Automated segmentation of target volumes in breast cancer radiotherapy, impact on target size and dose to organs at risk

Abstract

Introduction: Target volume delineation is crucial in breast cancer radiotherapy planning but involves significant interobserver variability. Deep learning (DL) models may reduce this variability, saving time and costs. However, current DL-models do not consider clinical data, such as tumor location and patient comorbidity, to adjust the target and reduce dose to organs at risk (OAR). This study compares clinically defined target volumes to those generated by a DL-model in terms of size, geometric overlap, and dose to OAR.

Method: For a hypothetical breast cancer patient, we compared target volumes constructed by Swedish radiotherapy clinics and two DL-models, Raystation and MVision. Geometrical overlap was evaluated, as well as the impact of differences in target delineation on dose to OAR. Treatment plans for locoregional vs. breast-only 3D-conformal radiotherapy were generated.

Results: CTV-structures for the breast, lymph nodes level I-IV, and internal mammary nodes were available for 10, 11, and 14 centers respectively. Volume of the CTV-breasts varied between 770-890cc, and the total CTV-volumes (breast + lymph nodes) between 875-1003cc. The DL-models did not constitute the largest nor smallest breast or total CTV-volumes, and geometric overlap between structures was relatively good. Evaluating dose to OAR from dose plans based on the respective CTV-volumes for locoregional radiotherapy, this was comparable between the DL-models and the mean of the CTVs generated by the clinics. In radiotherapy of only the breast, the CTV-breasts constructed by the DL-models gave the highest heart doses due to their proximity to the chest wall, affecting field angle choices. No difference was seen in dose to the ipsilateral lung, thyroid gland, or humeral head.

Conclusion: DL-models for target delineation have great potential. However, their introduction must be closely monitored since even small differences compared to clinical standards may affect doses to OAR in 3D conformal breast cancer radiotherapy.

Keywords: AI contouring; Breast cancer; Deep learning segmentation; Dosimetric data; Radiotherapy; Target volume delineation.

Fig. 1

A-d. Variability in target volume in CTV-breast (A), CTV lymph node level I-IV including the interpectoral nodes (CTVN) (B), CTV internal mammary chain (IMN) (C) and CTV-breast + CTVN + CTV-IMN (D).

Fig. 2

Axial CT image and topogram showing CTV breast, CTVN and CTV-IMN for all included centers. Raystation’s CTV is black, MVision’s CTV white and the Mean-CTV-structure green. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 3

CTV-breast dose plans illustrating the difference in dose to the heart between the delineations contributing with maximum (left picture, RS-DL) respective minimum heart dose (right picture, Center A).