Is a Deep Learning Based Segmentation Model Trained on planning CTs Transferable for Segmentation of Organs at Risk in Replanning CTs?

Abstract

To evaluate the differences in the segmentation of organs at risk (OARs) in planning and replanning radiotherapy CT images, and to assess the feasibility of using deep learning segmentation models trained on planning radiotherapy CTs for the contouring of OARs in replanning radiotherapy CTs. A total of 82 pairs of corresponding planning and replanning CT images from clinics were collected for contouring OARs in nasopharyngeal carcinoma patients. 14 of these were selected as the test set, and 20 OARs were selected for analysis. The deep learning model utilized in this study was the medical image segmentation framework, nnUNet. The test set of 14 replanning radiotherapy CT images was processed using different models trained on three training strategies: (A) 68 sets of planning CTs; (B) 68 sets of replanning CTs; (C) a mixed set of both 34 planning and replanning CTs. Additionally, the model trained by strategy A was also tested on the test set of 14 planning CT images. The segmentation results were evaluated using the Dice Similarity Coefficient (DSC). The average DSCs of the models trained using strategies A, B, and C on the test set of replanning CTs were (A) 0.54±0.28; (B) 0.57±0.28; (C) 0.56±0.27, respectively. On the test set of planning CTs, the average DSC of the model trained using strategy A was 0.64±0.25. These showed that when processing replanning CTs, the segmentation accuracy of the model trained using strategy A was significantly lower than that of the model trained using strategy B (p < 0.01), while the accuracy of the model trained using strategy C was improved compared to that of strategy A but still inferior to that of strategy B. Furthermore, the model trained on planning radiotherapy CTs alone (strategy A) showed a large difference in accuracy when processing planning and replanning CTs separately (p < 0.001). There is a significant difference in the segmentation of OARs in planning and replanning radiotherapy CT images, and the deep learning segmentation model constructed based on planning radiotherapy CTs is not suitable for the segmentation of OARs in replanning radiotherapy CT images. This highlights the need for re-modeling based on replanning CTs and also inspires us to incorporate the prior information contained in planning CTs and their labels into the OARs contouring of corresponding replanning radiotherapy CTs. These will, to some extent, provide insights into potential avenues for enhancing the future segmentation efficacy of adaptive radiotherapy.