 % and 5.1 %, respectively. The mean DSC of the nnUNet trained on planning CTs for segmenting re-planning CTs was 62.5 %, significantly lower than that of its segmentation of planning CT (72.9 %). Conclusions: The framework significantly enhances the segmentation accuracy of re-planning CTs by integrating patient-specific prior information from planning CTs through guided registration within a population-based DL segmentation model. It has considerable potential to improve the clinical outcomes of radiation therapy plans.